Pith. sign in

REVIEW 3 major objections 4 minor 55 references

Efficient randomized algorithms for the fixed Tucker-rank problem of Tucker decomposition with adaptive shifts

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding a dynamically updated shift to the power iterations of randomized T-HOSVD and ST-HOSVD yields Tucker approximations with accuracy comparable to deterministic methods while cutting runtime.

desk verdict Randomized Tucker with adaptive shifts is a reasonable algorithmic extension with extensive experiments, but the main error bound as written relies on a wrong Gaussian norm estimate and an unproved lemma. read the letter →

arxiv 2506.04840 v1 pith:C4WS4TRE submitted 2025-06-05 math.NA cs.NA

Maolin Che , Yimin Wei , Chong Wu , Hong Yan This is my paper · ORCID
classification math.NAcs.NA MSC 65F5568W2015A1815A69
keywords TuckerdecompositionfixedTucker-rankrandomizedalgorithmshiftedpoweriterationT-HOSVDST-HOSVDprobabilisticerrorboundsingularvaluegap
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that inserting a dynamically updated shift into the power iterations of randomized T-HOSVD and ST-HOSVD yields approximate Tucker decompositions with accuracy comparable to, and in some settings better than, existing deterministic and randomized methods, while reducing runtime. The shift refines the singular value gap of each mode's unfolding, so the power iteration converges faster, and one shifted iteration can match several unshifted ones. The paper proves probabilistic error bounds of the form that the Frobenius error is bounded by sums of tail singular values multiplied by products of shifted singular-value ratios, which decay to zero as the number of power iterations grows. If correct, these algorithms solve the fixed Tucker-rank problem efficiently with $q=1$ power iteration and controlled error.

What carries the argument

The central object is the adaptive shifted power iteration on the mode-$k$ Gram matrix $A_{(k)}A_{(k)}^{\top}$. In each iteration the algorithm computes the SVD of $A_{(k)}(A_{(k)}^{\top}Q_k)-\alpha Q_k$ and updates $\alpha=(\Sigma_k(l_k,l_k)+\alpha)/2$ when the smallest sampled singular value exceeds it. This shift makes the ratios of consecutive singular values smaller, accelerating the convergence of the subspace iteration, and the error analysis bounds the mode-$k$ error through Lemma 4.11, which decomposes it into a range-approximation term and a projection term, each controlled by products of shifted singular-value ratios.

What would settle it

Run Algorithm 3 on a small tensor with known singular values and a flat-tailed spectrum, compute the quantities in inequality (4.7) for several shift sequences, and check whether the claimed inequality holds for all of them; a single counterexample with $\alpha>0$ would refute the theorem's proof, while verifying it on random tensors would support the claim.

Watch

Extended reading notes

Core claim

The central claim is that the mode-$k$ approximation error of a randomized Tucker decomposition can be reduced by replacing the plain power step $A_{(k)}A_{(k)}^{\top}Q_k$ with $A_{(k)}A_{(k)}^{\top}Q_k - \alpha Q_k$, where the shift $\alpha$ is updated adaptively as the smallest computed singular value grows. The paper establishes, in Theorems 4.1 and 4.5, that for Gaussian embedding matrices the Frobenius-norm error of the resulting Tucker approximation is bounded, with probability at least $1-\sum_k \Phi_k$, by a sum of tail singular values weighted by products of shifted singular-value ratios $(\sigma_{j+1}^2-\alpha)/(\sigma_j^2-\alpha)$, and that these ratio products tend to zero as the power parameter $q\to\infty$. The numerical experiments indicate that with $q=1$ the shifted algorithms match the relative error of deterministic T-HOSVD and ST-HOSVD while running faster.

Load-bearing premise

The load-bearing premise is Lemma 4.11, an inequality bounding the mode-$k$ error by two shifted power-iteration terms whose proof the paper leaves to the reader; if the shifted version of that inequality holds only under conditions not stated, the probabilistic error bound and its $q\to\infty$ convergence claim do not follow as proven.

Editorial extensions

If this is right

  • A single shifted power iteration per mode suffices to reach the accuracy of several unshifted iterations, so the algorithms reduce the constant in front of the dominant tensor-matrix multiply cost.
  • The error bound has the same tail-singular-value form as the deterministic T-HOSVD and ST-HOSVD guarantees, so the randomized methods inherit the classical approximation behavior in the large-$q$ limit.
  • Because the shift only uses the smallest sampled singular value from the previous iterate, the method needs no prior knowledge of the spectrum and can be implemented with the same per-iteration cost as the unshifted power scheme.
  • The ST-HOSVD variant (Algorithm 4) is the fastest of the compared randomized and deterministic methods on the tested real and synthetic tensors, suggesting a practical default for large-scale fixed-rank approximation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The shift update rule behaves like a Rayleigh-quotient shift on the sampled subspace, so it may combine naturally with block-Krylov subspace acceleration; a testable extension would compare shifted power iteration against a shifted block-Krylov scheme with the same per-iteration cost.
  • The benefit of the shift likely concentrates on spectra with moderate gaps; for flat spectra the ratio products decay slowly, and an adaptive stopping rule based on estimated gaps could decide when further shifts no longer help.
  • An automatic stopping rule based on the per-vector-error criterion could be augmented by the shift to terminate earlier; the paper's Appendix B suggests this is feasible, but the interaction between the shift and the per-vector-error bound is not analyzed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes two randomized algorithms (Algorithms 3 and 4) for the fixed Tucker-rank problem, obtained by inserting adaptive shifted power iterations into randomized T-HOSVD and ST-HOSVD. It states probabilistic Frobenius-norm error bounds (Theorems 4.1 and 4.5) that express the error in terms of sums of tail singular values multiplied by products of shifted singular-value ratios, which decay as the power parameter q grows. The numerical section compares the methods against several deterministic and randomized Tucker algorithms on synthetic and real tensors, reporting that the new algorithms achieve comparable accuracy with reduced runtime.

Significance. The algorithmic idea is natural and the experiments are extensive: the comparison set includes HOOI, T-HOSVD, ST-HOSVD, and a range of randomized variants (R-T-HOSVD, R-ST-HOSVD, RP-HOSVD, rSTHOSVDkron, Tucker-Sketch, etc.), and the reported runtimes show clear gains for the ST-HOSVD variant. If the error bounds were proved, the paper would provide a theoretically grounded way to reduce the number of power iterations while preserving accuracy. However, the main probabilistic guarantee is currently not established: Theorem 4.14 contains a dimensionally incorrect Gaussian norm bound, and the proof of Lemma 4.11 is omitted. These are load-bearing issues, not presentation details.

major comments (3)
  1. [Theorem 4.14, equation (4.9)] In Theorem 4.14 and its proof, the bound ||Phi_k||_2 <= sqrt(2 min{nhat_k, l_k} gamma_k) is dimensionally wrong in the regime used by the paper. The matrix Phi_k is (nhat_k - r_k) x l_k, and under the hypothesis l_k = r_k + s_k <= nhat_k - r_k it has at least as many rows as columns. Lemma 4.6 (from [40]) bounds the largest singular value of a Gaussian matrix with fewer rows than columns using the larger dimension; applying it to Phi_k^T gives ||Phi_k||_2 <= sqrt(2(nhat_k - r_k) gamma_k), not sqrt(2 min{nhat_k, l_k} gamma_k). Consequently the probability term in (4.9) must also use nhat_k - r_k rather than min{nhat_k, l_k}. As printed, the claimed event fails with probability essentially one in the experimental regime (e.g., nhat_k = 600, r_k = 10, l_k = 20 gives a 590 x 20 Gaussian matrix whose largest singular value is near sqrt(590)+sqrt(20) ~ 28.8, while the printed bound is sqrt(40) gamma ~ 6.4 gamma for gamma near 1). Since Theorem 4.1 is assembled directly from (4.10) and (4.11), the main error bound is not proved. The same min{., l_k} expression appears for Psi_k in Theorem 4.5, so that theorem inherits the flaw.
  2. [Lemma 4.11 and Remark 4.12] Lemma 4.11 states the key inequality (4.7) without proof. The paper says the proof is analogous to [40, Lemma A.1] and leaves it to the reader. Lemma 4.11 is the bridge that turns the mode-wise approximation error ||A x_k (U_k U_k^T) - A||_F into the two shifted power-iteration terms; both Theorem 4.1 and Theorem 4.5 rely on it. The present setting uses the Frobenius norm and shifted matrices, so the analogy to the unshifted spectral-norm result is not automatic. The authors should include a complete proof or provide a precise citation that covers this exact inequality.
  3. [Theorems 4.1 and 4.5, hypothesis on Phi_k and Psi_k] Theorems 4.1 and 4.5 require the existence of j_k, beta_k, gamma_k with 0 < sum Phi_k < 1 (resp. 0 < sum Psi_k < 1), but no numerical example reports such parameters or verifies the condition. Without this verification, the experimental section does not demonstrate that the proved bounds are applicable to any of the tested tensors. Please either provide admissible parameter values for the test cases or discuss the range of validity of the condition.
minor comments (4)
  1. [Abstract and Table 2] The abstract claims 'superior accuracy compared to deterministic approaches,' but Table 2 shows Algorithms 3 and 4 with slightly larger RE than T-HOSVD and ST-HOSVD (e.g., Yale: 8.48e-2 and 8.40e-2 vs 7.95e-2). The conclusion in Section 6 correctly describes the accuracy as 'comparable'; please align the abstract with the data.
  2. [Algorithms 2, 3, 4, A.1, A.2, B.3] In the first line of these algorithms, 'k = 1,2,...,n' should be 'k = 1,2,...,d' (the tensor order). In Section 3.2, 'rank(A(n))' should be 'rank(A(k))'.
  3. [Section 5, experimental protocol] The paragraph before Section 5.1 states that the power parameter is 10 for all randomized variants, but Section 5.4 says the power parameter is set to 1. Please make the experimental protocol consistent.
  4. [Table B.1] The column headers for tol = 0.01 read '(p1,p2,p3)' but should be '(q1,q2,q3)'.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular dependency found: the error bounds are expressed in terms of the input tensor's singular values and standard Gaussian concentration inequalities, and the shift parameter is adaptively updated from the sample rather than fitted to a target error.

full rationale

The central claim is that Algorithms 3 and 4 produce approximate Tucker decompositions with errors bounded by the displayed probabilistic estimates. In Theorem 4.1 and Theorem 4.5, the error is bounded by tail sums of singular values of the input (or sequential) unfoldings multiplied by ratios of shifted singular values. These quantities are intrinsic to the input tensor and the chosen Gaussian embedding; they are not defined in terms of the computed factors, the reported relative error, or any user-supplied target tolerance. The shift alpha is initialized to zero and updated by alpha = (Sigma_k(l_k,l_k)+alpha)/2 only when the current estimate exceeds alpha; this is a data-dependent acceleration heuristic from [21], not a parameter fitted to make the output match a desired error. The proof chain imports Gaussian concentration bounds from [40] (Lemmas 4.6-4.7) and low-rank approximation inequalities from [10,11]; although two of those lemmas are cited to the authors' own prior work, they are standard matrix inequalities with independent content, and the main structural lemma, Lemma 4.11, is explicitly deferred to [40, Lemma A.1] rather than being assumed by definition. Remark 4.12 leaves the verification of Lemma 4.11 to the reader; this is a proof gap that affects soundness, but it is not circularity, since the lemma is external to the paper's claimed new result. The condition 0 < sum Phi_k < 1 is a hypothesis on beta_k and gamma_k about the validity of the concentration events, not a fitted input. The numerical comparison against T-HOSVD, ST-HOSVD, and their randomized variants is an external benchmark, so the empirical claim is not derived from the theorem. The skeptic's dimensional objection to the Gaussian-norm estimate in Theorem 4.14 concerns correctness of the displayed proof, not equivalence of the conclusion to the assumptions. Overall, the derivation is self-contained in the sense that its inputs do not presuppose the target error bound; only a minor self-citation pattern appears in the supporting lemmas, justifying a low score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No invented physical or mathematical entities are introduced. The central theoretical bounds are expressed directly in terms of the input tensor's singular values. The only hand-picked quantities are experimental hyperparameters s and q, and the main unproved step is Lemma 4.11.

free parameters (2)
  • Oversampling per mode s_k = 10 in all experiments
    Selected from Figure 1 parameter sweeps as a default trade-off between relative error and runtime; not part of the theoretical bounds.
  • Power parameter q = 1 in main comparisons
    Chosen because Figures 1 and 2 show similar relative error for q=1 through q=8; the practical claim that one power iteration suffices depends on this choice.
assumptions (5)
  • standard math Gaussian concentration inequalities for the largest and smallest singular values of standard Gaussian matrices (Lemmas 4.6 and 4.7).
    Imported from Rokhlin et al. [40] and used throughout Theorems 4.13 and 4.14.
  • ad hoc to paper Lemma 4.11: the mode-wise Frobenius error is bounded by shifted range-finder terms.
    Stated without proof; the proof is deferred to an analogy with [40, Lemma A.1] in Remark 4.12. This lemma is load-bearing for Theorem 4.1.
  • standard math Interlacing relationship sigma_i(B_k) <= sigma_i(A_(k)) from [11, Lemma 9].
    Used in Remark 4.4 and Theorem 4.5 to replace singular values of the intermediate tensor B_k by those of the input unfolding A_(k).
  • domain assumption Existence of integers j_k and scalars beta_k, gamma_k > 1 such that 0 < sum Phi_k < 1.
    Required for the probability guarantees in Theorems 4.1 and 4.5; this condition is not verified in the numerical experiments.
  • domain assumption Random embeddings are standard Gaussian matrices.
    The theory is proved only for Gaussian embeddings; uniform and Khatri-Rao variants are tested only numerically in Appendix C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient randomized algorithms for the fixed Tucker-rank problem of Tucker decomposition with adaptive shifts." pith.science (2026). https://pith.science/paper/C4WS4TRE

@misc{pith2026250604840,
  author       = {Pith},
  title        = {Pith review of: Efficient randomized algorithms for the fixed Tucker-rank problem of Tucker decomposition with adaptive shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4WS4TRE}},
  note         = {Machine review of arXiv:2506.04840}
}
read the original abstract

Randomized numerical linear algebra is proved to bridge theoretical advancements to offer scalable solutions for approximating tensor decomposition. This paper introduces fast randomized algorithms for solving the fixed Tucker-rank problem of Tucker decomposition, through the integration of adaptive shifted power iterations. The proposed algorithms enhance randomized variants of truncated high-order singular value decomposition (T-HOSVD) and sequentially T-HOSVD (ST-HOSVD) by incorporating dynamic shift strategies, which accelerate convergence by refining the singular value gap and reduce the number of required power iterations while maintaining accuracy. Theoretical analyses provide probabilistic error bounds, demonstrating that the proposed methods achieve comparable or superior accuracy compared to deterministic approaches. Numerical experiments on synthetic and real-world datasets validate the efficiency and robustness of the proposed algorithms, showing a significant decline in runtime and approximation error over state-of-the-art techniques.

Figures

Figures reproduced from arXiv: 2506.04840 by the authors.

Figure 4
Figure 4. ]) and the sequentially T-HOSVD (ST-HOSVD) (see [47, Algorith [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 2
Figure 2. From these two figures, we conclude that: (a) with the same r, different choices of s and q are comparable in terms of RE; and (b) when fixing q = 1 (or s = 10), with the same r, CPU running time increases as s increases (or q increases). Hence, if there are no special requirements, we assume that s = 10 and q = 1. 2The extended Yale Face Database B is available at http://vision.ucsd.edu/~iskwak/ ExtYaleDatabase/Ext… view at source ↗
Figure 1
Figure 1. When fixing q = 1, numerical simulation results of Algorithms 3 and 4 with different r and s to the test tensor A. 5.3. Comparison with different processing orders in Algorithm 4. The processing order in Algorithm 4 is set to (1, 2, . . . , d). In general, any processing order in Sd is suitable for Algorithm 4. Now, with the same r, we will consider the efficiency of Algorithm 4 with different processing orders via … view at source ↗
Figures from the paper (5 more)
Figure 2
Figure 2. Figure 2: When fixing s = 10, numerical simulation results of Algorithms 3 and 4 with different r and q to the test tensor A. orders (1, 2, 3) and (1, 3, 2) are two bad unreasonable choices in terms of RE and CPU running time. 5.4. Comparison with different algorithms. In this s…
Figure 3
Figure 3. Figure 3: When fixing s = 10 and q = 1, numerical simulation results of Algorithms 3 and 4 with different r and processing orders to the test tensor A. Algorithm 4 is the fastest one, and Algorithm 3 is faster than R-T-HOSVD and RP-HOSVD, similar to Tucker-Sketch-SP and Tucker-S…
Figure 4
Figure 4. Figure 4: When fixing s = 10 and q = 1, numerical simulation results of Algorithms 3 and 4 with different r and processing orders on the test tensor C. 6. Conclusions In this paper, we discussed efficient randomized variants of T-HOSVD and ST￾HOSVD for the fixed Tucker-rank prob…
Figure 5
Figure 5. Figure 5: When fixing r = 10 and q = 1, numerical simulation results of Algorithms 3 and 4, and the existing randomized variants of T-HOSVD and ST-HOSVD to the test tensors A and B [PITH_FULL_IMAGE:figures/full_fig_p029_5.png]
Figure 6
Figure 6. Figure 6: When fixing r = 10 and q = 1, numerical simulation results of Algorithms 3 and 4, and the existing randomized variants of T-HOSVD and ST-HOSVD to the test tensors from three real databases. applying the randomized SVD algorithm (see [26]) to each mode unfolding matrix,…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 54 canonical work pages

  1. [40]

    3, 1100–1124

    Vladimir Rokhlin, Arthur Szlam, and Mark Tygert, A randomized algorithm for principal component analysis , SIAM Journal on Matrix Analysis and Applications 31 (2010), no. 3, 1100–1124

    work page 2010

  2. [1]

    Salman Ahmadi-Asl, Stanislav Abukhovich, Maame G Asante -Mensah, Andrzej Cichocki, Anh Huy Phan, Tohishisa Tanaka, and Ivan Oseledets, Randomized algorithms for compu- tation of Tucker decomposition and higher order SVD (HOSVD) , IEEE Access 9 (2021), 28684–28706

    work page 2021

  3. [2]

    4, 635–653

    Brett W Bader and Tamara G Kolda, Algorithm 862: MATLAB tensor classes for fast algorithm prototyping , ACM Transactions on Mathematical Software (TOMS) 32 (2006), no. 4, 635–653

    work page 2006

  4. [3]

    Bader and Tamara G

    Brett W. Bader and Tamara G. Kolda, MATLAB tensor toolbox version 3.6 (r2023b) , Available online, September 2023, URL:https://gitlab.com/tensors/tensor_toolbox/-/ releases/v3.6

    work page 2023

  5. [4]

    185, 1–36

    HanQin Cai, Keaton Hamm, Longxiu Huang, and Deanna Needel l, Mode-wise tensor decom- positions: Multi-dimensional generalizations of CUR deco mpositions, The Journal of Machine Learning Research 22 (2021), no. 185, 1–36

    work page 2021

  6. [5]

    3, 852–877

    Maolin Che, Juefei Chen, and Yimin W ei, Perturbations of the Tcur decomposition for tensor valued data in the Tucker format , Journal of Optimization Theory and Applications 194 (2022), no. 3, 852–877

    work page 2022

  7. [6]

    1, 395–428

    Maolin Che and Yimin W ei, Randomized algorithms for the approximations of Tucker and the tensor train decompositions , Advances in Computational Mathematics 45 (2019), no. 1, 395–428

    work page 2019

  8. [7]

    2, 373–403

    Maolin Che, Yimin W ei, and Yanwei Xu, Randomized algorithms for the computation of multilinear rank-( µ1, µ2, µ3) approximations , Journal of Global Optimization 87 (2023), no. 2, 373–403

    work page 2023

Show all 55 references
  1. [8]

    2, 605–636

    Maolin Che, Yimin W ei, and Hong Yan, The computation of low multilinear rank approxima- tions of tensors via power scheme and random projection , SIAM Journal on Matrix Analysis and Applications 41 (2020), no. 2, 605–636

    2020

  2. [9]

    , An efficient randomized algorithm for computing the approxim ate Tucker decompo- sition, Journal of Scientific Computing 88 (2021), no. 2, 32

    2021

  3. [10]

    , Randomized algorithms for the low multilinear rank approxi mations of tensors , Jour- nal of Computational and Applied Mathematics 390 (2021), 113380

    2021

  4. [11]

    , Efficient algorithms for Tucker decomposition via approxima te matrix multiplication , Advances in Computational Mathematics 51 (2025), article number 20

    2025

  5. [12]

    1, 256–297

    , Efficient randomized algorithms for fixed precision problem o f approximate tucker decomposition, SIAM Journal on Matrix Analysis and Applications 46 (2025), no. 1, 256–297

    2025

  6. [13]

    4-5, 249–429

    Andrzej Cichocki, Namgil Lee, Ivan Oseledets, Anh-Huy P han, Qibin Zhao, Danilo P Mandic, et al., Tensor networks for dimensionality reduction and large-sc ale optimization: Part 1 low- rank tensor decompositions, Foundations and Trends® in Machine Learning 9 (2016), no. 4-5, 249–429

    2016

  7. [14]

    6, 431–673

    Andrzej Cichocki, Anh-Huy Phan, Qibin Zhao, Namgil Lee, Ivan Oseledets, Masashi Sugiyama, Danilo P Mandic, et al., Tensor networks for dimensionality reduction and large- scale optimization: Part 2 applications and future perspec tives, Foundations and Trends ® in Machine Lear...

    2017

  8. [15]

    , rn) approximation of higher-order tensors , SIAM Journal on Matrix Analysis and Applications 21 (2000), no

    Lieven De Lathauwer, Bart De Moor, and Joos Vandewalle, On the best rank-1 and rank- (r1, r2, . . . , rn) approximation of higher-order tensors , SIAM Journal on Matrix Analysis and Applications 21 (2000), no. 4, 1324–1342. 38 MAOLIN CHE, YIMIN WEI, CHONG WU, AND HONG YAN 10 2...

    2000

  9. [16]

    1, 3475–3506

    Petros Drineas, Malik Magdon-Ismail, Michael W Mahoney , and David P W oodruff, Fast approximation of matrix coherence and statistical leverag e, The Journal of Machine Learning Research 13 (2012), no. 1, 3475–3506

    2012

  10. [17]

    Mahoney, A randomized algorithm for a tensor-based general- ization of the singular value decomposition , Linear Algebra and its Applications 420 (2007), no

    Petros Drineas and Michael W. Mahoney, A randomized algorithm for a tensor-based general- ization of the singular value decomposition , Linear Algebra and its Applications 420 (2007), no. 2-3, 553–571

    2007

  11. [18]

    1, A139–A163

    Virginie Ehrlacher, Laura Grigori, Damiano Lombardi, a nd Hao Song, Adaptive hierarchical subtensor partitioning for tensor compression , SIAM Journal on Scientific Computing 43 (2021), no. 1, A139–A163

    2021

  12. [19]

    3, 1315–1347

    Lars Eld´ en and Maryam Dehghan, A Krylov-Schur-like method for computing the best rank- (r1, r2, r3) approximation of large and sparse tensors , Numerical Algorithms 91 (2022), no. 3, 1315–1347

    2022

  13. [20]

    2, 248–271

    Lars Eld´ en and Berkant Savas, A Newton-Grassmann method for computing the best mul- tilinear rank- (r1, r2, r3) approximation of a tensor , SIAM Journal on Matrix Analysis and Applications 31 (2009), no. 2, 248–271

    2009

  14. [21]

    Xu Feng, W enjian Yu, Yuyang Xie, and Jie Tang, Algorithm 1043: Faster randomized SVD with dynamic shifts , ACM Transactions on Mathematical Software (TOMS) 50 (2024), no. 2, 1–27

    2024

  15. [22]

    6, 643–660

    Athinodoros S Georghiades, Peter N Belhumeur, and David Kriegman, From few to many: illumination cone models for face recognition under variab le lighting and pose , IEEE Trans- actions on Pattern Analysis and Machine Intelligence 23 (2001), no. 6, 643–660

    2001

  16. [23]

    Golub and Charles F

    Gene H. Golub and Charles F. Van Loan, Matrix Computations , fourth ed., Johns Hopkins University Press, Baltimore, MD, 2013

    2013

  17. [24]

    S. A. Goreinov, I. V. Oseledets, and D. V. Savostyanov, Wedderburn rank reduction and Krylov subspace method for tensor approximation. Part 1: Tu cker case , SIAM Journal on Scientific Computing 34 (2012), no. 1, A1–A27

    2012

  18. [25]

    1, 53–78

    Lars Grasedyck, Daniel Kressner, and Christine Tobler, A literature survey of low-rank tensor approximation techniques, GAMM-Mitteilungen 36 (2013), no. 1, 53–78

    2013

  19. [26]

    2, 217–288

    Nathan Halko, Per-Gunnar Martinsson, and Joel A Tropp, Finding structure with random- ness: Probabilistic algorithms for constructing approxim ate matrix decompositions , SIAM Review 53 (2011), no. 2, 217–288

    2011

  20. [27]

    2, 1054–1075

    Mohammad Hamed and Reshad Hosseini, Riemannian preconditioned coordinate descent for low multilinear rank approximation , SIAM Journal on Matrix Analysis and Applications 45 (2024), no. 2, 1054–1075

    2024

  21. [28]

    Behnam Hashemi and Yuji Nakatsukasa, RTSMS: Randomized Tucker with single-mode sketching, arXiv preprint arXiv:2311.14873 (2023)

    2023 arXiv

  22. [29]

    Mariya Ishteva, P.-A. Absil, Sabine Van Huffel, and Lieve n De Lathauwer, Best low multilinear rank approximation of higher-order tensors, based on the Ri emannian trust-region scheme , SIAM Journal on Matrix Analysis and Applications 32 (2011), no. 1, 115–135

    2011

  23. [30]

    3, 455–500

    Tamara G Kolda and Brett W Bader, Tensor decompositions and applications , SIAM Review 51 (2009), no. 3, 455–500

    2009

  24. [31]

    Ran zato, A

    Linjian Ma and Edgar Solomonik, Fast and accurate randomized algorithms for low-rank tensor decompositions , Advances in Neural Information Processing Systems (M. Ran zato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. W ortman Vaugh an, eds.), vol. 34, Curran Associates, Inc...

    2021

  25. [32]

    Ben gio, H

    Osman Asif Malik and Stephen Becker, Low-rank tucker decomposition of large tensors using tensorsketch, Advances in Neural Information Processing Systems (S. Ben gio, H. W allach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, eds.), vol. 31, Curran Asso- ciates, ...

    2018

  26. [33]

    1, 47–68

    Per-Gunnar Martinsson, Vladimir Rokhlin, and Mark Tyge rt, A randomized algorithm for the decomposition of matrices , Applied and Computational Harmonic Analysis 30 (2011), no. 1, 47–68

    2011

  27. [34]

    5, S485–S507

    Per-Gunnar Martinsson and Sergey Voronin, A randomized blocked algorithm for efficiently computing rank-revealing factorizations of matrices , SIAM Journal on Scientific Computing 38 (2016), no. 5, S485–S507

    2016

  28. [35]

    2, A1186–A1213

    Rachel Minster, Zitong Li, and Grey Ballard, Parallel randomized Tucker decomposition al- gorithms, SIAM Journal on Scientific Computing 46 (2024), no. 2, A1186–A1213. EFFICIENT RANDOMIZED ALGORITHMS FOR THE FIXED TUCKER-RANK PROBLEM OF TUCKER DECOMPOSITION WITH

    2024

  29. [36]

    1, 189–215

    Rachel Minster, Arvind K Saibaba, and Misha E Kilmer, Randomized algorithms for low-rank tensor decompositions in the Tucker format , SIAM Journal on Mathematics of Data Science 2 (2020), no. 1, 189–215

    2020

  30. [37]

    Cortes, N

    Cameron Musco and Christopher Musco, Randomized block Krylov methods for stronger and faster approximate singular value decomposition , Advances in Neural Information Processing Systems (C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. G arnett, eds.), vol. 28, Curran Assoc...

    2015

  31. [38]

    Rasmus Pagh, Compressed matrix multiplication, ACM Transactions on Computation Theory (TOCT) 5 (2013), no. 3, 1–17

    2013

  32. [39]

    Yichun Qiu, W eijun Sun, Guoxu Zhou, and Qibin Zhao, Towards efficient and accurate ap- proximation: tensor decomposition based on randomized blo ck Krylov iteration , Signal, Image and Video Processing 18 (2024), 6287–6297

    2024

  33. [41]

    Saibaba, HOID: higher order interpolatory decomposition for tensor s based on Tucker representation, SIAM Journal on Matrix Analysis and Applications 37 (2016), no

    Arvind K. Saibaba, HOID: higher order interpolatory decomposition for tensor s based on Tucker representation, SIAM Journal on Matrix Analysis and Applications 37 (2016), no. 3, 1223–1249

    2016

  34. [42]

    2, 891–918

    Berkant Savas and Lars Eld´ en,Krylov-type methods for tensor computations I , Linear Algebra and its Applications 438 (2013), no. 2, 891–918

    2013

  35. [43]

    6, 3352– 3393

    Berkant Savas and Lek-Heng Lim, Quasi-Newton methods on Grassmannians and multilinear approximations of tensors , SIAM Journal on Scientific Computing 32 (2010), no. 6, 3352– 3393

    2010

  36. [44]

    4, 1123–1150

    Yiming Sun, Yang Guo, Charlene Luo, Joel Tropp, and Madel eine Udell, Low-rank Tucker ap- proximation of a tensor from streaming data , SIAM Journal on Mathematics of Data Science 2 (2020), no. 4, 1123–1150

    2020

  37. [45]

    4, 1454–1485

    Joel A Tropp, Alp Yurtsever, Madeleine Udell, and Volkan Cevher, Practical sketching algo- rithms for low-rank matrix approximation , SIAM Journal on Matrix Analysis and Applications 38 (2017), no. 4, 1454–1485

    2017

  38. [46]

    3, 279–311

    Ledyard R Tucker, Some mathematical notes on three-mode factor analysis , Psychometrika 31 (1966), no. 3, 279–311

    1966

  39. [47]

    2, A1027–A1052

    Nick Vannieuwenhoven, Raf Vandebril, and Karl Meerberg en, A new truncation strategy for the higher-order singular value decomposition , SIAM Journal on Scientific Computing 34 (2012), no. 2, A1027–A1052

    2012

  40. [48]

    Vervliet, O

    N. Vervliet, O. Debals, L. Sorber, M. Van Barel, and L. De L athauwer, Tensorlab 3.0, Avail- able online, March 2016, http://tensorlab.net

    2016

  41. [49]

    3, A1714–A1746

    Mengyu W ang, Yajie Yu, and Hanyu Li, Randomized tensor wheel decomposition , SIAM Journal on Scientific Computing 46 (2024), no. 3, A1714–A1746

    2024

  42. [50]

    Chuanfu Xiao and Chao Yang, RA-HOOI: Rank-adaptive higher-order orthogonal iteratio n for the fixed-accuracy low multilinear-rank approximation of t ensors, Applied Numerical Mathe- matics 201 (2024), 290–300

    2024

  43. [51]

    3, 1339–1359

    W enjian Yu, Yu Gu, and Yaohang Li, Efficient randomized algorithms for the fixed-precision low-rank matrix approximation , SIAM Journal on Matrix Analysis and Applications 39 (2018), no. 3, 1339–1359

    2018

  44. [52]

    Qibin Zhao, Guoxu Zhou, Shengli Xie, Liqing Zhang, and An drzej Cichocki, Tensor ring decomposition, arXiv preprint arXiv:1606.05535 (2016)

    2016 arXiv

  45. [53]

    35, 2021, pp

    Yu-Bang Zheng, Ting-Zhu Huang, Xi-Le Zhao, Qibin Zhao, a nd Tai-Xiang Jiang, Fully- connected tensor network decomposition and its applicatio n to higher-order tensor completion , Proceedings of the AAAI conference on artificial intelligen ce, vol. 35, 2021, pp. 11071–11078

    2021

  46. [54]

    6, 2928–2940

    Guoxu Zhou, Andrzej Cichocki, and Shengli Xie, Fast nonnegative matrix/tensor factorization based on low-rank approximation , IEEE Transactions on Signal Processing 60 (2012), no. 6, 2928–2940

    2012

  47. [55]

    42 MAOLIN CHE, YIMIN WEI, CHONG WU, AND HONG YAN School of Mathematics and Statistics and State Key Laborato ry of Public Big Data, Guizhou University, Guiyang, Guizhou, P

    , Decomposition of big tensors with low multilinear rank , arXiv preprint arXiv:1412.1885 (2014). 42 MAOLIN CHE, YIMIN WEI, CHONG WU, AND HONG YAN School of Mathematics and Statistics and State Key Laborato ry of Public Big Data, Guizhou University, Guiyang, Guizhou, P. R. of ...

    2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.