Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

A Scalable Factorization Approach for High-Order Structured Tensor Recovery

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Riemannian gradient descent on the Stiefel manifold recovers Tucker, tensor-train, and orthogonal CP tensors with convergence guarantees that scale polynomially with the tensor order.

desk verdict The Tucker/TT RRC framework is solid and worth refereeing, but the headline 4r-RIP sensing result fails as proven: Appendix D conflates a rank-restricted norm with the full Frobenius norm. read the letter →

arxiv 2506.16032 v1 pith:DYBVLJ5C submitted 2025-06-19 cs.LG eess.SPmath.OC

Zhen Qin , Michael B. Wakin , Zhihui Zhu This is my paper · ORCID
classification cs.LGeess.SPmath.OC MSC 15A6965K1090C26
keywords RiemanniangradientdescentStiefelmanifoldtensor-traindecompositionTuckerrestrictedisometrypropertynonconvexoptimizationtensorsensinglinearconvergence
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

The paper aims to prove that one factorized optimization strategy, Riemannian gradient descent on the Stiefel manifold, can recover high-order tensors in Tucker, tensor-train, and orthogonal CP formats with guarantees that degrade only polynomially in the tensor order $N$. Previous analyses either targeted order-3 tensors, allowed exponential dependence on $N$, or required a restricted isometry property whose order grows with $N$. The authors show that any loss satisfying a restricted correlated gradient condition yields a Riemannian regularity condition for the factorized objective, and that RGD then converges linearly to the ground-truth tensor given a suitable initialization. In tensor sensing this reduces the RIP requirement from $(N+3)r$-RIP to $4r$-RIP, so the number of measurements can scale linearly with the degrees of freedom of the decomposition. If the proof holds, the paper supplies one common convergence analysis for several tensor formats instead of case-by-case treatments.

What carries the argument

The load-bearing objects are the Stiefel manifold of columnwise orthonormal matrices and the rotation-invariant distance $\mathrm{dist}^2(\{Z_i\},\{Z_i^*\}) = \min_{R_i} \sum_{i<\tilde N} \sigma^2(X^*)\lVert Z_i - Z_i^*R_i\rVert_F^2 + \lVert Z_{\tilde N} - Z_{\tilde N}^*R_{\tilde N}\rVert_F^2$, which handles the inherent rotation ambiguity in Tucker and tensor-train factorizations. RGD updates orthonormal factors by projecting the Euclidean gradient onto the tangent space of the Stiefel manifold and using a polar-decomposition retraction, while the remaining factor is updated in Euclidean space with a rescaled learning rate. The Riemannian regularity condition (RRC) is the key inequality: it combines a strong-convexity-like term along the factorization trajectories with a gradient-norm term, and the paper proves RRC from the restricted correlated gradient condition through polynomial-in-$N$ constants. This machinery is what converts the RCG condition into a linear-contraction proof that remains polynomial in the tensor order.

What would settle it

Run RGD on a tensor-sensing instance with random Gaussian measurements at a sample count where the $4r$-RIP constant is small but the $(N+3)r$-RIP constant is large; if the algorithm fails to converge linearly from the spectral initialization for some $N \ge 3$, the claimed relaxation would be contradicted. Alternatively, search numerically for a low-rank pair $X, X^*$ where the RCG condition holds with $αβ > 1/4$ yet the RRC contraction in (17) fails, since the proof uses $αβ \le 1/4$ in equations (54) and (71).

Watch

Extended reading notes

Core claim

The central claim is that the factorized tensor-recovery problem becomes tractable when all but one factor are constrained to be columnwise orthonormal. Under this canonical form, the authors establish a Riemannian regularity condition (Definition 1) for the factorized objective and prove in Theorem 2 that RGD with step size $\mu \le 2a_3$ and initialization satisfying $\mathrm{dist}^2(\{Z_i^{(0)}\},\{Z_i^*\}) \le a_1$ contracts the distance by the factor $1-2a_2\mu$ at every iteration. Lemmas 10 and 11 show that the restricted correlated gradient (RCG) condition implies this Riemannian regularity condition for both Tucker and tensor-train formats, with constants $a_1,a_2,a_3$ that depend polynomially on $N$ and on condition numbers. Theorem 4 derives the RCG condition from the $4r$-RIP with explicit constants $\alpha=(1-\delta_{2r})/2$ and $\beta=(1-\delta_{2r})/(2(1+\delta_{4r})^2)$. For tensor-train sensing, this relaxes the $(N+3)r_{tt}$-RIP required by prior work to $4r_{tt}$-RIP.

Load-bearing premise

The whole argument rests on the loss satisfying the Restricted Correlated Gradient condition with parameters $α,β$ and the product bound $αβ \le 1/4$ on the low-rank set near the truth; in tensor sensing this means the $4r$-RIP constant must be small enough, and the spectral initialization must land inside the basin $a_1$.

Editorial extensions

If this is right

  • For Tucker and tensor-train tensor sensing, RGD converges linearly to the ground truth when the measurement operator satisfies the $4r$-RIP and the spectral initialization satisfies the $3r$-RIP condition.
  • Both the initialization radius and the convergence rate depend polynomially on the tensor order $N$, so the algorithm avoids the exponential deterioration in $N$ that appeared in earlier Tucker and CP analyses.
  • Orthogonal CP decomposition is covered as a special case of the Tucker analysis, since its orthonormal factor matrices fit the same Stiefel-manifold update.
  • The same Riemannian regularity machinery applies to tensor factorization and to tensor completion with incoherent sampling, giving RCG constants for those losses as well.
  • The measurement count required for the guarantees scales linearly with the degrees of freedom of the decomposition up to a factor of $\log N$, matching the information-theoretic order of the parameter count.

Reading between the lines

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

  • A testable next question is whether the RIP requirement can be pushed below $4r$, since the proof uses the $2r$ inner-product preservation in Lemma 12 and the $4r$ norm comparison in (89); a $3r$-based argument, if it exists, would further reduce the sample complexity.
  • The same RRC-from-RCG route should extend to hierarchical Tucker and tensor-network formats whose canonical forms have exactly one non-orthonormal factor, but the paper only sketches this extension and does not compute the resulting constants.
  • The polynomial constants contain factors such as $N^2$ and condition-number squares, so the guarantees are polynomial but not small; practical users should expect the iteration count to grow noticeably with $N$ even though it does not explode exponentially.
  • The exponential sample complexity observed in tensor completion is inherent to the incoherence-based bound and is not removed by the polynomial convergence constants, so the optimization guarantee and the statistical requirement should be tracked separately.
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

4 major / 6 minor

Summary. The manuscript develops a unified local convergence theory for Riemannian gradient descent applied to factorized low-rank tensor recovery. The factors are constrained to the Stiefel manifold (all but the core in Tucker; all but the last factor in TT), and the loss is assumed to satisfy a Restricted Correlated Gradient (RCG) condition. The authors prove a Riemannian regularity condition (RRC) for the Tucker and TT parameterizations with explicit constants depending polynomially on the tensor order N, and from it derive linear convergence of RGD under a suitable initialization. For tensor sensing, they claim that the 4r-RIP suffices to imply the RCG condition, improving on the (N+3)rtt-RIP used in prior work, and they provide numerical experiments for sensing and completion.

Significance. If the main claims held, the paper would be a useful unification: one proof strategy covers Tucker, TT, and (claimed) orthogonal CP, with polynomial dependence on N and a RIP order independent of N. Strengths include explicit constants in Lemmas 10 and 11, a clear modular structure (RCG to RRC to convergence), and numerical verification of linear convergence. However, the proof of the headline RIP improvement has a concrete gap, and the orthogonal CP claim is not actually analyzed; these issues need to be fixed before the paper can be relied upon.

major comments (4)
  1. [Appendix D, Eq. (89)] The proof of Theorem 4 contains an invalid equality: the last line of Eq. (89) asserts that ∥∇G(X)−∇G(X*)∥_{F,2r} = ∥∇G(X)−∇G(X*)∥_F. This is not implied by the restricted-norm definition in Eq. (88); the restricted norm is an extremum over rank-≤2r unit tensors and is generally strictly smaller than the full Frobenius norm. For the sensing loss, ∇G(X)−∇G(X*) = (1/m)∑_{k}⟨A_k, X−X*⟩A_k is a generic linear combination of full-rank measurement tensors and has full tensor rank with probability one. Consequently Eq. (89) only yields a bound on the restricted gradient norm, not on the full gradient norm that appears in the RCG condition (16) and that is subsequently used in the RRC upper bounds (57) and (76). Theorem 4's claim that 4r-RIP implies RCG, which is the central advertised improvement over [46], is therefore not established by the manuscript as written.
  2. [Section 1, 'Special case: Orthogonal CP decomposition'] The manuscript claims orthogonal CP as a covered case (abstract, Table 1), but the text explicitly says 'we opt to omit the discussion on the orthogonal CP decomposition, as the convergence analysis of the Tucker format can be applied to it without the need for additional exploration.' This is not justified: the Tucker RRC in Lemma 10 allows an arbitrary core tensor S, while orthogonal CP requires a diagonal core, and the core update in Eq. (12) is an unconstrained gradient step that does not preserve diagonality. No CP-specific derivation, diagonal-core projection, or identifiability argument is provided. The claim that orthogonal CP follows from the Tucker analysis is therefore unsupported, and Table 1 should either be revised or the analysis completed.
  3. [Definition 2 and Lemma 2 / Table 2] The RCG condition in Definition 2 is stated with arbitrary α,β>0, but the proofs of Lemma 10 and Lemma 11 use the additional inequality αβ≤1/4: it appears in Eq. (54) ('where we use the inequality αβ≤1/4 [43]') and again in Eq. (71). This condition is not stated in Definition 2, Lemma 2, Table 2, or the informal Theorem 1. For a loss such as h(X)=∥X−X*∥²_F with α=β=1, the product violates the condition and the stated constants and contraction proof do not apply. The authors should either add αβ≤1/4 to the RCG definition or state explicitly in Lemma 2 and Table 2 that the constants hold for RCG parameters satisfying this product bound; otherwise the general unified claim is overstated.
  4. [Theorem 4, Eq. (23)] Even if the restricted/full gradient equality in Eq. (89) were corrected, the proof as written derives only a bound involving the rank-restricted gradient norm, whereas the RCG condition (16) and the subsequent RRC derivations in Appendices B and C use the full Frobenius norm of ∇h(X)−∇h(X*). A repair would require showing that the factor gradients in Eqs. (55)-(56) and (72) depend only on the low-rank projection of the tensor gradient, or an otherwise modified argument. As it stands, the 4r-RIP-to-RCG link is a load-bearing gap in the paper's main improvement over prior work.
minor comments (6)
  1. [Section 2.4] There is a typo: 'decompostions' should be 'decompositions'.
  2. [Definition 1, Eq. (15)] The notation 'P_TZ fN St = I' is nonstandard and confusing; it should be clarified that the projection onto the tangent space of the last factor is the identity because that factor is unconstrained.
  3. [Definition 3] The phrase 'any tensor decomposition X with rank r' is vague; the authors should specify that r denotes the rank tuple for Tucker or TT format, as done elsewhere in the paper.
  4. [Lemma 12] Lemma 12 states an inner-product preservation property 'for any tensor X1 and X2', but the polarization argument requires that X1, X2, and X1+X2 have rank at most r; this restriction should be stated explicitly.
  5. [Appendix D, Eq. (89)] In Eq. (89) the constant δ_{4r} is written without the format subscript; it should read δ_{4rtk} or δ_{4rtt} to match the Tucker/TT distinction used elsewhere.
  6. [Appendix E] The initialization analysis is written only for the Tucker case; the TT case is delegated to [46, Theorem 3]. This is acceptable, but the main text should state this delegation more prominently near Eq. (25).

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the main caveat is a non-circular proof gap in Appendix D.

full rationale

The derivation chain is not circular. The paper's logic is: the RCG condition (Definition 2) is assumed for a loss h; Lemmas 10 and 11 verify the Riemannian regularity condition (Definition 1) for Tucker and TT formats with explicit polynomial constants; and Theorem 2 then shows that RRC implies linear convergence of RGD. The constants a1, a2, a3 are produced by algebra from alpha, beta, N, r, sigma, and kappa; they are not fitted to data nor renamed predictions. In tensor sensing, Theorem 4 attempts to derive RCG from the 4r-RIP, and this is where the main caveat lies. In Appendix D, eq. (89) contains the step ||∇G(X)-∇G(X*)||_{F,2r} = ||∇G(X)-∇G(X*)||_F, but definition (88) defines the restricted norm as a maximum over rank-2r unit tensors, which is generally strictly smaller than the full Frobenius norm for the generic full-rank gradient of the sensing loss. This is a serious missing-support or proof gap in the claimed 4r-RIP-to-RCG implication, but it is a correctness gap rather than a circular identification: the RCG conclusion is not already contained in the RIP premise by construction, and no parameter is fitted and then reported as a prediction. The self-citations to the authors' JMLR paper [46] (TT distance lemmas and spectral initialization Theorem 3) are prior published statements with assumptions that do not include the RCG condition or the target 4r-RIP result; under the stated rules they count as independent support and do not make the framework circular. The score of 2 reflects only the presence of minor self-cited technical lemmas in the TT branch; no load-bearing circularity is established.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four layers: canonical orthonormal parameterization, RCG with the extra alpha*beta <= 1/4 product bound, RIP-based RCG for sensing, and the non-projected treatment of orthogonal CP. The TT distance lemmas are imported from the authors' prior work [46], while the Tucker RRC is derived fresh. The unstated alpha*beta bound and the CP omission are the weakest paid-for inputs.

free parameters (2)
  • step size mu = 0.5 in experiments; theory requires mu <= 2 a3
    Chosen by hand. The convergence rate depends on mu, but existence of a suitable small mu is established analytically rather than by fitting to data.
  • learning-rate ratio gamma = gamma = sigma^2(X*)
    Set to the squared largest singular value of the unknown target tensor to balance Stiefel-factor and core updates. This is a theoretical choice that requires estimation in practice.
assumptions (4)
  • domain assumption The target tensor admits a canonical decomposition with all but one factor column-wise orthonormal.
    Section 1.1, eq. (7). The whole Stiefel-manifold formulation and the distance metric in eq. (13) are defined against this parameterization.
  • ad hoc to paper The loss satisfies RCG with alpha*beta <= 1/4 on the relevant restricted set.
    Definition 2 only requires alpha, beta > 0, but Lemma 10 (eq. (54)) and Appendix C use alpha*beta <= 1/4. This extra product bound is not part of the stated RCG condition and fails for the paper's own example with alpha=beta=1.
  • domain assumption For tensor sensing, the measurement operator satisfies a sufficiently small 4r-RIP for RCG and a sufficiently small 3r-RIP for spectral initialization.
    Theorem 4 (eq. (23)) and spectral initialization bound (25). If the RIP constants are not small, the basin condition and the contraction argument fail.
  • ad hoc to paper Orthogonal CP can be treated as a special case of the Tucker analysis without modifying the core update.
    Section 1.1 states that the Tucker convergence analysis applies to orthogonal CP 'without the need for additional exploration'. No CP-specific RRC constants, gradient formulas, or projection onto a diagonal core are provided, and a free Euclidean core update does not preserve the diagonal CP constraint.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Scalable Factorization Approach for High-Order Structured Tensor Recovery." pith.science (2026). https://pith.science/paper/DYBVLJ5C

@misc{pith2026250616032,
  author       = {Pith},
  title        = {Pith review of: A Scalable Factorization Approach for High-Order Structured Tensor Recovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYBVLJ5C}},
  note         = {Machine review of arXiv:2506.16032}
}
abstract

Tensor decompositions, which represent an $N$-order tensor using approximately $N$ factors of much smaller dimensions, can significantly reduce the number of parameters. This is particularly beneficial for high-order tensors, as the number of entries in a tensor grows exponentially with the order. Consequently, they are widely used in signal recovery and data analysis across domains such as signal processing, machine learning, and quantum physics. A computationally and memory-efficient approach to these problems is to optimize directly over the factors using local search algorithms such as gradient descent, a strategy known as the factorization approach in matrix and tensor optimization. However, the resulting optimization problems are highly nonconvex due to the multiplicative interactions between factors, posing significant challenges for convergence analysis and recovery guarantees. In this paper, we present a unified framework for the factorization approach to solving various tensor decomposition problems. Specifically, by leveraging the canonical form of tensor decompositions--where most factors are constrained to be orthonormal to mitigate scaling ambiguity--we apply Riemannian gradient descent (RGD) to optimize these orthonormal factors on the Stiefel manifold. Under a mild condition on the loss function, we establish a Riemannian regularity condition for the factorized objective and prove that RGD converges to the ground-truth tensor at a linear rate when properly initialized. Notably, both the initialization requirement and the convergence rate scale polynomially rather than exponentially with $N$, improving upon existing results for Tucker and tensor-train format tensors.

Figures

Figures reproduced from arXiv: 2506.16032 by the authors.

Figure 1
Figure 1. Performance comparison of the RGD, in the tensor sensing, (a) for different [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Performance comparison of the RGD, in the tensor completion, (a) for different [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Unified Framework for Sample Complexity of Structured Quantum State Tomography under Noisy Observations

    quant-ph 2026-08 conditional novelty 6.0 of 10

    A unified sample-complexity bound for structured quantum state tomography under depolarizing state preparation and measurement noise, with a proof that noise-unaware estimators suffer an irreducible bias.

Reference graph

Works this paper leans on

92 extracted references · 68 canonical work pages · cited by 1 Pith paper

  1. [46]

    Guaranteed nonconvex factorization approach for tensor train recovery.Journal of Machine Learning Research, 25(383):1–48, 2024

    Zhen Qin, Michael B Wakin, and Zhihui Zhu. Guaranteed nonconvex factorization approach for tensor train recovery.Journal of Machine Learning Research, 25(383):1–48, 2024

    work page 2024

  2. [43]

    An optimal statistical and computational framework for generalized tensor estimation.The Annals of Statistics, 50(1):1–29, 2022

    Rungang Han, Rebecca Willett, and Anru R Zhang. An optimal statistical and computational framework for generalized tensor estimation.The Annals of Statistics, 50(1):1–29, 2022

    work page 2022

  3. [1]

    Interior-point method for nuclear norm approximation with application to system identification.SIAM Journal on Matrix Analysis and Applications, 31(3):1235–1256, 2010

    Zhang Liu and Lieven Vandenberghe. Interior-point method for nuclear norm approximation with application to system identification.SIAM Journal on Matrix Analysis and Applications, 31(3):1235–1256, 2010

    2010

  4. [2]

    Designing tensor-train deep neural networks for time-varying mimo channel estimation.IEEE Journal of Selected Topics in Signal Processing, 15(3):759–773, 2021

    Jing Zhang, Xiaoli Ma, Jun Qi, and Shi Jin. Designing tensor-train deep neural networks for time-varying mimo channel estimation.IEEE Journal of Selected Topics in Signal Processing, 15(3):759–773, 2021

    2021

  5. [3]

    Fast and robust quantum state tomography from few basis measurements

    Daniel Stilck Franc ¸a, Fernando GS Brand ˜ao, and Richard Kueng. Fast and robust quantum state tomography from few basis measurements. In16th Conference on the Theory of Quantum Computation, Communication and Cryptography (TQC 2021). Schloss Dagstuhl-Leibniz-Zentrum f¨ur Informatik, 2021

    2021

  6. [4]

    Quantum state tomography with tensor train cross approximation.arXiv preprint arXiv:2207.06397, 2022

    Alexander Lidiak, Casey Jameson, Zhen Qin, Gongguo Tang, Michael B Wakin, Zhihui Zhu, and Zhexuan Gong. Quantum state tomography with tensor train cross approximation.arXiv preprint arXiv:2207.06397, 2022

    arXiv 2022

  7. [5]

    Quantum state tomography for matrix product density operators.IEEE Transactions on Information Theory, 70(7):5030–5056, 2024

    Zhen Qin, Casey Jameson, Zhexuan Gong, Michael B Wakin, and Zhihui Zhu. Quantum state tomography for matrix product density operators.IEEE Transactions on Information Theory, 70(7):5030–5056, 2024

    2024

  8. [6]

    Mixture-rank matrix approximation for collaborative filtering.Advances in Neural Information Processing Systems, 30, 2017

    Dongsheng Li, Chao Chen, Wei Liu, Tun Lu, Ning Gu, and Stephen Chu. Mixture-rank matrix approximation for collaborative filtering.Advances in Neural Information Processing Systems, 30, 2017

    2017

Show all 92 references
  1. [7]

    Recovery of simultaneous low rank and two-way sparse coefficient matrices, a nonconvex approach.Electronic Journal of Statistics, 14(1):413–457, 2020

    Ming Yu, Varun Gupta, and Mladen Kolar. Recovery of simultaneous low rank and two-way sparse coefficient matrices, a nonconvex approach.Electronic Journal of Statistics, 14(1):413–457, 2020

    2020

  2. [8]

    Analysis of a complex of statistical variables into principal components.Journal of educational psychology, 24(6):417, 1933

    Harold Hotelling. Analysis of a complex of statistical variables into principal components.Journal of educational psychology, 24(6):417, 1933

    1933

  3. [9]

    Microbial co-occurrence relationships in the human microbiome.PLoS computational biology, 8(7):e1002606, 2012

    Karoline Faust, J Fah Sathirapongsasuti, Jacques Izard, Nicola Segata, Dirk Gevers, Jeroen Raes, and Curtis Huttenhower. Microbial co-occurrence relationships in the human microbiome.PLoS computational biology, 8(7):e1002606, 2012

    2012

  4. [10]

    Latent space models for dynamic networks.Journal of the American Statistical Association, 110(512):1646–1657, 2015

    Daniel K Sewell and Yuguo Chen. Latent space models for dynamic networks.Journal of the American Statistical Association, 110(512):1646–1657, 2015. 24

    2015

  5. [11]

    Poisson noise reduction with non-local pca.Journal of mathematical imaging and vision, 48(2):279–294, 2014

    Joseph Salmon, Zachary Harmany, Charles-Alban Deledalle, and Rebecca Willett. Poisson noise reduction with non-local pca.Journal of mathematical imaging and vision, 48(2):279–294, 2014

    2014

  6. [12]

    Expressive power of recurrent neural networks

    Valentin Khrulkov, Alexander Novikov, and Ivan Oseledets. Expressive power of recurrent neural networks. arXiv preprint arXiv:1711.00811, 2017

    2017 arXiv

  7. [13]

    Supervised learning with tensor networks.Advances in neural infor- mation processing systems, 29, 2016

    Edwin Stoudenmire and David J Schwab. Supervised learning with tensor networks.Advances in neural infor- mation processing systems, 29, 2016

    2016

  8. [14]

    Tensor-train recurrent neural networks for video classifica- tion

    Yinchong Yang, Denis Krompass, and V olker Tresp. Tensor-train recurrent neural networks for video classifica- tion. InInternational Conference on Machine Learning, pages 3891–3900. PMLR, 2017

    2017

  9. [15]

    Tensorizing neural networks

    Alexander Novikov, Dmitrii Podoprikhin, Anton Osokin, and Dmitry P Vetrov. Tensorizing neural networks. Advances in neural information processing systems, 28, 2015

    2015

  10. [16]

    Compressing recurrent neural network with tensor train

    Andros Tjandra, Sakriani Sakti, and Satoshi Nakamura. Compressing recurrent neural network with tensor train. In2017 International Joint Conference on Neural Networks (IJCNN), pages 4451–4458. IEEE, 2017

    2017

  11. [17]

    Long-term forecasting using tensor-train rnns

    Rose Yu, Stephan Zheng, Anima Anandkumar, and Yisong Yue. Long-term forecasting using tensor-train rnns. Arxiv, 2017

    2017

  12. [18]

    A tensorized transformer for language modeling.Advances in neural information processing systems, 32, 2019

    Xindian Ma, Peng Zhang, Shuai Zhang, Nan Duan, Yuexian Hou, Ming Zhou, and Dawei Song. A tensorized transformer for language modeling.Advances in neural information processing systems, 32, 2019

    2019

  13. [19]

    Tensor methods and recommender systems.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 7(3):e1201, 2017

    Evgeny Frolov and Ivan Oseledets. Tensor methods and recommender systems.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 7(3):e1201, 2017

    2017

  14. [20]

    Some mathematical notes on three-mode factor analysis.Psychometrika, 31(3):279–311, 1966

    Ledyard R Tucker. Some mathematical notes on three-mode factor analysis.Psychometrika, 31(3):279–311, 1966

    1966

  15. [21]

    Oseledets

    I. Oseledets. Tensor-train decomposition.SIAM Journal on Scientific Computing, 33(5):2295–2317, 2011

    2011

  16. [22]

    Rasmus Bro. Parafac. Tutorial and applications.Chemometrics and intelligent laboratory systems, 38(2):149– 171, 1997

    1997

  17. [23]

    On the uniqueness of multilinear decomposition of n-way arrays

    Nicholas D Sidiropoulos and Rasmus Bro. On the uniqueness of multilinear decomposition of n-way arrays. Journal of Chemometrics: A Journal of the Chemometrics Society, 14(3):229–239, 2000

    2000

  18. [24]

    Parsimonious tensor response regression.Journal of the American Statistical Associa- tion, 112(519):1131–1146, 2017

    Lexin Li and Xin Zhang. Parsimonious tensor response regression.Journal of the American Statistical Associa- tion, 112(519):1131–1146, 2017

    2017

  19. [25]

    Tucker tensor regression and neuroimaging analysis.Statistics in Biosciences, 10:520–545, 2018

    Xiaoshan Li, Da Xu, Hua Zhou, and Lexin Li. Tucker tensor regression and neuroimaging analysis.Statistics in Biosciences, 10:520–545, 2018

    2018

  20. [26]

    Hyperspectral image, video compression using sparse tucker tensor decomposition.IET Image Processing, 15(4):964–973, 2021

    Samiran Das. Hyperspectral image, video compression using sparse tucker tensor decomposition.IET Image Processing, 15(4):964–973, 2021

    2021

  21. [27]

    Multilinear tensor regression for longitudinal relational data.The annals of applied statistics, 9(3):1169, 2015

    Peter D Hoff. Multilinear tensor regression for longitudinal relational data.The annals of applied statistics, 9(3):1169, 2015

    2015

  22. [28]

    Efficient quantum state tomography.Nature communications, 1(1):149, 2010

    Marcus Cramer, Martin B Plenio, Steven T Flammia, Rolando Somma, David Gross, Stephen D Bartlett, Olivier Landon-Cardinal, David Poulin, and Yi-Kai Liu. Efficient quantum state tomography.Nature communications, 1(1):149, 2010

    2010

  23. [29]

    Efficient tomography of a quantum many-body system.Nature Physics, 13(12):1158–1162, 2017

    BP Lanyon, C Maier, Milan Holz ¨apfel, Tillmann Baumgratz, C Hempel, P Jurcevic, Ish Dhand, AS Buyskikh, AJ Daley, Marcus Cramer, et al. Efficient tomography of a quantum many-body system.Nature Physics, 13(12):1158–1162, 2017

    2017

  24. [30]

    Scalable quantum tomography with fidelity estimation.Physical Review A, 101(3):032321, 2020

    Jun Wang, Zhao-Yu Han, Song-Bo Wang, Zeyang Li, Liang-Zhu Mu, Heng Fan, and Lei Wang. Scalable quantum tomography with fidelity estimation.Physical Review A, 101(3):032321, 2020. 25

    2020

  25. [31]

    Matrix product density operators: Simulation of finite-temperature and dissipative systems.Physical review letters, 93(20):207204, 2004

    Frank Verstraete, Juan J Garcia-Ripoll, and Juan Ignacio Cirac. Matrix product density operators: Simulation of finite-temperature and dissipative systems.Physical review letters, 93(20):207204, 2004

    2004

  26. [32]

    Matrix product operator representations

    Bogdan Pirvu, Valentin Murg, J Ignacio Cirac, and Frank Verstraete. Matrix product operator representations. New Journal of Physics, 12(2):025012, 2010

    2010

  27. [33]

    Positive tensor network approach for simulating open quantum many-body systems.Physical review letters, 116(23):237201, 2016

    Albert H Werner, Daniel Jaschke, Pietro Silvi, Martin Kliesch, Tommaso Calarco, Jens Eisert, and Simone Montangero. Positive tensor network approach for simulating open quantum many-body systems.Physical review letters, 116(23):237201, 2016

    2016

  28. [34]

    Efficient description of many-body systems with matrix product density operators.PRX Quantum, 1(1):010304, 2020

    Ji ˇr´ı Guth Jarkovsk `y, Andr´as Moln´ar, Norbert Schuch, and J Ignacio Cirac. Efficient description of many-body systems with matrix product density operators.PRX Quantum, 1(1):010304, 2020

    2020

  29. [35]

    On manifolds of tensors of fixed tt-rank.Nu- merische Mathematik, 120(4):701–731, 2012

    Sebastian Holtz, Thorsten Rohwedder, and Reinhold Schneider. On manifolds of tensors of fixed tt-rank.Nu- merische Mathematik, 120(4):701–731, 2012

    2012

  30. [36]

    A singular value thresholding algorithm for matrix completion.SIAM Journal on optimization, 20(4):1956–1982, 2010

    Jian-Feng Cai, Emmanuel J Cand `es, and Zuowei Shen. A singular value thresholding algorithm for matrix completion.SIAM Journal on optimization, 20(4):1956–1982, 2010

    1956

  31. [37]

    Tensor decompositions for signal processing applications: From two-way to multiway component analysis

    Andrzej Cichocki, Danilo Mandic, Lieven De Lathauwer, Guoxu Zhou, Qibin Zhao, Cesar Caiafa, and Huy Anh Phan. Tensor decompositions for signal processing applications: From two-way to multiway component analysis. IEEE signal processing magazine, 32(2):145–163, 2015

    2015

  32. [38]

    Canonical polyadic de- composition with a columnwise orthonormal factor matrix.SIAM Journal on Matrix Analysis and Applications, 33(4):1190–1213, 2012

    Mikael Sørensen, Lieven De Lathauwer, Pierre Comon, Sylvie Icart, and Luc Deneire. Canonical polyadic de- composition with a columnwise orthonormal factor matrix.SIAM Journal on Matrix Analysis and Applications, 33(4):1190–1213, 2012

    2012

  33. [39]

    Probabilistic tensor canonical polyadic decomposition with orthogonal factors.IEEE Transactions on Signal Processing, 65(3):663–676, 2016

    Lei Cheng, Yik-Chung Wu, and H Vincent Poor. Probabilistic tensor canonical polyadic decomposition with orthogonal factors.IEEE Transactions on Signal Processing, 65(3):663–676, 2016

    2016

  34. [40]

    Revisit cp tensor decomposition: Statistical opti- mality and fast convergence.arXiv preprint arXiv:2505.23046, 2025

    Runshi Tang, Julien Chhor, Olga Klopp, and Anru R Zhang. Revisit cp tensor decomposition: Statistical opti- mality and fast convergence.arXiv preprint arXiv:2505.23046, 2025

    2025 arXiv

  35. [41]

    Nonconvex low-rank tensor completion from noisy data.Advances in neural information processing systems, 32, 2019

    Changxiao Cai, Gen Li, H Vincent Poor, and Yuxin Chen. Nonconvex low-rank tensor completion from noisy data.Advances in neural information processing systems, 32, 2019

    2019

  36. [42]

    Sparse tensor additive regression.Journal of machine learning research, 22(64):1–43, 2021

    Botao Hao, Boxiang Wang, Pengyuan Wang, Jingfei Zhang, Jian Yang, and Will Wei Sun. Sparse tensor additive regression.Journal of machine learning research, 22(64):1–43, 2021

    2021

  37. [44]

    On polynomial time methods for exact low-rank tensor completion.Foundations of Computational Mathematics, 19(6):1265–1313, 2019

    Dong Xia and Ming Yuan. On polynomial time methods for exact low-rank tensor completion.Foundations of Computational Mathematics, 19(6):1265–1313, 2019

    2019

  38. [45]

    Scaling and scalability: Provable nonconvex low-rank tensor estimation from incomplete measurements.Journal of Machine Learning Research, 23(163):1–77, 2022

    Tian Tong, Cong Ma, Ashley Prater-Bennette, Erin Tripp, and Yuejie Chi. Scaling and scalability: Provable nonconvex low-rank tensor estimation from incomplete measurements.Journal of Machine Learning Research, 23(163):1–77, 2022

    2022

  39. [47]

    Provable tensor-train format tensor completion by riemannian opti- mization.Journal of Machine Learning Research, 23(123):1–77, 2022

    Jian-Feng Cai, Jingyang Li, and Dong Xia. Provable tensor-train format tensor completion by riemannian opti- mization.Journal of Machine Learning Research, 23(123):1–77, 2022

    2022

  40. [48]

    Deep transfer tensor decomposition with orthogonal constraint for recommender systems

    Zhengyu Chen, Ziqing Xu, and Donglin Wang. Deep transfer tensor decomposition with orthogonal constraint for recommender systems. InProceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 4010–4018, 2021

    2021

  41. [49]

    Tensor decompositions for learning latent variable models.Journal of machine learning research, 15:2773–2832, 2014

    Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models.Journal of machine learning research, 15:2773–2832, 2014. 26

    2014

  42. [50]

    Low-rank matrix completion using alternating mini- mization

    Prateek Jain, Praneeth Netrapalli, and Sujay Sanghavi. Low-rank matrix completion using alternating mini- mization. InProceedings of the forty-fifth annual ACM symposium on Theory of computing, pages 665–674, 2013

    2013

  43. [51]

    Low-rank solutions of linear matrix equations via procrustes flow

    Stephen Tu, Ross Boczar, Max Simchowitz, Mahdi Soltanolkotabi, and Ben Recht. Low-rank solutions of linear matrix equations via procrustes flow. InInternational Conference on Machine Learning, pages 964–973. PMLR, 2016

    2016

  44. [52]

    A unified computational and statistical framework for noncon- vex low-rank matrix estimation

    Lingxiao Wang, Xiao Zhang, and Quanquan Gu. A unified computational and statistical framework for noncon- vex low-rank matrix estimation. InArtificial Intelligence and Statistics, pages 981–990. PMLR, 2017

    2017

  45. [53]

    Global optimality in low-rank matrix optimiza- tion.IEEE Transactions on Signal Processing, 66(13):3614–3628, 2018

    Zhihui Zhu, Qiuwei Li, Gongguo Tang, and Michael B Wakin. Global optimality in low-rank matrix optimiza- tion.IEEE Transactions on Signal Processing, 66(13):3614–3628, 2018

    2018

  46. [54]

    Nonconvex robust low-rank matrix recovery.SIAM Journal on Optimization, 30(1):660–686, 2020

    Xiao Li, Zhihui Zhu, Anthony Man-Cho So, and Rene Vidal. Nonconvex robust low-rank matrix recovery.SIAM Journal on Optimization, 30(1):660–686, 2020

    2020

  47. [55]

    Accelerating ill-conditioned low-rank matrix estimation via scaled gradient descent.J

    Tian Tong, Cong Ma, and Yuejie Chi. Accelerating ill-conditioned low-rank matrix estimation via scaled gradient descent.J. Mach. Learn. Res., 22:150–1, 2021

    2021

  48. [56]

    Beyond procrustes: Balancing-free gradient descent for asymmetric low-rank matrix sensing.IEEE Transactions on Signal Processing, 69:867–877, 2021

    Cong Ma, Yuanxin Li, and Yuejie Chi. Beyond procrustes: Balancing-free gradient descent for asymmetric low-rank matrix sensing.IEEE Transactions on Signal Processing, 69:867–877, 2021

    2021

  49. [57]

    Fast and provable tensor robust principal component analysis via scaled gradient descent.Information and Inference: A Journal of the IMA, 12(3):iaad019, 2023

    Harry Dong, Tian Tong, Cong Ma, and Yuejie Chi. Fast and provable tensor robust principal component analysis via scaled gradient descent.Information and Inference: A Journal of the IMA, 12(3):iaad019, 2023

    2023

  50. [58]

    Tensor completion by alternating minimization under the tensor train (tt) model.arXiv preprint arXiv:1609.05587, 2016

    Wenqi Wang, Vaneet Aggarwal, and Shuchin Aeron. Tensor completion by alternating minimization under the tensor train (tt) model.arXiv preprint arXiv:1609.05587, 2016

    2016 arXiv

  51. [59]

    High-order tensor completion via gradient-based optimization under tensor train format.Signal Processing: Image Communication, 73:53–61, 2019

    Longhao Yuan, Qibin Zhao, Lihua Gui, and Jianting Cao. High-order tensor completion via gradient-based optimization under tensor train format.Signal Processing: Image Communication, 73:53–61, 2019

    2019

  52. [60]

    Low rank tensor recovery via iterative hard threshold- ing.Linear Algebra and its Applications, 523:220–262, 2017

    Holger Rauhut, Reinhold Schneider, and ˇZeljka Stojanac. Low rank tensor recovery via iterative hard threshold- ing.Linear Algebra and its Applications, 523:220–262, 2017

    2017

  53. [61]

    Tensor completion in hierarchical tensor representa- tions

    Holger Rauhut, Reinhold Schneider, and ˇZeljka Stojanac. Tensor completion in hierarchical tensor representa- tions. InCompressed sensing and its applications, pages 419–450. Springer, 2015

    2015

  54. [62]

    Non-convex projected gradient descent for generalized low-rank tensor regression.The Journal of Machine Learning Research, 20(1):172–208, 2019

    Han Chen, Garvesh Raskutti, and Ming Yuan. Non-convex projected gradient descent for generalized low-rank tensor regression.The Journal of Machine Learning Research, 20(1):172–208, 2019

    2019

  55. [63]

    Low-rank tensor completion by riemannian optimization.BIT Numerical Mathematics, 54(2):447–468, 2014

    Daniel Kressner, Michael Steinlechner, and Bart Vandereycken. Low-rank tensor completion by riemannian optimization.BIT Numerical Mathematics, 54(2):447–468, 2014

    2014

  56. [64]

    Tensor train completion: local recovery guarantees via rieman- nian optimization.arXiv preprint arXiv:2110.03975, 2021

    Stanislav Budzinskiy and Nikolai Zamarashkin. Tensor train completion: local recovery guarantees via rieman- nian optimization.arXiv preprint arXiv:2110.03975, 2021

    2021 arXiv

  57. [65]

    Tensor completion using low-rank tensor train decomposition by riemannian optimization

    Junli Wang, Guangshe Zhao, Dingheng Wang, and Guoqi Li. Tensor completion using low-rank tensor train decomposition by riemannian optimization. In2019 Chinese Automation Congress (CAC), pages 3380–3384. IEEE, 2019

    2019

  58. [66]

    Low-rank tensor estimation via riemannian gauss-newton: Statistical optimality and second-order convergence.arXiv preprint arXiv:2104.12031, 2021

    Yuetian Luo and Anru R Zhang. Low-rank tensor estimation via riemannian gauss-newton: Statistical optimality and second-order convergence.arXiv preprint arXiv:2104.12031, 2021

    2021 arXiv

  59. [67]

    Tensor-on-tensor regression: Riemannian optimization, over-parameterization, statistical-computational gap, and their interplay.arXiv preprint arXiv:2206.08756, 2022

    Yuetian Luo and Anru R Zhang. Tensor-on-tensor regression: Riemannian optimization, over-parameterization, statistical-computational gap, and their interplay.arXiv preprint arXiv:2206.08756, 2022

    2022 arXiv

  60. [68]

    Computational and statistical guarantees for tensor-on-tensor regression with tensor train decomposition.arXiv preprint arXiv:2406.06002, 2024

    Zhen Qin and Zhihui Zhu. Computational and statistical guarantees for tensor-on-tensor regression with tensor train decomposition.arXiv preprint arXiv:2406.06002, 2024. 27

    2024 arXiv

  61. [69]

    Princeton University Press, 2008

    P-A Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization algorithms on matrix manifolds. Princeton University Press, 2008

    2008

  62. [70]

    Cambridge University Press, 2017

    Xian-Da Zhang.Matrix analysis and applications. Cambridge University Press, 2017

    2017

  63. [71]

    Solving random quadratic systems of equations is nearly as easy as solving linear systems.Advances in Neural Information Processing Systems, 28, 2015

    Yuxin Chen and Emmanuel Candes. Solving random quadratic systems of equations is nearly as easy as solving linear systems.Advances in Neural Information Processing Systems, 28, 2015

    2015

  64. [72]

    Phase retrieval via wirtinger flow: Theory and algorithms.IEEE Transactions on Information Theory, 61(4):1985–2007, 2015

    Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase retrieval via wirtinger flow: Theory and algorithms.IEEE Transactions on Information Theory, 61(4):1985–2007, 2015

    1985

  65. [73]

    Zhihui Zhu, Tianyu Ding, Daniel Robinson, Manolis Tsakiris, and Ren´e Vidal. A linearly convergent method for non-smooth non-convex optimization on the grassmannian with applications to robust subspace and dictionary learning.Advances in Neural Information Processing Systems, 32, 2019

    2019

  66. [74]

    Nonconvex optimization meets low-rank matrix factorization: An overview.IEEE Transactions on Signal Processing, 67(20):5239–5269, 2019

    Yuejie Chi, Yue M Lu, and Yuxin Chen. Nonconvex optimization meets low-rank matrix factorization: An overview.IEEE Transactions on Signal Processing, 67(20):5239–5269, 2019

    2019

  67. [75]

    A deterministic theory for exact non-convex phase retrieval.IEEE Transac- tions on Signal Processing, 68:4612–4626, 2020

    Bariscan Yonel and Birsen Yazici. A deterministic theory for exact non-convex phase retrieval.IEEE Transac- tions on Signal Processing, 68:4612–4626, 2020

    2020

  68. [76]

    Hierarchical singular value decomposition of tensors.SIAM journal on matrix analysis and applications, 31(4):2029–2054, 2010

    Lars Grasedyck. Hierarchical singular value decomposition of tensors.SIAM journal on matrix analysis and applications, 31(4):2029–2054, 2010

    2010

  69. [77]

    Optimization on the hierarchical tucker manifold–applications to tensor completion.Linear Algebra and its Applications, 481:131–173, 2015

    Curt Da Silva and Felix J Herrmann. Optimization on the hierarchical tucker manifold–applications to tensor completion.Linear Algebra and its Applications, 481:131–173, 2015

    2015

  70. [78]

    Tensor networks for dimensionality reduction and large-scale optimization: Part 1 low-rank tensor decompositions

    Andrzej Cichocki, Namgil Lee, Ivan Oseledets, Anh-Huy Phan, Qibin Zhao, Danilo P Mandic, et al. Tensor networks for dimensionality reduction and large-scale optimization: Part 1 low-rank tensor decompositions. Foundations and Trends® in Machine Learning, 9(4-5):249–429, 2016

    2016

  71. [79]

    Tensor networks for complex quantum systems.Nature Reviews Physics, 1(9):538–550, 2019

    Rom ´an Or´us. Tensor networks for complex quantum systems.Nature Reviews Physics, 1(9):538–550, 2019

    2019

  72. [80]

    Tensor learning for regression.IEEE Transactions on Image Processing, 21(2):816–827, 2011

    Weiwei Guo, Irene Kotsia, and Ioannis Patras. Tensor learning for regression.IEEE Transactions on Image Processing, 21(2):816–827, 2011

    2011

  73. [81]

    Sparse and low-rank tensor estimation via cubic sketchings

    Botao Hao, Anru R Zhang, and Guang Cheng. Sparse and low-rank tensor estimation via cubic sketchings. In International Conference on Artificial Intelligence and Statistics, pages 1319–1330. PMLR, 2020

    2020

  74. [82]

    Tensor regression with applications in neuroimaging data analysis.Journal of the American Statistical Association, 108(502):540–552, 2013

    Hua Zhou, Lexin Li, and Hongtu Zhu. Tensor regression with applications in neuroimaging data analysis.Journal of the American Statistical Association, 108(502):540–552, 2013

    2013

  75. [83]

    Compressed sensing.IEEE Transactions on information theory, 52(4):1289–1306, 2006

    David L Donoho. Compressed sensing.IEEE Transactions on information theory, 52(4):1289–1306, 2006

    2006

  76. [84]

    Robust uncertainty principles: Exact signal reconstruc- tion from highly incomplete frequency information.IEEE Transactions on information theory, 52(2):489–509, 2006

    Emmanuel J Cand `es, Justin Romberg, and Terence Tao. Robust uncertainty principles: Exact signal reconstruc- tion from highly incomplete frequency information.IEEE Transactions on information theory, 52(2):489–509, 2006

    2006

  77. [85]

    An introduction to compressive sampling.IEEE signal processing magazine, 25(2):21–30, 2008

    Emmanuel J Cand `es and Michael B Wakin. An introduction to compressive sampling.IEEE signal processing magazine, 25(2):21–30, 2008

    2008

  78. [86]

    Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization.SIAM review, 52(3):471–501, 2010

    Benjamin Recht, Maryam Fazel, and Pablo A Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization.SIAM review, 52(3):471–501, 2010

    2010

  79. [87]

    Iterative hard thresholding for low cp-rank tensor models.Linear and Multilinear Algebra, pages 1–17, 2021

    Rachel Grotheer, Shuang Li, Anna Ma, Deanna Needell, and Jing Qin. Iterative hard thresholding for low cp-rank tensor models.Linear and Multilinear Algebra, pages 1–17, 2021

    2021

  80. [88]

    Modulus of continuity of some conditionally sub-gaussian fields, applica- tion to stable random fields

    Hermine Bierm ´e and C´eline Lacaux. Modulus of continuity of some conditionally sub-gaussian fields, applica- tion to stable random fields. 2015

    2015

  81. [89]

    Islet: Fast and optimal low-rank tensor regression via importance sketching.SIAM journal on mathematics of data science, 2(2):444–479, 2020

    Anru R Zhang, Yuetian Luo, Garvesh Raskutti, and Ming Yuan. Islet: Fast and optimal low-rank tensor regression via importance sketching.SIAM journal on mathematics of data science, 2(2):444–479, 2020. 28

    2020

  82. [90]

    A multilinear singular value decomposition.SIAM journal on Matrix Analysis and Applications, 21(4):1253–1278, 2000

    Lieven De Lathauwer, Bart De Moor, and Joos Vandewalle. A multilinear singular value decomposition.SIAM journal on Matrix Analysis and Applications, 21(4):1253–1278, 2000

    2000

  83. [91]

    Weakly convex optimization over Stiefel manifold using Riemannian subgradient-type methods.SIAM Journal on Optimization, 31(3):1605–1634, 2021

    Xiao Li, Shixiang Chen, Zengde Deng, Qing Qu, Zhihui Zhu, and Anthony Man-Cho So. Weakly convex optimization over Stiefel manifold using Riemannian subgradient-type methods.SIAM Journal on Optimization, 31(3):1605–1634, 2021

    2021

  84. [92]

    Springer, 2012

    Wolfgang Hackbusch.Tensor spaces and numerical tensor calculus, volume 42. Springer, 2012. 29

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.