Pith. sign in

REVIEW 2 major objections 3 minor 49 references

This paper claims that tensor ring (TR) decomposition admits a deterministic finite-step exact algorithm, BLOSTR, which recovers all TR cores from O(r²(n₁+⋯+n_d)) observed entries under a genericity condition.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 19:22 UTC pith:PRPRCWY4

load-bearing objection Genuinely promising TR decomposition algorithm, but Lemma 1 as stated is not proven: the proof silently cancels B_α and B_β, which only works when Γ_α=Γ_β; the likely fix (same Γ per spectral probe) is simple but absent from the paper. the 2 major comments →

arxiv 2512.01016 v2 pith:PRPRCWY4 submitted 2025-11-30 math.NA cs.NA

A Provably Efficient Method for Tensor Ring Decomposition and Its Applications

Han Chen , Sitan Chen , Anru R. Zhang This is my paper
classification math.NA cs.NA MSC 15A6965F99
keywords tensor ring decompositionsimultaneous diagonalizationtensor networksfinite-step algorithmsample complexitysymmetric tensor decompositionmatrix product statesmethod of moments
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims to settle a long-open question: tensor ring (TR) decomposition — expressing a d-way tensor as a trace of a product of d cores — admits a deterministic, finite-step exact algorithm. The method, blockwise simultaneous diagonalization, recovers all TR cores up to the model's inherent gauge freedom from about r²(n₁+⋯+n_d) observed entries, the same order as the number of free parameters. A sympathetic reader would care because previous TR algorithms were iterative or heuristic, and this brings TR in line with Tucker and tensor-train formats, which already had finite-step constructions. The paper also extends the construction to symmetric TR tensors, couples it with alternating least squares for noisy data, and derives consequences for matrix product state tomography and for learning distributions with quadratic cyclic transformations.

Core claim

BLOSTR treats a TR tensor as glued together by products R_α = Q₂^{(α₂)}⋯Q_{d-1}^{(α_{d-1})}. For two middle indices α and β, the matrix T(:,α,Γ_α) T(:,β,Γ_β)^† has nonzero eigenvalues equal to those of R_α(R_β)⁻¹, each repeated r times, and its eigenvectors carry the first core Q₁ up to a block-diagonal gauge. By taking two such spectral probes, aligning their eigenbases, fixing the gauge through one block, and then using circular mode shifts to peel off the remaining cores, the algorithm reconstructs all cores in a fixed number of algebraic steps. Theorem 1 states that when each mode size is at least r² and the core entries are drawn from any measure absolutely continuous with respect to Le

What carries the argument

The central object is the blockwise simultaneous diagonalization of two 'spectral probes': for a fixed pair of middle indices (α,β), the ratio matrix R_α(R_β)⁻¹ is read off as the repeated eigenvalue pattern of T(:,α,Γ_α) T(:,β,Γ_β)^†. Taking a second probe (α′,β′) and forming F = E†E′ rewrites the unknown gauge as an r×r block-structured matrix with entries (U⁻¹V)_{j,k} K_j⁻¹ K′_k, so fixing one block and normalizing the first row and column of U⁻¹V determines Q₁; circular mode permutation then yields the rest. The whole argument rides on the genericity that these ratios are diagonalizable with distinct eigenvalues and that the gauge-fixing entries do not vanish.

Load-bearing premise

The whole construction assumes a random-genericity condition: every intermediate product must be invertible, every ratio of two such products must have all distinct eigenvalues, and certain diagonal entries used to fix the gauge must be nonzero; if any of these fails, the spectral probes cannot be aligned and the algorithm has no stated fallback.

What would settle it

Draw many random TR tensors with n_k = r² and run Algorithm 1 in high-precision arithmetic; the theorem predicts failure probability zero, so any single draw that fails exact recovery to numerical precision — or any measurable family of continuous core distributions on which the algorithm fails with positive probability — would falsify the probability-one claim. A targeted probe is to set the middle core slices so that R_α(R_β)⁻¹ has a repeated eigenvalue for one pair (α,β), which lies outside the theorem's event but is still a valid input; observing failure there confirms the guarantee is dis

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Exact TR decomposition is achievable in finitely many steps, so under the stated genericity and dimension conditions, iterative TR-ALS is no longer needed for exact recovery from clean sparse samples.
  • The exact recovery guarantee uses O((n₁+⋯+n_d) r²) entries, which is order-optimal because it matches the number of free TR parameters up to gauge.
  • Symmetric TR decomposition reduces parameter complexity from O(d n r²) to O(n r²), and the paper gives a finite-step solver that observes only O(n r²) entries.
  • Noisy recovery can be initialized by the spectral-block procedure and refined by alternating least squares; the numerical experiments show faster convergence and lower error than random-initialized ALS.
  • For matrix product states with bond dimension r and n_k ≥ r², the paper shows recovery from a constant number of 3-body marginals for generic MPS, up to gauge, and applies the same machinery to learn cyclic quadratic transformations from moment tensors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because exactness uses only a handful of fibers, a regularized or averaged version could become a practical tensor-ring completion primitive for large high-order tensors — but only if the spectral gap between the r repeated eigenvalue groups exceeds the noise level, which the paper does not quantify.
  • Editorial inference: the distinct-eigenvalue/genericity condition is the practical bottleneck; a testable prediction is that exact recovery degrades smoothly as the eigenvalue clusters of R_α(R_β)⁻¹ merge, and joint-diagonalization variants might extend the method to degenerate spectra.
  • Editorial inference: the symmetric-TR root-of-unity ambiguity in Algorithm 2 (Lemma 5) means the recovered core is identified only up to discrete phase choices; this is inherent to the model but should be made explicit when the core is used to interpret exchangeable data.
  • The paper's own remarks flag that quantitative noise bounds and MPS sample complexity are deferred (Remark 4) and that fully heterogeneous TR ranks remain open; a reader should not read the noiseless exactness theorem as a statistical guarantee.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper proposes BLOSTR, a closed-form, finite-step algorithm for exact tensor ring decomposition from a sparse set of observed entries, under the conditions n_k ≥ r^2 and a genericity assumption on the cores. The method is based on forming contractions T(:,α,Γα)T(:,β,Γβ)^† and extracting the TR cores from their block-simultaneous-diagonalization structure. The paper further gives a symmetric variant (BLOSTR-S), a noise-robust variant coupled with ALS, applications to MPS tomography and to learning cyclic quadratic transformations, and numerical experiments demonstrating machine-precision recovery in exact cases and improved performance over random-initialized ALS in noisy cases.

Significance. If the main theorem were correct, this would be a substantial advance: it would give the first finite-step exact TR decomposition algorithm with near-optimal sample complexity O((∑ n_k)r^2), matching the parameter count up to constants. The algebraic strategy is elegant and the numerical tables are consistent with exact recovery for the tested settings. However, the central lemma on which Theorem 1 rests contains a false cancellation of different column-subset factors, so the main claim is not established as written. The issue appears locally repairable, but it is load-bearing and must be fixed before the paper can be considered for publication.

major comments (2)
  1. Lemma 1 as stated allows Γα and Γβ to be two different subsets of [n_d], but the proof in Appendix B cancels Q_d[1]^*(:,Γα) against Q_d[1]^*(:,Γβ). Writing Bγ = Q_d[1]^*(:,Γγ), the exact product is T(:,α,Γα)T(:,β,Γβ)^† = Q1⟨1⟩(I⊗Rα) Bα Bβ^{-1} (I⊗Rβ^{-1}) Q1⟨1⟩^†. If Γα ≠ Γβ, the factor BαBβ^{-1} is a generic invertible r^2×r^2 matrix and does not cancel; the nonzero spectrum is not generally {λ_1,...,λ_r} each with multiplicity r. For instance, already for r=2 one can take RαRβ^{-1}=diag(2,3) and generic C=BαBβ^{-1}; the eigenvalues of (I⊗Rα)C(I⊗Rβ^{-1}) are not 2,2,3,3. Thus Lemma 1 is false as stated, and Steps 1–3 of Algorithm 1 together with Theorem 1 are unproven. The proof would be valid if the statement required a common subset Γα=Γβ=Γ for the two factors in each spectral probe, and Algorithm 1's notation should be changed accordingly; this local fix would preserve the claimed sa
  2. The symmetric algorithm's loop 'repeat ... until Q satisfies (15)' searches over d^{r-1} choices of roots of unity. The paper does not analyze the cost of this search or give any indication that it is efficient. Since the title and abstract emphasize a 'provably efficient' method, this is a gap in the symmetric-extensions section. A bound on the expected number of iterations, or an alternative deterministic choice, should be provided.
minor comments (3)
  1. The text states that BLOSTR only needs to inspect O(d max_k n_k^2) entries, which is inconsistent with Theorem 1's O((∑ n_k) r^2). This appears to be a typo; it should be O(d max_k n_k r^2) or similar.
  2. The paper repeatedly calls the algorithm 'deterministic' while the assumptions on the cores are probabilistic (absolutely continuous measure). This is standard, but the distinction between a deterministic algorithm and a probabilistic identifiability/genericity assumption should be stated more carefully.
  3. The construction of M_{j,k}^{h,v} is notationally dense and the proof of the reshaping step is only sketched. A more explicit derivation would improve readability, especially since Theorem 3 relies on this construction.

Circularity Check

0 steps flagged

No significant circularity: BLOSTR's derivation is a self-contained algebraic inversion; flagged issues are proof-level gaps or non-load-bearing context.

full rationale

The central derivation (Theorem 1 via Lemma 1 and Steps 1–4, Section 3) is self-contained: the recovered cores are obtained by closed-form spectral algebra from the observed matrices T(:,α,Γ_α)T(:,β,Γ_β)†, and the theorem's conclusion (recovery up to gauge (5)) is not an input to the algorithm. No parameter of the conclusion is defined in terms of the observed entries; the genericity premises (absolute continuity of the core distribution, distinct spectra, invertibility of R_α and Q_{d[1]}^*(:,Γ)) are distributional conditions on the input, not fitted values. The sample-complexity claim O((Σn_k)r²) is matched to a free-parameter count, not to a fitted constant. The one overlapping-author citation [8] (two of three present authors overlap; STOC 2023) appears only as motivation in Section 7.2 ('Building upon an observation from [8]') and Remark 6; the moment-to-TR identity needed there is proved in-paper as Lemma 7 ('whose proof is immediate from Wick calculus'), so the self-citation is not load-bearing and, being a published, externally checkable result, does not raise the score. The applications (MPS recovery, Section 7.1; pushforward learning, Section 7.2) reuse the same BLOSTR probes on marginals or moments — they are applications of the method, not circular validations. The paper honestly flags its own limitations: Remark 4 defers quantitative MPS sample complexity; Remark 5 concedes worst-case states are unrecoverable from local measurements; Remark 6 leaves the symmetric-matrix variant open; Remark 2 leaves unequal ranks open. What the reviewer should weigh as correctness risk, not circularity: (a) Lemma 1 states Γ_α ≠ Γ_β, but its proof (Appendix B) writes both factors with the same subset Γ_d and cancels Q_{d[1]}^*(:,Γ_d) with its own inverse, silently omitting B_αB_β^{-1}; for genuinely distinct subsets the asserted eigenvalue identity need not hold — a proof gap in the current arXiv version; (b) Algorithm 2's termination check verifies (15) for all n^d tuples while the stated observation budget is O(nr²); (c) Algorithm 2's update Q̄^{(j)} ← (Q̄^{(1)})^{-(d-1)} Q̂^{(1)}_1 Q̂^{(j)}_d omits the middle factors Q̂^{(1)}_2…Q̂^{(1)}_{d-1} that (15) requires. None of these reduce the paper's conclusions to its inputs by construction; they are mathematical-support issues, so under the seven circularity patterns the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 1 invented entities

The central claim rests on (a) a genericity assumption on the cores that is stated but not verified on real data, (b) the n_k ≥ r² dimension condition, (c) known rank r, and (d) exact-access assumptions in the applications that are explicitly deferred to future work. No free parameters are fitted to data; r is an input, and the algorithm's steps are parameter-free. The invented entity is just the cyclic quadratic transformation model, which has no independent falsifiable handle.

axioms (6)
  • domain assumption The TR cores Q_k are drawn from a measure absolutely continuous w.r.t. the standard Euclidean measure (mathematical genericity)
    Theorem 1, Lemma 1: licenses 'with probability one' invertibility of R_α, distinct eigenvalues of R_α(R_β)⁻¹, full column rank of Q₁⟨1⟩, and nonzero gauge-fixing entries. Load-bearing: degenerate cores are outside the guarantee.
  • domain assumption Dimension condition n_k ≥ r² for all k ∈ [d]
    Theorem 1; required so the n_k×r² fiber probes read in (13) have full rank and eigenbases carry the needed block structure. Remark 1 relaxes to some k via contraction (Appendix C) at higher sample cost, but the relaxed version still needs the product of the remaining modes ≥ r².
  • domain assumption The TR rank r is known a priori and equal across cores
    Section 2 (final remark) and Remark 2: rank estimation deferred to [47]; mixed-rank rings left open.
  • domain assumption For Theorem 3, exact access to O(d) 3-body marginals of the MPS (and, implicitly, that the marginals identify the τ terms in (26))
    Section 7.1, Lemma 6, Remark 4: the quantum access model is replaced by exact marginal entries; the stochastic measurement error is explicitly deferred.
  • standard math Standard Moore–Penrose identities (Lemma S1) and circular-shift invariance of TR (Lemma 2, cited from [47, Thm 2.1])
    Used throughout Appendix A for the pseudoinverse manipulations in Steps 3–4.
  • standard math Wick calculus: Gaussian moments of degree d in cyclic quadratic forms equal the TR-trace (Lemma 7)
    Section 7.2: E[z₁...z_d] = tr[Q₁^(α₁)...Q_d^(α_d)]; standard Isserlis-type computation, stated as immediate.
invented entities (1)
  • Cyclic quadratic transformation (generative model in Definition 4) no independent evidence
    purpose: Connects the TR decomposition to a learnable distribution family: moments are exactly TR tensors, so BLOSTR can parameter-learn from moment estimates.
    A new model definition, not a falsifiable physical entity; used to motivate the pushforward-learning application. No prediction outside the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 30217 in / 44124 out tokens · 393556 ms · 2026-08-03T19:22:01.517197+00:00 · methodology

0 comments
read the original abstract

We present the first deterministic, finite-step algorithm for exact tensor ring (TR) decomposition. Our method leverages blockwise simultaneous diagonalization to recover TR cores from a limited number of tensor observations, under a dimension condition requiring each mode size to be at least quadratic in the TR rank and under a genericity assumption on the cores, thereby providing both algebraic insight and practical efficiency. We extend the approach to the symmetric TR setting, where parameter complexity is significantly reduced and applications arise naturally in physics-based modeling and exchangeable data analysis. To handle noisy observations, we develop a robust recovery scheme that couples our initialization with alternating least squares, achieving faster convergence and improved accuracy compared to classical methods. As applications, we obtain new algorithms for questions in other domains where tensor ring decomposition is a key primitive, namely matrix product state tomography in quantum information and provable learning of pushforward distributions in the foundations of machine learning. These contributions advance the algorithmic foundations of TR decomposition and open new opportunities for scalable tensor network computation.

Figures

Figures reproduced from arXiv: 2512.01016 by Anru R. Zhang, Han Chen, Sitan Chen.

Figure 1
Figure 1. Figure 1: Illustration of Π2,3 operated on a 6-by-6 matrix X. Elements in the same sub-matrices are labeled with same colors. Order-2 TR Decomposition. It is helpful to first explore the TR decomposition for order-2 tensors, i.e., matrices, before diving into more general scenarios. Suppose we observe a matrix T ∈ C n1×n2 that admits a TR decomposition: T(α1, α2) = tr n Q (α1) 1 Q (α2) 2 o = Q1⟨1⟩(α1, :) Q2[1](α2, :… view at source ↗
Figure 2
Figure 2. Figure 2: Illustrative example of a possible choice of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Average relative error (with standard deviation in log scale) of [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average relative error (with standard deviation in log scale) of [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Proportion of successful recoveries (relative error below [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 5 linked inside Pith

  1. [1]

    Unsupervised multiway data analysis: A literature survey.IEEE transactions on knowledge and data engineering, 21(1):6–20, 2008

    Evrim Acar and Bülent Yener. Unsupervised multiway data analysis: A literature survey.IEEE transactions on knowledge and data engineering, 21(1):6–20, 2008

    2008

  2. [2]

    A survey on the complexity of learning quantum states

    Anurag Anshu and Srinivasan Arunachalam. A survey on the complexity of learning quantum states. Nature Reviews Physics, 6(1):59–69, 2024

    2024

  3. [3]

    Rigorous rg algorithms and area laws for low energy eigenstates in 1d.Communications in Mathematical Physics, 356(1):65–105, 2017

    Itai Arad, Zeph Landau, Umesh Vazirani, and Thomas Vidick. Rigorous rg algorithms and area laws for low energy eigenstates in 1d.Communications in Mathematical Physics, 356(1):65–105, 2017

    2017

  4. [4]

    Springer, 2003

    Adi Ben-Israel and Thomas NE Greville.Generalized inverses: theory and applications. Springer, 2003

    2003

  5. [5]

    Constrained k-means clustering.Microsoft Research, Redmond, 20(0):0, 2000

    Paul S Bradley, Kristin P Bennett, and Ayhan Demiriz. Constrained k-means clustering.Microsoft Research, Redmond, 20(0):0, 2000. 17

    2000

  6. [6]

    Hand-waving and interpretive dance: an introductory course on tensor networks.Journal of physics A: Mathematical and theoretical, 50(22):223001, 2017

    Jacob C Bridgeman and Christopher T Chubb. Hand-waving and interpretive dance: an introductory course on tensor networks.Journal of physics A: Mathematical and theoretical, 50(22):223001, 2017

    2017

  7. [7]

    eckart-young

    J Douglas Carroll and Jih-Jie Chang. Analysis of individual differences in multidimensional scaling via an n-way generalization of “eckart-young” decomposition.Psychometrika, 35(3):283–319, 1970

    1970

  8. [8]

    Learning polynomial transformations via gen- eralized tensor decompositions

    Sitan Chen, Jerry Li, Yuanzhi Li, and Anru R Zhang. Learning polynomial transformations via gen- eralized tensor decompositions. InProceedings of the 55th Annual ACM Symposium on Theory of Computing, pages 1671–1684, 2023

    2023

  9. [9]

    Tensor decompositions for signal processing applications: From two-way to multiway component analysis.IEEE Signal Processing Magazine, 32(2):145–163, 2015

    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

  10. [10]

    Decomposition of quantics in sums of powers of linear forms

    Pierre Comon and Bernard Mourrain. Decomposition of quantics in sums of powers of linear forms. Signal Processing, 53(2-3):93–107, 1996

    1996

  11. [11]

    Tensor decomposition of eeg signals: a brief review.Journal of neuroscience methods, 248:59–69, 2015

    Fengyu Cong, Qiu-Hua Lin, Li-Dan Kuang, Xiao-Feng Gong, Piia Astikainen, and Tapani Ristaniemi. Tensor decomposition of eeg signals: a brief review.Journal of neuroscience methods, 248:59–69, 2015

    2015

  12. [12]

    Springer, 4th edition, 2015

    David Cox, John Little, and Donal O’Shea.Ideals, Varieties, and Algorithms: An Introduction to Computational Algebraic Geometry and Commutative Algebra. Springer, 4th edition, 2015

    2015

  13. [13]

    Efficient quantum state tomogra- phy.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 tomogra- phy.Nature communications, 1(1):149, 2010

    2010

  14. [14]

    A survey of tensor methods

    Lieven De Lathauwer. A survey of tensor methods. In2009 IEEE international symposium on circuits and systems, pages 2773–2776. IEEE, 2009

    2009

  15. [15]

    On the best rank-1 and rank-(r 1, r 2,..., rn) approximation of higher-order tensors.SIAM journal on Matrix Analysis and Applications, 21(4): 1324–1342, 2000

    Lieven De Lathauwer, Bart De Moor, and Joos Vandewalle. On the best rank-1 and rank-(r 1, r 2,..., rn) approximation of higher-order tensors.SIAM journal on Matrix Analysis and Applications, 21(4): 1324–1342, 2000

    2000

  16. [16]

    A multilinear singular value decomposition

    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

  17. [17]

    Computing large-scale matrix and tensor decomposition with structured factors: A unified nonconvex optimization perspective

    Xiao Fu, Nico Vervliet, Lieven De Lathauwer, Kejun Huang, and Nicolas Gillis. Computing large-scale matrix and tensor decomposition with structured factors: A unified nonconvex optimization perspective. IEEE Signal Processing Magazine, 37(5):78–94, 2020

    2020

  18. [18]

    Harshman

    Richard A. Harshman. Foundations of the parafac procedure: Models and conditions for an explanatory multimodal factor analysis. Technical Report 16, UCLA Working Papers in Phonetics, 1970. (University Microfilms, Ann Arbor, Michigan, No. 10,085)

    1970

  19. [19]

    A survey on tensor techniques and applications in machine learning.IEEE Access, 7:162950–162990, 2019

    Yuwang Ji, Qiang Wang, Xuan Li, and Jie Liu. A survey on tensor techniques and applications in machine learning.IEEE Access, 7:162950–162990, 2019

    2019

  20. [20]

    Efficient construction of tensor ring representations from sampling.Multiscale Modeling & Simulation, 19(3):1261–1284, 2021

    Yuehaw Khoo, Jianfeng Lu, and Lexing Ying. Efficient construction of tensor ring representations from sampling.Multiscale Modeling & Simulation, 19(3):1261–1284, 2021

    2021

  21. [21]

    Orthogonal tensor decompositions.SIAM Journal on Matrix Analysis and Applica- tions, 23(1):243–255, 2001

    Tamara G Kolda. Orthogonal tensor decompositions.SIAM Journal on Matrix Analysis and Applica- tions, 23(1):243–255, 2001

    2001

  22. [22]

    Tensor decompositions and applications.SIAM review, 51(3): 455–500, 2009

    Tamara G Kolda and Brett W Bader. Tensor decompositions and applications.SIAM review, 51(3): 455–500, 2009

    2009

  23. [23]

    Tensor regression networks.Journal of Machine Learning Research, 21(123):1–21, 2020

    Jean Kossaifi, Zachary C Lipton, Arinbjorn Kolbeinsson, Aran Khanna, Tommaso Furlanello, and Anima Anandkumar. Tensor regression networks.Journal of Machine Learning Research, 21(123):1–21, 2020. 18

    2020

  24. [24]

    A polynomial time algorithm for the ground state of one-dimensional gapped local hamiltonians.Nature Physics, 11(7):566–569, 2015

    Zeph Landau, Umesh Vazirani, and Thomas Vidick. A polynomial time algorithm for the ground state of one-dimensional gapped local hamiltonians.Nature Physics, 11(7):566–569, 2015

    2015

  25. [25]

    Efficient direct tomography for matrix product states.arXiv preprint arXiv:1002.4632, 2010

    Olivier Landon-Cardinal, Yi-Kai Liu, and David Poulin. Efficient direct tomography for matrix product states.arXiv preprint arXiv:1002.4632, 2010

    Pith/arXiv arXiv 2010

  26. [26]

    A decomposition for three-way arrays.SIAM Journal on Matrix Analysis and Applications, 14(4):1064–1083, 1993

    Sue E Leurgans, Robert T Ross, and Rebecca B Abel. A decomposition for three-way arrays.SIAM Journal on Matrix Analysis and Applications, 14(4):1064–1083, 1993

    1993

  27. [27]

    k-means-constrained, April 2018

    Josh Levy-Kramer. k-means-constrained, April 2018. URLhttps://github.com/joshlk/ k-means-constrained

    2018

  28. [28]

    Characterizing spatiotemporal transcriptome of the human brain via low-rank tensor decomposition.Statistics in Biosciences, 14(3):485–513, 2022

    Tianqi Liu, Ming Yuan, and Hongyu Zhao. Characterizing spatiotemporal transcriptome of the human brain via low-rank tensor decomposition.Statistics in Biosciences, 14(3):485–513, 2022

    2022

  29. [29]

    Symmetric quantum states: a review of recent progress.arXiv preprint arXiv:2506.10185, 2025

    Carlo Marconi, Guillem Müller-Rigat, Jordi Romero-Pallejà, Jordi Tura, and Anna Sanpera. Symmetric quantum states: a review of recent progress.arXiv preprint arXiv:2506.10185, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    The complexity of the word problems for commutative semigroups and polynomial ideals.Advances in mathematics, 46(3):305–329, 1982

    Ernst W Mayr and Albert R Meyer. The complexity of the word problems for commutative semigroups and polynomial ideals.Advances in mathematics, 46(3):305–329, 1982

    1982

  31. [31]

    Parallel factor analysis as an exploratory tool for wavelet transformed event-related eeg.NeuroImage, 29(3): 938–947, 2006

    Morten Mørup, Lars Kai Hansen, Christoph S Herrmann, Josef Parnas, and Sidse M Arnfred. Parallel factor analysis as an exploratory tool for wavelet transformed event-related eeg.NeuroImage, 29(3): 938–947, 2006

    2006

  32. [32]

    Cambridge university press, 2010

    Michael A Nielsen and Isaac L Chuang.Quantum computation and quantum information. Cambridge university press, 2010

    2010

  33. [33]

    A practical introduction to tensor networks: Matrix product states and projected entan- gled pair states.Annals of physics, 349:117–158, 2014

    Román Orús. A practical introduction to tensor networks: Matrix product states and projected entan- gled pair states.Annals of physics, 349:117–158, 2014

    2014

  34. [34]

    Tensor-traindecomposition.SIAM Journal on Scientific Computing, 33(5):2295–2317, 2011

    Ivan V Oseledets. Tensor-traindecomposition.SIAM Journal on Scientific Computing, 33(5):2295–2317, 2011

    2011

  35. [35]

    Tensor methods in computer vision and deep learning.Proceedings of the IEEE, 109(5):863–890, 2021

    Yannis Panagakis, Jean Kossaifi, Grigorios G Chrysos, James Oldfield, Mihalis A Nicolaou, Anima Anandkumar, and Stefanos Zafeiriou. Tensor methods in computer vision and deep learning.Proceedings of the IEEE, 109(5):863–890, 2021

    2021

  36. [36]

    Tensors for data mining and data fusion: Models, applications, and scalable algorithms.ACM Transactions on Intelligent Systems and Technology (TIST), 8(2):1–44, 2016

    Evangelos E Papalexakis, Christos Faloutsos, and Nicholas D Sidiropoulos. Tensors for data mining and data fusion: Models, applications, and scalable algorithms.ACM Transactions on Intelligent Systems and Technology (TIST), 8(2):1–44, 2016

    2016

  37. [37]

    Introduction to tensor decompositions and their applications in machine learning.arXiv preprint arXiv:1711.10781, 2017

    Stephan Rabanser, Oleksandr Shchur, and Stephan Günnemann. Introduction to tensor decompositions and their applications in machine learning.arXiv preprint arXiv:1711.10781, 2017

    Pith/arXiv arXiv 2017

  38. [38]

    Thedensity-matrixrenormalization group intheageofmatrix product states.Annals of physics, 326(1):96–192, 2011

    UlrichSchollwöck. Thedensity-matrixrenormalization group intheageofmatrix product states.Annals of physics, 326(1):96–192, 2011

    2011

  39. [39]

    Tensor methods in biomedical image analysis.Journal of Medical Signals & Sensors, 14(6):16, 2024

    Farnaz Sedighin. Tensor methods in biomedical image analysis.Journal of Medical Signals & Sensors, 14(6):16, 2024

    2024

  40. [40]

    Tensordecompositionforsignalprocessingandmachinelearning.IEEE Transactions on Signal Processing, 65(13):3551–3582, 2017

    Nicholas D Sidiropoulos, Lieven De Lathauwer, Xiao Fu, Kejun Huang, Evangelos E Papalexakis, and ChristosFaloutsos. Tensordecompositionforsignalprocessingandmachinelearning.IEEE Transactions on Signal Processing, 65(13):3551–3582, 2017

    2017

  41. [41]

    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

  42. [42]

    Matrix product states, projected entangled pair states, and variational renormalization group methods for quantum spin systems.Advances in physics, 57(2):143–224, 2008

    Frank Verstraete, Valentin Murg, and J Ignacio Cirac. Matrix product states, projected entangled pair states, and variational renormalization group methods for quantum spin systems.Advances in physics, 57(2):143–224, 2008. 19

    2008

  43. [43]

    Efficient low rank tensor ring completion

    Wenqi Wang, Vaneet Aggarwal, and Shuchin Aeron. Efficient low rank tensor ring completion. In Proceedings of the IEEE International Conference on Computer Vision, pages 5697–5705, 2017

    2017

  44. [44]

    Tensor network ranks.arXiv preprint arXiv:1801.02662, 2018

    Ke Ye and Lek-Heng Lim. Tensor network ranks.arXiv preprint arXiv:1801.02662, 2018

    Pith/arXiv arXiv 2018

  45. [45]

    Tensor ring decomposition with rank minimization on latent space: An efficient approach for tensor completion

    Longhao Yuan, Chao Li, Danilo Mandic, Jianting Cao, and Qibin Zhao. Tensor ring decomposition with rank minimization on latent space: An efficient approach for tensor completion. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9151–9158, 2019

    2019

  46. [46]

    Tensor svd: Statistical and computational limits.IEEE Transactions on Information Theory, 64(11):7311–7338, 2018

    Anru Zhang and Dong Xia. Tensor svd: Statistical and computational limits.IEEE Transactions on Information Theory, 64(11):7311–7338, 2018

    2018

  47. [47]

    Tensor ring decomposition

    Qibin Zhao, Guoxu Zhou, Shengli Xie, Liqing Zhang, and Andrzej Cichocki. Tensor ring decomposition. arXiv preprint arXiv:1606.05535, 2016

    Pith/arXiv arXiv 2016

  48. [48]

    Learning with hypergraphs: Clustering, classification, and embedding.Advances in neural information processing systems, 19, 2006

    Dengyong Zhou, Jiayuan Huang, and Bernhard Schölkopf. Learning with hypergraphs: Clustering, classification, and embedding.Advances in neural information processing systems, 19, 2006

    2006

  49. [49]

    Optimal high-order tensor svd via tensor- train orthogonal iteration.IEEE transactions on information theory, 68(6):3991–4019, 2022

    Yuchen Zhou, Anru R Zhang, Lili Zheng, and Yazhen Wang. Optimal high-order tensor svd via tensor- train orthogonal iteration.IEEE transactions on information theory, 68(6):3991–4019, 2022. 20 Appendices A Proofs of Theorems A.1 Auxiliary Lemmas The first lemma is a classic conclusion on the Moore-Penrose inverse [4]. Lemma S1.SupposeAandBare two matrices ...

    2022