REVIEW 3 major objections 4 minor 88 references
Efficient Tensor Decomposition via Moment Matrix Extension
T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read The paper shows that the moment matrix extension algorithm decomposes generic order-4 symmetric tensors of rank up to 2n+1 efficiently, reduces the problem to a linear system, and conjectures the same up to rank O(n^2).
desk verdict Solid new rank thresholds for generic order-4 tensor decomposition via linear systems; the monomial parameterization is genuinely weaker and should be flagged as conditional. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The machinery is the moment matrix (Hankel matrix) of a tensor extended by unknown moment variables, together with determinantal relations that force the extended tensor to admit a rank-$s$ decomposition; equivalently, commutation of multiplication matrices. The monomial set $B$ selects an invertible submatrix, and its degree bounds the regularity of the decomposition. The decisive object is the coefficient matrix $A$ of the linear determinantal relations: when $A$ is full column rank, the moment variables are determined by a linear system, an eigendecomposition recovers the points, and a second linear system recovers the coefficients. Regularity $\rho(Z)$, the first $k$ for which the $k$-th Veronese images of the decomposition points are linearly independent, controls whether the relations are linear rather than quadratic.
What would settle it
Take $n = 18$ and the conjectured rank bound $r$ selected by the condition $|E_1| \ge |Y|$, draw $r$ random points over a finite field $\mathbb{F}_p$, form $\varphi = \sum_i z_i^{\otimes 4}$, and compute the rank of the coefficient matrix $A$ of linear determinantal relations. If $\mathrm{rank}(A) < |Y|$ for any such generic specialization, the conjecture's sufficiency claim is false; the same computation for other $n$ and $r > 2n+1$ would locate exactly where the linear system ceases to be generically full rank.
Extended reading notes
Core claim
The central claim is that $(n,r)$ is an efficient format for $n \ge 3$ and $r \le 2n+1$: for a generic tensor in $S^4\mathbb{C}^{n+1}$ of rank $r$, the coefficient matrix $A$ formed from the linear determinantal relations of moment matrix extension is generically full column rank. Consequently Algorithm 4 decomposes such tensors in polynomially many linear-algebra operations, and $A$ having full column rank provides an effective criterion for specific identifiability (Theorem 4.17). For monomials in canonical form, the paper establishes that the free parameters in the extension are in bijection with the dimension of the decomposition space, yielding an explicit parameterization of every minimal Waring decomposition. The regularity of a decomposition is identified as the key complexity measure throughout.
Load-bearing premise
The load-bearing premise is that the dimension of the monomial decomposition space is exactly the Hilbert-function value from [12], and that all determinantal equations defining a monomial extension are equivalent to the graded relations used in Algorithm 5; the paper explicitly leaves the second equivalence to future study and provides no standalone proof of the first.
Editorial extensions
If this is right
- Generic order-4 symmetric tensors of rank at most $2n+1$ can be decomposed in polynomial time using only linear algebra, going beyond the simultaneous diagonalization threshold.
- For a specific tensor satisfying the format conditions, full column rank of $A$ is a checkable certificate that its decomposition is unique.
- If Conjecture 1 holds, efficient decomposition extends to generic tensors of rank $O(n^2)$, with leading coefficient greater than $0.432$; the paper verifies the efficient formats for $n = 2,\dots,17$ via finite-field specializations.
- Nonidentifiable tensors, including order-4 tensors arising from three collinear points and arbitrary monomials, are efficiently decomposable, and monomial decompositions are explicitly parameterized.
Reading between the lines
- If regularity, not uniqueness, is the true complexity driver, similar speedups may apply to other extension-based decomposition algorithms by tracking regularity instead of rank.
- The checkable full-column-rank criterion for $A$ could serve as a practical uniqueness certificate or stopping rule in numerical tensor pipelines, since it is evaluated by ordinary linear algebra on a given tensor.
- The monomial parameterization suggests that other structured sparse polynomials may admit free-parameter decomposition families of the same kind, potentially reducing them to linear algebra as well.
- Because the paper found that some second-type determinantal equations are inherently dependent, a refined equation selection or a different flattening could push the conjectured rank threshold beyond the stated leading coefficient.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper refines the moment-matrix extension algorithm of Brachat, Comon, Mourrain, and Tsigaridas for symmetric CP decomposition. It introduces the regularity of a target decomposition as the key complexity parameter, shows that for even-order tensors low regularity can reduce decomposition to solving linear systems, and proves that for order-4 tensors generic ranks up to r = 2n+1 are efficiently decomposable (Theorem 4.17), exceeding simultaneous diagonalization bounds. It further formulates Conjecture 1 for ranks O(n^2), provides finite-field computer-assisted verification for n = 2,...,17, and treats nonidentifiable tensors, including tensors with three collinear points and monomials. For monomials, the paper claims an explicit parameterization of the full space of minimal decompositions via Algorithm 5.
Significance. If the main results hold, this is a meaningful advance in algebraic tensor decomposition: it gives a clean regularity-based explanation of when moment matrix extension is efficient, proves new generic rank thresholds for order-4 tensors, supplies an effective identifiability criterion, and is one of the first works to handle nonidentifiable tensors efficiently. The generic-order-4 theorem (Theorem 4.17) appears well supported by detailed algebra, explicit specializations in the appendix, and a clear induction argument. The computational verifications over finite fields are exact and therefore constitute genuine evidence. The monomial parameterization, however, is the least secure part of the paper: its central theorem depends on a combinatorial equivalence stated in Remark 5.11 as a matter for future study, and on the external dimension formula Proposition 5.9. The paper is commendably explicit about this dependence, but as written the advertised parameterization of all minimal decompositions of monomials is conditional.
major comments (3)
- [Section 5, Remark 5.11 and Theorem 5.13] Theorem 5.13 asserts that Algorithm 5 produces all minimal decompositions of a monomial, but a load-bearing step is explicitly deferred. Remark 5.11 states that the paper does not prove that all determinantal equations whose leading variable is the same x^gamma impose the same constraint after rewriting in terms of YP, nor that E2 is automatically satisfied. Proposition 5.8's induction assumes that the distinct equations for a variable in Yk determine it consistently, and Lemma 5.12's surjectivity of pi_2 presupposes precisely this equivalence. As written, Theorem 5.13 therefore does not establish that every decomposition arises from a choice of YP values, and Algorithm 5 may parameterize only a subset. This needs either a complete proof of the combinatorial equivalence or a reformulation of Theorem 5.13 as a conditional statement.
- [Section 5, Lemma 5.12] Lemma 5.12 is used to prove that generic values of YP yield decompositions, but its proof has a circular aspect. The assertion that pi_2 is surjective 'because every decomposition of phi has a corresponding choice of free parameters' already assumes the missing statement that every solution of the full determinantal system is uniquely recoverable from its YP projection by the graded solving procedure. The equality of dimensions from Theorem 5.10 does not fill this gap, because Theorem 5.10 itself depends on Proposition 5.9 and the same deferred equivalence. The lemma should be rewritten to make explicit which of its hypotheses are established independently of the claim being proved.
- [Section 4.2.2, Lemma 4.16] The induction step in Lemma 4.16 is central to Theorem 4.17 for ranks n+1 < r < 2n+1, but its key assertion that the embedded matrix A(n,r) 'is actually the same submatrix' as A(n,r) under the format (n,r), after the specialization of points, is stated without a full verification. In particular, the claim that the new equations S have zero support on all other new variables in Y(n+1,r+1) \ Y(n,r) and the claim that the relevant minors m^xdelta_x^gamma are unchanged need to be checked explicitly, as they justify the block triangular reduction [A(n,r) 0; * mI]. Please expand this part of the proof, either in the main text or in the appendix, so that the induction is fully verifiable.
minor comments (4)
- [Algorithm 4] The while loop in Algorithm 4 has no explicit termination condition for the case where E is nonempty but no new moment variables can be eliminated; add a stopping rule or a note that the loop is finite by monotone elimination of variables.
- [Equation (7)] The symbol Y is used both for the set of moment variables and, with an abuse of notation, for the vector of moment variables in the linear system AY = b; using a different symbol such as y for the vector would improve readability.
- [Section 4.2.2, Table 3] The text says the verified ranks r_n are 'slightly better' than the conjecture-based bounds r'_n, but for small n the gap is not uniform (for n=3, r_n=7 versus r'_n=4); a sentence explaining this discrepancy would be helpful.
- [Section 5.3.1] The explicit choice of values for YP that yields a decomposition would benefit from a short summary of why the constructed multiplication matrices are nondefective; the computation is clear but the reader must reconstruct the eigenvector argument from the displayed matrices.
Circularity Check
No significant circularity: central claims are derived from the Hankel/determinantal system and external results, with only an explicitly deferred combinatorial check in the monomial parameterization.
full rationale
The paper's derivation chain is self-contained in the senses that matter for circularity. The main efficient-decomposition results (Theorem 4.17 and the computer-assisted Theorem 4.20) are obtained by analyzing the determinantal relations coming from a Hankel matrix with formal moment variables; the coefficient matrix A is constructed explicitly from minors of Vandermonde matrices of a hypothetical decomposition, and generic full column rank is established by exhibiting explicit point specializations over finite fields and by induction on the format. No parameter is fitted to the target decomposition and then renamed as a prediction: the linear system AY = b is derived from the tensor and the chosen monomial basis, and its full column rank is an open condition verified by evaluation. The monomial section does rely on Proposition 5.9 and Proposition 5.2 of Buczyńska–Buczyński–Teitler [12] for the dimension of the space of decompositions and for nonvanishing of the first coordinate; these are external prior results with no author overlap, and they are used as inputs, not as conclusions smuggled from this paper. Remark 5.11 explicitly concedes that the combinatorial equivalence of determinantal equations is deferred to future study and that standalone proofs of Propositions 5.2 and 5.9 are not given; this is a genuine completeness gap and a correctness risk for Theorem 5.13, but it is not circularity, because the missing verification is independent of the theorem being proved. Self-citations in the paper ([44], [45], [72]) are contextual or auxiliary and are not load-bearing for the main derivations. The conjecture on O(n^2) rank is clearly labeled as open and supported by independent finite-field checks, not by assuming the conclusion. Accordingly, no specific reduction of a claimed result to its own inputs can be exhibited, and the appropriate finding is no significant circularity.
Assumptions & free parameters
assumptions (4)
- domain assumption Theorem 3.7 of [11]: a size-s decomposition exists iff a solution to the determinantal/commuting relations with nondefective multiplication matrices exists.
- domain assumption Theorem 7.34 of [31]: common eigenvectors of commuting multiplication matrices are Vandermonde vectors z_B.
- domain assumption Proposition 5.2 of [12]: monomial minimal decompositions have all first coordinates nonzero.
- domain assumption Proposition 5.9 of [12]: dimension of the space of Waring decompositions of a monomial is a sum of Hilbert function values.
Cite this review
Pith. "Pith review of Efficient Tensor Decomposition via Moment Matrix Extension." pith.science (2026). https://pith.science/paper/ZFKWDKSZ
@misc{pith2026250622564,
author = {Pith},
title = {Pith review of: Efficient Tensor Decomposition via Moment Matrix Extension},
year = {2026},
howpublished = {\url{https://pith.science/paper/ZFKWDKSZ}},
note = {Machine review of arXiv:2506.22564}
}
abstract
Motivated by a flurry of recent work on efficient tensor decomposition algorithms, we show that the celebrated moment matrix extension algorithm of Brachat, Comon, Mourrain, and Tsigaridas for symmetric tensor canonical polyadic (CP) decomposition can be made efficient under the right conditions. We first show that the crucial property determining the complexity of the algorithm is the regularity of a target decomposition. This allows us to reduce the complexity of the vanilla algorithm, while also unifying results from previous works. We then show that for tensors in $S^d\mathbb{C}^{n+1}$ with $d$ even, low enough regularity can reduce finding a symmetric tensor decomposition to solving a system of linear equations. For order-$4$ tensors we prove that generic tensors of rank up to $r=2n+1$ can be decomposed efficiently via moment matrix extension, exceeding the rank threshold allowed by simultaneous diagonalization. We then formulate a conjecture that states for generic order-$4$ tensors of rank $r=O(n^2)$ the induced linear system is sufficient for efficient tensor decomposition, matching the asymptotics of existing algorithms and in fact improving the leading coefficient. Towards this conjecture we give computer assisted proofs that the statement holds for $n=2, \dots, 17$. Next we demonstrate that classes of nonidentifiable tensors can be decomposed efficiently via the moment matrix extension algorithm, bypassing the usual need for uniqueness of decomposition. Of particular interest is the class of monomials, for which the extension algorithm is not only efficient but also improves on existing theory by explicitly parameterizing the space of decompositions. Code for implementations of the efficient algorithm for generic tensors and monomials are provided, along with several numerical examples.
Reference graph
Works this paper leans on
-
[1]
Polynomial interpolation in several vari- ables
James Alexander and André Hirschowitz. “Polynomial interpolation in several vari- ables”.In: Journal of Algebraic Geometry4.2(1995),pp.201–222. issn:1056-3911,1534- 7486. 46
1995
-
[2]
Identifiability of param- eters in latent structure models with many observed variables
Elizabeth S. Allman, Catherine Matias, and John A. Rhodes. “Identifiability of param- eters in latent structure models with many observed variables”. In:Annals of Statistics 37.6A (2009), pp. 3099–3132.issn: 0090-5364,2168-8966. doi: 10.1214/09-AOS689
-
[3]
The Hilbert scheme of points and its link with border basis
Mariemi Alonso, Jerome Brachat, and Bernard Mourrain. “The Hilbert scheme of points and its link with border basis”. In:arXiv preprint arXiv:0911.3503(2009). doi: 10.48550/arXiv.0911.3503
work page Pith review arXiv doi:10.48550/arxiv.0911.3503 2009
-
[4]
Tensor decompositions for learning latent variable models
Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M. Kakade, and Matus Telgar- sky. “Tensor decompositions for learning latent variable models”. In:Journal of Ma- chine Learning Research15 (2014), pp. 2773–2832. issn: 1532-4435,1533-7928. doi: 10.21236/ada604494
-
[5]
Waring decompositions of special ternary forms with different Hilbert functions
Elena Angelini, Luca Chiantini, and Alessandro Oneto. “Waring decompositions of special ternary forms with different Hilbert functions”. In:Deformation of Artinian algebras and Jordan type. Vol. 805. Contemp. Math. Amer. Math. Soc., [Providence], RI, 2024, pp. 77–93.doi: 10.1090/conm/805/16127
-
[6]
Identifiability beyond Kruskal’s bound for symmetric tensors of degree 4
Elena Angelini, Luca Chiantini, and Nick Vannieuwenhoven. “Identifiability beyond Kruskal’s bound for symmetric tensors of degree 4”. In:Rendiconti Lincei, Matematica e Applicazioni29.3 (June 2018), pp. 465–485.issn: 1720-0768.doi: 10.4171/rlm/817
-
[7]
Spectral decomposition of a 4th-order covariance tensor: Applications to diffusion tensor MRI
Peter J. Basser and Sinisa Pajevic. “Spectral decomposition of a 4th-order covariance tensor: Applications to diffusion tensor MRI”. In:Signal Processing87.2 (2007). Tensor Signal Processing, pp. 220–236.issn: 0165-1684. doi: 10.1016/j.sigpro.2006.02. 050
-
[8]
Pencil-based algorithms for tensor rank decomposition are not stable
Carlos Beltrán, Paul Breiding, and Nick Vannieuwenhoven. “Pencil-based algorithms for tensor rank decomposition are not stable”. In:SIAM Journal on Matrix Analysis and Applications40.2 (2019), pp. 739–773.issn: 0895-4798,1095-7162. doi: 10.1137/ 18M1200531
2019
Show all 88 references
-
[9]
Tensor decomposition and homotopy continuation
Alessandra Bernardi, Noah S. Daleo, Jonathan D. Hauenstein, and Bernard Mourrain. “Tensor decomposition and homotopy continuation”. In:Differential Geometry and its Applications 55 (2017), pp. 78–105.doi: 10.1016/j.difgeo.2017.07.009
2017 doi
-
[10]
Waring, tangential and cactus decompo- sitions
Alessandra Bernardi and Daniele Taufer. “Waring, tangential and cactus decompo- sitions”. In: Journal de Mathématiques Pures et Appliquées. Neuvième Série (9)143 (2020), pp. 1–30.issn: 0021-7824,1776-3371. doi: 10.1016/j.matpur.2020.07.003
2020 doi
-
[11]
Symmet- ric tensor decomposition
Jerome Brachat, Pierre Comon, Bernard Mourrain, and Elias Tsigaridas. “Symmet- ric tensor decomposition”. In:Linear Algebra and its Applications433.11-12 (2010), pp. 1851–1872. issn: 0024-3795,1873-1856. doi: 10.1016/j.laa.2010.06.046
2010 doi
-
[12]
Waring decompositions of monomials
Weronika Buczyńska, Jarosław Buczyński, and Zach Teitler. “Waring decompositions of monomials”. In: Journal of Algebra 378 (2013), pp. 45–57. issn: 0021-8693,1090- 266X. doi: 10.1016/j.jalgebra.2012.12.011. 47
2013 doi
- [13]
-
[14]
Blind signal separation: Statistical principles
Jean-François Cardoso. “Blind signal separation: Statistical principles”. In:Proceedings of the IEEE86.10 (1998), pp. 2009–2025.doi: 10.1109/5.720250
1998 doi
-
[15]
Super-symmetric decomposition of the fourth-order cumulant tensor. Blind identification of more sources than sensors
Jean-François Cardoso. “Super-symmetric decomposition of the fourth-order cumulant tensor. Blind identification of more sources than sensors”. In:[Proceedings] ICASSP 91: 1991 International Conference on Acoustics, Speech, and Signal Processing. 1991, 3109–3112 vol.5. doi: 10....
1991
-
[16]
Reducing the number of variables of a polynomial
Enrico Carlini. “Reducing the number of variables of a polynomial”. In:Algebraic Ge- ometry and Geometric Modeling. Math. Vis. Springer, Berlin, 2006, pp. 237–247.doi: 10.1007/978-3-540-33275-6\_15
2006 doi
-
[17]
Waring loci and the Strassen conjecture
Enrico Carlini, Maria Virginia Catalisano, and Alessandro Oneto. “Waring loci and the Strassen conjecture”. In:Advances in Mathematics314 (2017), pp. 630–662.issn: 0001-8708,1090-2082. doi: 10.1016/j.aim.2017.05.008
2017 doi
-
[18]
Analysis of individual differences in multidi- mensional scaling via anN-way generalization of “Eckart-Young
J. Douglas Carroll and Jih-Jie Chang. “Analysis of individual differences in multidi- mensional scaling via anN-way generalization of “Eckart-Young” decomposition”. In: Psychometrika 35.3 (1970), pp. 283–319.doi: 10.1007/BF02310791
1970 doi
-
[19]
Optimal separation of independent narrow-band sources: Concept and performance
Pascal Chevalier. “Optimal separation of independent narrow-band sources: Concept and performance”. In:Signal Processing73.1 (1999), pp. 27–47.issn: 0165-1684. doi: 10.1016/S0165-1684(98)00183-2
1999 doi
-
[20]
An Algorithm For Generic and Low-Rank Specific Identifiability of Complex Tensors
Luca Chiantini, Giorgio Ottaviani, and Nick Vannieuwenhoven. “An Algorithm For Generic and Low-Rank Specific Identifiability of Complex Tensors”. In:SIAM Jour- nal on Matrix Analysis and Applications35.4 (2014), pp. 1265–1287.doi: 10.1137/ 140961389
2014
-
[21]
Effective criteria for specific identifiability of tensors and forms
Luca Chiantini, Giorgio Ottaviani, and Nick Vannieuwenhoven. “Effective criteria for specific identifiability of tensors and forms”. In:SIAM Journal on Matrix Analysis and Applications 38.2 (2017), pp. 656–681.doi: 10.1137/16m1090132
2017 doi
-
[22]
On generic identifi- ability of symmetric tensors of subgeneric rank
Luca Chiantini, Giorgio Ottaviani, and Nick Vannieuwenhoven. “On generic identifi- ability of symmetric tensors of subgeneric rank”. In:Trans. Amer. Math. Soc.369.6 (2017), pp. 4021–4042.issn: 0002-9947,1088-6850. doi: 10.1090/tran/6762
2017 doi
-
[23]
On the rank of a binary form
Gonzalo Comas and Malena Seiguer. “On the rank of a binary form”. In:Foundations of Computational Mathematics11.1 (2011), pp. 65–78.issn: 1615-3375,1615-3383.doi: 10.1007/s10208-010-9077-x
2011 doi
-
[24]
Symmetric Ten- sors and Symmetric Tensor Rank
Pierre Comon, Gene Golub, Lek-Heng Lim, and Bernard Mourrain. “Symmetric Ten- sors and Symmetric Tensor Rank”. In:SIAM Journal on Matrix Analysis and Appli- cations 30.3 (2008), pp. 1254–1279.doi: 10.1137/060661569. 48
2008 doi
- [25]
-
[26]
Tensor-based techniques for the blind separation of DS–CDMA signals
Lieven De Lathauwer and Joséphine Castaing. “Tensor-based techniques for the blind separation of DS–CDMA signals”. In:Signal Processing87.2 (2007), pp. 322–336.doi: 10.1016/j.sigpro.2005.12.015
2007 doi
-
[27]
Fourth-order cumulant-based blind identification of underdetermined mixtures
Lieven De Lathauwer, Joséphine Castaing, and Jean-François Cardoso. “Fourth-order cumulant-based blind identification of underdetermined mixtures”. In:IEEE Trans- actions on Signal Processing55.6 (2007), pp. 2965–2973.issn: 1053-587X,1941-0476. doi: 10.1109/TSP.2007.893943
2007
-
[29]
Uncertainty principles and ideal atomic decom- position
David L. Donoho and Xiaoming Huo. “Uncertainty principles and ideal atomic decom- position”. In: IEEE Transactions on Information Theory47.7 (2001), pp. 2845–2862. issn: 0018-9448,1557-9654. doi: 10.1109/18.959265
2001 doi
-
[30]
The Geometry of Syzygies
David Eisenbud. The Geometry of Syzygies. Vol. 229. Graduate Texts in Mathematics. A second course in commutative algebra and algebraic geometry. Springer-Verlag, New York, 2005, pp. xvi+243.isbn: 0-387-22215-4. doi: 10.1007/b137572
2005 doi
-
[31]
Introduction à la Résolution des Systèmes Polynomiaux
Mohamed Elkadi and Bernard Mourrain. Introduction à la Résolution des Systèmes Polynomiaux. Vol. 59. Mathématiques & Applications (Berlin) [Mathematics & Appli- cations]. Springer, Berlin, 2007, pp. vi+305.doi: 10.1007/978-3-540-71647-1
2007 doi
-
[32]
Nemo/Hecke: Computer algebra and number theory packages for the Julia programming language
Claus Fieker, William Hart, Tommy Hofmann, and Fredrik Johansson. “Nemo/Hecke: Computer algebra and number theory packages for the Julia programming language”. In: ISSAC’17—Proceedings of the 2017 ACM International Symposium on Symbolic and Algebraic Computation. ACM, New York...
2017
- [33]
-
[34]
Learning mixtures of Gaussians in high dimensions [extended abstract]
Rong Ge, Qingqing Huang, and Sham M. Kakade. “Learning mixtures of Gaussians in high dimensions [extended abstract]”. In:STOC’15—Proceedings of the 2015 ACM Symposium on Theory of Computing. ACM, New York, 2015, pp. 761–770.isbn: 978- 1-4503-3536-2. doi: 10.1145/2746539.2746616
2015
-
[35]
Decomposing overcomplete 3rd order tensors using sum-of- squares algorithms
Rong Ge and Tengyu Ma. “Decomposing overcomplete 3rd order tensors using sum-of- squares algorithms”. In:Approximation, Randomization, and Combinatorial Optimiza- tion. Algorithms and Techniques (APPROX/RANDOM 2015). Ed. by Naveen Garg, Klaus Jansen, Anup Rao, and José D. P. R...
2015 doi
-
[36]
PARAFAC: An “explanatory
Richard A. Harshman. “PARAFAC: An “explanatory” factor analysis procedure”. In: The Journal of the Acoustical Society of America 50.1A_Supplement (July 1971), pp. 117–117. issn: 1520-8524. doi: 10.1121/1.1977523
1971 doi
-
[37]
Most tensor problems are NP-hard
Christopher J. Hillar and Lek-Heng Lim. “Most tensor problems are NP-hard”. In: Journal of the ACM 60.6 (2013), Art. 45, 39. issn: 0004-5411,1557-735X. doi: 10. 1145/2512329
2013
-
[38]
A robust spectral algorithm for overcomplete tensor decomposition
Samuel B. Hopkins, Tselil Schramm, and Jonathan Shi. “A robust spectral algorithm for overcomplete tensor decomposition”. In:Proceedings of the Thirty-Second Confer- ence on Learning Theory. Ed. by Alina Beygelzimer and Daniel Hsu. Vol. 99. Proceed- ings of Machine Learning Re...
2019
-
[39]
Fast spectral algorithms from sum-of-squares proofs: Tensor decomposition and planted sparse vec- tors
Samuel B. Hopkins, Jonathan Shi, Tselil Schramm, and David Steurer. “Fast spectral algorithms from sum-of-squares proofs: Tensor decomposition and planted sparse vec- tors”. In: STOC’16—Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing. ACM, New York, ...
2016
-
[40]
Learning mixtures of spherical Gaussians: Moment methods and spectral decompositions
Daniel Hsu and Sham M. Kakade. “Learning mixtures of spherical Gaussians: Moment methods and spectral decompositions”. In: ITCS’13—Proceedings of the 2013 ACM Conference on Innovations in Theoretical Computer Science. ACM, New York, 2013, pp. 11–19. isbn: 978-1-4503-1859-4. do...
2013
-
[41]
Independent Component Analy- sis
Aapo Hyvärinen, Juha Karhunen, and Erkki Oja. Independent Component Analy- sis. English. United Kingdom: Wiley, 2001. isbn: 9780471405405. doi: 10 . 1002 / 0471221317
2001
-
[42]
Anthony Iarrobino and Vassil Kanev.Power Sums, Gorenstein Algebras, and Determi- nantal Loci. Vol. 1721. Lecture Notes in Mathematics. Appendix C by Iarrobino and Steven L. Kleiman. Springer-Verlag, Berlin, 1999, pp. xxxii+345.isbn: 3-540-66766-0. doi: 10.1007/BFb0093426
1999 doi
-
[43]
Computing linear sections of varieties: Quantum entanglement, tensor decompositions and be- yond
Nathaniel Johnston, Benjamin Lovitz, and Aravindan Vijayaraghavan. “Computing linear sections of varieties: Quantum entanglement, tensor decompositions and be- yond”. In:2023 IEEE 64th Annual Symposium on Foundations of Computer Science— FOCS 2023. IEEE Computer Soc., Los Alam...
2023
-
[44]
Landscape analysis of an improved power method for tensor decomposition
Joe Kileel, Timo Klock, and João M. Pereira. “Landscape analysis of an improved power method for tensor decomposition”. In:Advances in Neural Information Process- ing Systems 34 (2021), pp. 6253–6265. url: https : / / proceedings . neurips . cc / paper/2021/hash/31784d9fc1fa0d...
2021
- [45]
-
[46]
An efficient uniqueness theorem for overcomplete tensor decomposi- tion
Pascal Koiran. “An efficient uniqueness theorem for overcomplete tensor decomposi- tion”. In: Proceedings of the 2025 Annual ACM-SIAM Symposium on Discrete Algo- rithms (SODA). SIAM, Philadelphia, PA, 2025, pp. 1909–1932.isbn: 978-1-61197-832-
2025
-
[47]
doi: 10.1137/1.9781611978322.60
-
[48]
On the uniqueness and computation of commuting extensions
Pascal Koiran. “On the uniqueness and computation of commuting extensions”. In: Linear Algebra and its Applications703 (2024), pp. 645–666. issn: 0024-3795,1873-
2024
-
[49]
Tensor decompositions and applications
Tamara G. Kolda and Brett W. Bader. “Tensor decompositions and applications”. In: SIAM Review 51.3 (2009), pp. 455–500. issn: 0036-1445,1095-7200. doi: 10 . 1137 / 07070111X
2009
-
[50]
Numerical optimization for symmetric tensor decomposition
Tamara G. Kolda. “Numerical optimization for symmetric tensor decomposition”. In: Mathematical Programming151.1(2015),pp.225–248. issn:0025-5610,1436-4646. doi: 10.1007/s10107-015-0895-0
2015 doi
-
[51]
Overcomplete tensor decomposition via Koszul-Young flattenings
Pravesh K. Kothari, Ankur Moitra, and Alexander S. Wein. “Overcomplete tensor decomposition via Koszul-Young flattenings”. In: arXiv preprint arXiv:2411.14344 (2024). doi: 10.48550/arXiv.2411.14344
2024 doi
-
[52]
Shifted power method for computing ten- sor eigenpairs
Tamara G. Kolda and Jackson R. Mayo. “Shifted power method for computing ten- sor eigenpairs”. In:SIAM Journal on Matrix Analysis and Applications32.4 (2011), pp. 1095–1124. issn: 0895-4798,1095-7162. doi: 10.1137/100801482
2011 doi
-
[53]
J. M. Landsberg. Tensors: Geometry and Applications. Vol. 128. Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2012, pp. xx+439. isbn: 978-0-8218-6907-9. doi: 10.1090/gsm/128
2012 doi
-
[54]
Three-way arrays: Rank and uniqueness of trilinear decomposi- tions, with application to arithmetic complexity and statistics
Joseph B. Kruskal. “Three-way arrays: Rank and uniqueness of trilinear decomposi- tions, with application to arithmetic complexity and statistics”. In:Linear Algebra and its Applications18.2 (1977), pp. 95–138.doi: 10.1016/0024-3795(77)90069-6
1977 doi
-
[55]
A decomposition for three-way arrays
Sue E. Leurgans, Robert T. Ross, and Robert B. Abel. “A decomposition for three-way arrays”. In:SIAM Journal on Matrix Analysis and Applications14.4 (1993), pp. 1064–
1993
-
[56]
Equations for secant varieties of Veronese and other varieties
J. M. Landsberg and Giorgio Ottaviani. “Equations for secant varieties of Veronese and other varieties”. In:Ann. Mat. Pura Appl. (4)192.4 (2013), pp. 569–606. issn: 0373-3114,1618-1891. doi: 10.1007/s10231-011-0238-6
2013 doi
-
[57]
Polynomial-time tensor decompositions with sum-of-squares
Tengyu Ma, Jonathan Shi, and David Steurer. “Polynomial-time tensor decompositions with sum-of-squares”. In:57th Annual IEEE Symposium on Foundations of Computer Science—FOCS 2016. IEEE Computer Soc., Los Alamitos, CA, 2016, pp. 438–446. isbn: 978-1-5090-3933-3. doi: 10.1109/f...
2016 doi
-
[58]
On minimal decompositions of low rank symmetric tensors
Bernard Mourrain and Alessandro Oneto. “On minimal decompositions of low rank symmetric tensors”. In:Linear Algebra and its Applications607 (2020), pp. 347–377. issn: 0024-3795,1873-1856. doi: 10.1016/j.laa.2020.06.029
2020 doi
-
[59]
A generalization of Kruskal’s theorem on tensor decomposition
Benjamin Lovitz and Fedor Petrov. “A generalization of Kruskal’s theorem on tensor decomposition”. In:Forum of Mathematics, Sigma11 (2023), e27.doi: 10.1017/fms. 2023.20. 51
2023 doi
-
[60]
The A-truncated K-moment problem
Jiawang Nie. “The A-truncated K-moment problem”. In: Foundations of Computa- tional Mathematics 14.6 (2014), pp. 1243–1276.doi: 10.1007/s10208-014-9225-9
2014 doi
-
[61]
EigenvectorsoftensorsandalgorithmsforWaring decomposition
LukeOedingandGiorgioOttaviani.“EigenvectorsoftensorsandalgorithmsforWaring decomposition”. In:Journal of Symbolic Computation54 (2013), pp. 9–35.issn: 0747- 7171,1095-855X. doi: 10.1016/j.jsc.2012.11.005
2013 doi
-
[62]
Generatingpolynomialsandsymmetrictensordecompositions
JiawangNie.“Generatingpolynomialsandsymmetrictensordecompositions”.In: Foun- dations of Computational Mathematics17.2 (2017), pp. 423–465.issn: 1615-3375,1615-
2017
-
[63]
Overcomplete independent component analysis via SDP
Anastasia Podosinnikova, Amelia Perry, Alexander S. Wein, Francis Bach, Alexandre d’Aspremont, and David Sontag. “Overcomplete independent component analysis via SDP”. In: Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics.Ed.by...
2019
-
[64]
Varieties of sums of powers
Kristian Ranestad and Frank-Olaf Schreyer. “Varieties of sums of powers”. In:Jour- nal für die Reine und Angewandte Mathematik525 (2000), pp. 147–181.issn: 0075- 4102,1435-5345. doi: 10.1515/crll.2000.064
2000 doi
-
[65]
Decomposing tensors via rank-one approximations
Alvaro Ribot, Emil Horobet, Anna Seigal, and Ettore Teixeira Turatti. “Decomposing tensors via rank-one approximations”. In: arXiv preprint arXiv:2411.15935 (2024). doi: 10.48550/arXiv.2411.15935
2024 doi
- [66]
-
[67]
Explicit projective embeddings of stan- dard opens of the Hilbert scheme of points
Roy Mikael Skjelnes and Gustav Sædén Ståhl. “Explicit projective embeddings of stan- dard opens of the Hilbert scheme of points”. In:Journal of Algebra590 (2022), pp. 254–
2022
-
[68]
Sur une extension d’un théorème de Clebsch relatif aux courbes du quatrième degré
James J. Sylvester. “Sur une extension d’un théorème de Clebsch relatif aux courbes du quatrième degré”. In:Comptes rendus de l’Académie des Sciences Paris102 (1886), pp. 1532–1534. 52
-
[69]
A normal form algorithm for tensor rank decomposition
Simon Telen and Nick Vannieuwenhoven. “A normal form algorithm for tensor rank decomposition”. In:ACM Transactions on Mathematical Software48.4 (2022), Art. 38,
2022
-
[70]
Blind PARAFAC receivers for DS-CDMA systems
Nicholas D. Sidiropoulos, Georgios B. Giannakis, and Rasmus Bro. “Blind PARAFAC receivers for DS-CDMA systems”. In:IEEE Transactions on Signal Processing48.3 (2000), pp. 810–823.doi: 10.1109/78.824675
2000 doi
-
[71]
An analytical constant modulus algorithm
Alle-Jan van der Veen and Arogyaswami Paulraj. “An analytical constant modulus algorithm”. In: IEEE Transactions on Signal Processing44.5 (1996), pp. 1136–1155. doi: 10.1109/78.502327
1996 doi
-
[72]
Moment estimation for nonparametric mixture models through implicit tensor decomposition
Yifan Zhang and Joe Kileel. “Moment estimation for nonparametric mixture models through implicit tensor decomposition”. In: SIAM Journal on Mathematics of Data Science 5.4 (2023), pp. 1130–1159.doi: 10.1137/22M153879X. 53 Appendix A Additional Proofs and Examples for Section 4...
2023 doi
- [75]
-
[76]
On the Waring rank of binary forms
Neriman Tokcan. “On the Waring rank of binary forms”. In:Linear Algebra and its Applications 524 (2017), pp. 250–262. issn: 0024-3795,1873-1856. doi: 10.1016/j. laa.2017.03.007
2017 doi
-
[79]
Then B = {1, x1, x2, x2 1}
Let r = 4 with some decompositionZ. Then B = {1, x1, x2, x2 1}. Then Y = {x5 1, x4 1x2} and E1 = {(x3 1, x1x2)}, E2 = {[(x3 1, x2 2), (x2 1x2, x1x2)]}. Therefore, A = "−mx1x2 x2 1 m −m x2 2 x2 1 mx1x2 x2 1 # . If m · m x2 2 x2 1 − mx1x2 x2 1 2 ̸= 0 as a polynomial in the entri...
-
[80]
Then B = {1, x1, x2, x2 1, x1x2}
Let r = 5 with some decompositionZ. Then B = {1, x1, x2, x2 1, x1x2}. Then Y = {x5 1, x4 1x2, x3 1x2 2, x2 1x3 2} and E1 = {(x3 1, x2 2), (x2 1x2, x2 2)}, E2 = ∅. There are four moment variables but only two linear equations, so linear equations do not suffice. For n = 3, it s...
-
[81]
There are three moment variables Y = {x5 1, x4 1x2, x4 1x3}
Let Z = 1 0 0 0 1 1 0 0 1 0 1 0 1 0 0 1 1 −1 2 2 , so Z2 = 1 0 0 0 0 0 0 0 0 0 1 1 0 0 1 0 0 0 0 0 1 0 1 0 0 0 0 1 0 0 1 0 0 1 0 0 0 0 0 1 1 −1 2 2 1 −2 −2 4 4 4 . There are three moment variables Y = {x5 1, x4 1x2, x4 1x3}. There are many equat...
-
[82]
Let Z = 1 0 0 0 1 1 0 0 1 0 1 0 1 0 0 1 1 −1 2 2 1 −1 −1 2 . There are seven variables, and a7 × 7 full-rank submatrix ofA is 6 0 6 0 0 0 0 −6 0 0 6 0 0 0 0 12 0 0 6 0 0 0 6 0 0 6 0 0 0 −6 0 0 0 6 0 0 0 0 12 0 0 6 0 0 0 6 0 0 6
-
[83]
Let Z = 1 0 0 0 1 1 0 0 1 0 1 0 1 0 0 1 1 −1 2 2 1 −1 −1 2 1 −1 −1 −1 . There are ten variables, and a10 × 10 full-rank submatrix ofA is −18 0 0 18 0 0 0 0 0 0 −18 36 −18 0 18 0 0 0 0 0 −18 0 0 0 0 18 0 0 0 0 0 −18 0 0 0 0 18 0 0 0 0 −...
-
[84]
Proof of Lemma 4.15.By Lemma 4.13, forn ≥ 4 |E1| ≥ |Y |
For r ≥ 8 there are simply not enough linear equations by a naive count. Proof of Lemma 4.15.By Lemma 4.13, forn ≥ 4 |E1| ≥ |Y |. We give a family of special- izations. Construct z1 = e0, zk = e0 + ek−1, k= 2, . . . , n+ 1, zk = 1 −1(k−(n+1))×1 2(2n+1−k)×1 , k= n + 2, ...
-
[85]
Take the dot product of this column with the row (x2 1xj, xkxℓ)
In the casec = 1 such a xγ corresponds to xγ = x4 1xi, i≥ 1. Take the dot product of this column with the row (x2 1xj, xkxℓ). This is the sum nX b=2 (−1)r+1+pos(x1xb)(−1)r+1+pos(x1xi)mxkxℓ x1xb mxj xb x1xi = nX b=2 (−1)pos(xixb)+pos(x1xi)mxkxℓ x1xb mxj xb x1xi . Taking now the...
-
[86]
− col(x1) we obtain the matrix 1 G 1 0(n+1)×n 1 G 3 G′ , where G′ = 2 −2 . . .−2 2 1 . . .−2 ... ... 2 1 . . .−2 2 1 . . . 1 . Then det (ZB) = det (G′) = 2 det (G) = 2 · 3n−1. To computem x2 k x2 1 for 2 ≤ k ≤ n, we first see thatG(1,k,k) 4 is the lastn − 1 c...
-
[87]
Now assumej < k
− col (x1) and col x2 j ← col x2 j − col (xj) leads to a repeated column, so the determinant must be zero. Now assumej < k. Take col (x2
-
[88]
Gx1xi−1 Gx1xi+1
− col (x1), so we are left with determining the determinant of the matrix 2G2 x1 . . .Gx1xi−1 Gx1xi+1 . . .Gx1xn vjk . We instead cyclevjk into x1xi index and first calculate G′ = 2G2 x1 . . .Gx1xi−1 vjk Gx1xi+1 . . .Gx1xn , accounting for sign later. Then det (G′) = 2 det G +...
-
[89]
But the difference of the latter two equations is exactly the first equation
∈ E1. But the difference of the latter two equations is exactly the first equation. A more subtle dependency is as follows. Consider the following six equations inE1: (x3 1, x2 2), (x3 1, x2 3), (x2 1x2, x2 2), (x2 1x2, x2 3), (x2 1x3, x2 2), (x2 1x3, x2 3). It can be seen tha...
-
[276]
doi: 10.1016/j.jalgebra.2021.10.007
issn: 0021-8693,1090-266X. doi: 10.1016/j.jalgebra.2021.10.007
2021 doi
- [1083]
-
[1856]
doi: 10.1016/j.laa.2024.10.004
2024 doi
-
[3383]
doi: 10.1007/s10208-015-9291-7
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.