Identifiability of Nonnegative Tucker Decompositions -- Part I: Theory

Giovanni Barbarino; Nicolas Gillis; Subhayan Saha

REVIEW 1 major objections 8 minor 56 references

Identifiability of Nonnegative Tucker Decompositions -- Part I: Theory

T0 review · 1 major / 8 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper proves that nonnegative Tucker decompositions are identifiable, up to permutation, whenever the factor matrices are sufficiently scattered and the core tensor has full-rank slices or unfoldings, and that minimizing the core's…

desk verdict The core identifiability theorems for nonnegative Tucker decompositions are new and correct; the printed proofs have fixable typos, one in a load-bearing spot. read the letter →

arxiv 2505.12713 v1 pith:AUEJZWLW submitted 2025-05-19 math.NA cs.LGcs.NAeess.SPstat.ML

Subhayan Saha , Giovanni Barbarino , Nicolas Gillis This is my paper · ORCID

classification math.NAcs.LGcs.NAeess.SPstat.ML MSC 15A2315A69

keywords nonnegativeTuckerdecompositionidentifiabilitysufficientlyscatteredconditionminimum-volumefactorizationseparabilitymatrixtri-factorizationtensorslicesKroneckerproduct

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 establishes conditions under which a nonnegative Tucker decomposition (nTD) is essentially unique. The main idea is to import the sufficiently scattered condition from nonnegative matrix factorization: if each factor matrix is sufficiently spread out over the nonnegative orthant, and if the core tensor has enough full-rank slices or unfoldings, then minimizing the volume of the core recovers the true factors, up to permutation and trivial scaling. The results cover order-2 tri-factorizations, order-3 tensors with five distinct procedures, and higher-order generalizations. The paper also characterizes when the Kronecker product of two sufficiently scattered matrices is itself sufficiently scattered, which matters for the unfolding-based procedures.

What carries the argument

The central object is the sufficiently scattered condition (SSC): a nonnegative matrix $U$ satisfies the SSC when its row cone contains the pointed cone $C = \{x \ge 0 : e^\top x \ge \sqrt{r-1}\,\|x\|_2\}$ and also satisfies a dual-cone boundary condition. The SSC is what makes minimum-volume NMF identifiable, and the paper adapts that proof to the determinant of the core matrix. For slice-based procedures, the key mechanical step is rewriting a slice $T^{(3)}_j$ as $U_1 S_j U_2^\top$ with $S_j = \sum_k (U_3)_{j,k} G^{(3)}_k$, so that rank conditions on slices translate into rank conditions on linear combinations of core slices. For random-combination procedures, the machinery is a determinant polynomial $P(x) = \det\bigl(\sum_k ((U_3^\#)^\top x)_k G^\#_k\bigr)$ whose non-vanishing guarantees that a random slice combination has full rank with probability 1.

What would settle it

Build an order-3 tensor satisfying Assumption 5.13 with $r_1=r_2=4$, $r_3=2$, where the third-mode slice span has maximal rank but every individual slice is singular; if a positive-measure set of random coefficients $\alpha$ makes $T_\alpha$ rank-deficient, Theorem 5.14's probability-1 claim fails. Equivalently, check the proof's claimed point: the text uses $x_0 = (((U^\#_3)^\dagger)^\dagger)^\top t$, whereas $(U^\#_3)^\top x_0 = t$ requires $x_0 = (U^\#_3)^\dagger{}^\top t$; substituting the printed value invalidates the determinant-nonvanishing argument.

Watch

Extended reading notes

Core claim

The central discovery is that identifiability of an nTD splits into two independent ingredients: sparsity of the factor matrices, and rank abundance of the core. For order-2 nTD, the paper proves that if $X = U_1^\# G^\# U_2^{\#\top}$ has rank $r$, both $U_i^\#$ satisfy the sufficiently scattered condition, and $U_i^{\#\top} e = e$, then minimizing $|\det(G)|$ over all decompositions $X = U_1 G U_2^\top$ with $U_i \ge 0$ and $U_i^\top e = e$ recovers the factors up to permutation; this is Theorem 4.2. For order-3 tensors, the paper gives five procedures, based on unfoldings or on single slices, spans of slices, or random combinations of slices, that achieve the same kind of essential uniqueness under progressively weaker core assumptions. The same pattern is extended to arbitrary order-d tensors.

Load-bearing premise

The weakest link is the claim that a random linear combination of the core's slices has full rank with probability 1, a step that relies on a determinant polynomial not being identically zero and whose printed choice of evaluation point appears to contain a slip.

Editorial extensions

If this is right

If the theorems are correct, any exact nTD problem whose factors satisfy the SSC and whose core has the stated rank properties can be solved in principle by a single volume-minimization step, without needing nonnegativity of the core.
For order-3 tensors, the practical conditions are checkable: one can first verify the SSC of the factors and then test whether two slices, or the span of all slices in two modes, reach the required rank.
The randomized procedures imply that, for generic choices of the random coefficients, identifiability holds with probability 1 as long as the relevant matrix spaces have maximal rank, which weakens the requirement from a single full-rank slice to a full-rank span.
The Kronecker-product results show that the unfolding-based procedures are safe when one inner dimension is at most 2 or both are 3, but can fail in worst-case constructions when both dimensions are at least 3 and one is at least 4, so the unfolding approach needs an explicit SSC check in those cases.
For higher-order tensors, the same volume-minimization strategy generalizes by grouping modes into blocks and applying min-vol NMF or min-vol order-2 nTD to the corresponding unfoldings and slices.

Reading between the lines

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

One likely practical consequence, not spelled out in the paper, is that the SSC-based volume-minimization objective could serve as a regularizer for tensor completion or source separation, since it penalizes degenerate cores without requiring the core to be nonnegative.
A testable extension would be to study whether the randomized procedures remain identifiable when the core slices only span a matrix space of rank close to, but not equal to, the maximum; the current theory requires exact maximal rank.
The paper states that the construction and comparison of numerical algorithms is deferred to its second part; the identifiability theorems therefore guarantee uniqueness of global optima, but not that a particular nonconvex solver will reach them.
Since the worst-case counterexamples for Kronecker products need very large factor matrices, a natural open question is whether random SSC matrices, under a reasonable distribution, yield an SSC Kronecker product with high probability; the paper's numerical experiment suggests this may often be true.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know two things about this paper. First, the main theoretical contribution is real: Theorem 4.2 gives the first identifiability result for order-2 nonnegative matrix tri-factorization under the sufficiently scattered condition, and the slice/unfolding procedures in Sections 5 and 6 extend NMF-style uniqueness to Tucker decompositions in a systematic way. Second, the printed proofs have a handful of typos, one of them load-bearing, but the results survive once the typos are corrected.

What I checked most carefully is the randomized-slice argument in Theorem 5.14. The stress-test note is right: the proof chooses x0 = (((U3^†)^†)^T t, which does not give (U3)^T x0 = t. The correct choice is x0 = (U3^†)^T t. With that fix, P(x0) = det(sum t_k G_k) ≠ 0, and the 'with probability 1' claim goes through. So this is a fixable slip, not a gap in the theorem. Similarly, the C* in Theorem 4.2 is printed with an extra nonnegativity constraint that is not in Definition 2.7; the proof only needs the definition as given. Equation (10) has B^{-1} instead of B^{-⊤} in one display, but the determinant is the same, so that's cosmetic.

The paper is honest about its main limitation: the identifiability guarantees are for exact decompositions and assume you can solve the nonconvex min-vol problems. Algorithms are deferred to Part II. That is a scope limitation, not a hidden flaw. The reductions from NMF are clean, there are no fitted constants, and the citation pattern is normal for this line of work.

Who gets value: anyone working on uniqueness of nonnegative tensor decompositions, and people in signal processing or data analysis who use Tucker models and want interpretable factors. The paper deserves a serious referee. I would send it out. The typos need to be fixed, especially the pseudoinverse choice in Theorem 5.14, but the core results are new and appear correct.

Referee Report

1 major / 8 minor

Summary. The paper develops identifiability theory for nonnegative Tucker decompositions (nTDs). It first proves an order-2 result: under the sufficiently scattered condition (SSC) on the two factor matrices and a full-rank core, the minimum-volume problem (9) recovers the factors up to permutation (Theorem 4.2), alongside a separable analogue (Theorem 4.1). It then gives five order-3 procedures, one unfolding-based and four slice-based, and proves identifiability under SSC on factors plus rank conditions on slices, slice spans, or unfoldings of the core (Theorems 5.4, 5.7, 5.11, 5.14, 5.16, 5.18). These results are extended to order d in Section 6. Section 7 studies when a Kronecker product of SSC matrices satisfies SSC, with a positive low-rank corollary and a counterexample construction for larger dimensions.

Significance. Assuming the stated results hold after the corrections below, this is a useful and fairly complete theory paper. The order-2 minimum-volume identifiability theorem is new, and the slice and unfolding reductions convert NMF identifiability into conditions on core slices that are weaker than full unfolding rank conditions and are comparatively easy to state and check. Section 7 contributes nontrivial information about SSC preservation under Kronecker products, including a rigorous negative result. The paper is honest that algorithms are deferred to Part II, so the contribution is theoretical; that is appropriate for a Part I paper. No machine-checked proofs or code accompany the paper, but the proofs are mostly self-contained reductions to known NMF theorems.

major comments (1)

[Theorem 5.14 proof] The choice x0 = (((U#3)^dagger)^dagger)^T t is incorrect. Since (A^dagger)^dagger = A for the Moore-Penrose pseudoinverse, this gives (U#3)^T x0 = (U#3)^T (U#3)^T t, not t, so the displayed evaluation P(x0) = det(sum_k t_k G#_k) does not follow. The correct choice is x0 = (U#3^dagger)^T t; full column rank of U#3 gives (U#3)^T x0 = t. This step is load-bearing for the 'with probability 1' full-rank conclusion, and Theorem 5.18 and the higher-order analogues rely on the same argument. With the correction the polynomial non-vanishing argument is valid, so this is a fixable proof error rather than a false result. In the same proof, the sums over k in [r] should be over [r3] in the definition of S_alpha and P(x), since G# has r3 slices along the third mode.

minor comments (8)

[Equation (10)] In the proof of Theorem 4.2, the displayed determinant should use B^{-T}, not B^{-1}, because G* = A^{-1} G# B^{-T}; the determinant magnitude is unaffected, so the argument goes through once the typo is fixed.
[Equation (14)] The proof of Theorem 4.2 writes C* = {x in R^r_+ : e^T x >= ||x||_2}, while Definition 2.7 defines C* without the nonnegativity restriction. The proof only needs the nonnegativity-free definition together with the SSC2 boundary condition, but the mismatch should be corrected.
[Theorem 5.16] The statement says 'essentially unique with probability 1', but Procedure 3 is deterministic; the qualification 'with probability 1' should be removed.
[Theorem 6.13] The last displayed equation writes G* = (Pi_1^T, ..., Pi_4^T).G#, but the index should be d, not 4.
[Appendix A.1] In Theorem A.1, the claim that D_k is the identity because U_k has entries in [0,1] is false in general; for example, scaling a separable factor by 1/2 preserves separability and the [0,1] bounds. The theorem should include diagonal matrices, as in Theorem 4.1 and Definition 2.1.
[Theorem 4.2 proof] In the boundary argument, 'B[i,:]' should be 'B[:,i]' to match the notation used for the columns of B.
[Theorem 5.14 proof] In the second part of the proof, the phrase 'is a solution of (25)' should refer to problem (26), since that is the min-vol NMF problem solved for U3 and S_beta.
[Procedure 2] In step 2 of Procedure 2, the optimal solution is denoted S*_{i3}, but the variable being optimized is S_alpha; the subscript should be alpha for consistency.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: the paper derives nTD identifiability by reduction to external NMF theorems and newly introduced min-volume formulations; the only self-citations are to future Part II and are not load-bearing.

full rationale

The derivation chain is a sequence of reductions rather than a self-referential loop. Theorem 4.2 proves min-volume order-2 nTD identifiability from the sufficiently scattered condition and a determinant argument; Theorems 5.4, 5.11, 5.14, 5.16, 5.18, and the higher-order theorems reduce order-3 and order-d identifiability to Theorem 4.2 and to the external min-vol NMF theorem of Fu et al. (Theorem 2.8). The core tensor is not fitted to the factors and then re-predicted; volume minimization of unfoldings or slices is a newly proposed objective, and the proofs show that any optimizer must be a permutation of the ground truth. No fitted input is renamed as a prediction, and no theorem assumes the conclusion it establishes. The paper's self-citations are to Part II [44] for algorithms and to a same-group PhD thesis [50] for a background max-vol NMF result; neither carries the main uniqueness proofs. The pseudoinverse slip in the proof of Theorem 5.14 is a fixable typographical error: taking x0 = (U3^dagger)^T t gives (U3)^T x0 = t, restoring the non-vanishing polynomial argument; it is not a circular step. Likewise, the practical caveat that the procedures need global solutions of nonconvex min-volume problems is deferred to Part II and is a scope limitation, not a circularity. The paper is not self-consistent in every proof detail, but the central theoretical claims have independent content and do not reduce by construction to their inputs. The score of 1 reflects only the presence of minor, non-load-bearing self-citations; no significant circularity was found.

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

The paper introduces no fitted parameters or new postulated entities; all free choices are structural assumptions (SSC or separability, rank conditions) that define the identifiability theorems. The main reliance is on standard results from the NMF literature. The only ad hoc item is the polynomial non-vanishing argument in Theorem 5.14, whose proof has a pseudoinverse typo.

assumptions (5)

standard math Theorem 2.6 (separability implies unique NMF up to permutation and scaling, from [6])
Used as a black box in the proof of Theorem 4.1 and Theorem A.1.
standard math Theorem 2.8 (min-vol NMF identifiability under the SSC, from [17])
Used as a black box in Theorems 5.11, 5.16, 6.9, and 6.13.
domain assumption Factors U_i satisfy the SSC or separability (Definitions 2.5 and 2.7) and are normalized with U_i^T e = e
This is the paper's main hypothesis; if a factor is not separable or SSC, the volume-minimization identifiability arguments do not apply.
domain assumption Core tensor G# has specified slice or unfolding rank conditions, such as rank(G#_(3)) = r3 = r1 r2 or existence of maximum-rank slices
These conditions are required for the core rank to match the factor ranks and for the reduced matrix factorization problems to be full-rank.
ad hoc to paper The polynomial P(x) = det(sum_k ((U#_3)^T x)_k G#_k) is not identically zero when the span of the G#_k has maximal rank (used in Theorem 5.14)
This fact is asserted through a pseudoinverse computation that contains a typo; the claim is plausible but not rigorously established in the printed proof.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Identifiability of Nonnegative Tucker Decompositions -- Part I: Theory." pith.science (2026). https://pith.science/paper/AUEJZWLW

@misc{pith2026250512713,
  author       = {Pith},
  title        = {Pith review of: Identifiability of Nonnegative Tucker Decompositions -- Part I: Theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AUEJZWLW}},
  note         = {Machine review of arXiv:2505.12713}
}

read the original abstract

Tensor decompositions have become a central tool in data science, with applications in areas such as data analysis, signal processing, and machine learning. A key property of many tensor decompositions, such as the canonical polyadic decomposition, is identifiability: the factors are unique, up to trivial scaling and permutation ambiguities. This allows one to recover the groundtruth sources that generated the data. The Tucker decomposition (TD) is a central and widely used tensor decomposition model. However, it is in general not identifiable. In this paper, we study the identifiability of the nonnegative TD (nTD). By adapting and extending identifiability results of nonnegative matrix factorization (NMF), we provide uniqueness results for nTD. Our results require the nonnegative matrix factors to have some degree of sparsity (namely, satisfy the separability condition, or the sufficiently scattered condition), while the core tensor only needs to have some slices (or linear combinations of them) or unfoldings with full column rank (but does not need to be nonnegative). Under such conditions, we derive several procedures, using either unfoldings or slices of the input tensor, to obtain identifiable nTDs by minimizing the volume of unfoldings or slices of the core tensor.

Figures

Figures reproduced from arXiv: 2505.12713 by the authors.

**Figure 2.** Visualization of p-SSC in the case r = 3 on the plane {x ∈ R r | e ⊤x = 1}. On the left, the cones Cp with 1 < p < √ r − 1 and Cq with √ r − 1 < q < √ r. On the right, the cone C ≡ C√ r−1 and the nonnegative orthant R r + ≡ C1. Intuitively, p-SSC of U requires cone(U ⊤) to contain a larger cone than the one required by SSC1 when p < √ r − 1. For p = 1, p-SSC is equivalent to separability because C1 = R r +. For 29 … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 55 canonical work pages

[1]

SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

Abdolali, M., Gillis, N.: Simplex-structured matrix factorization: Sparsity-based identifiability and provably correct algorithms. SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

work page 2021
[2]

Journal of the American Statistical Association pp

Agterberg, J., Zhang, A.R.: Estimating higher-order mixed memberships via theℓ2,inf tensor per- turbation bound. Journal of the American Statistical Association pp. 1–11 (2024)

work page 2024
[3]

Advances in Neural Information Processing Systems21 (2008)

Airoldi, E.M., Blei, D., Fienberg, S., Xing, E.: Mixed membership stochastic blockmodels. Advances in Neural Information Processing Systems21 (2008)

work page 2008
[4]

In: Advances in Neural Information Processing Systems, vol

Anandkumar, A., Hsu, D.J., Janzamin, M., Kakade, S.M.: When are overcomplete topic models identifiable? uniqueness of tensor Tucker decompositions with structured sparsity. In: Advances in Neural Information Processing Systems, vol. 26 (2013)

work page 2013
[5]

In: International Conference on Machine Learning, vol

Arora, S., Ge, R., Halpern, Y., Mimno, D., Moitra, A., Sontag, D., Wu, Y., Zhu, M.: A practical algorithm for topic modeling with provable guarantees. In: International Conference on Machine Learning, vol. 28, pp. 280–288 (2013)

work page 2013
[6]

In: ACM Symposium on Theory of Computing, pp

Arora, S., Ge, R., Kannan, R., Moitra, A.: Computing a nonnegative matrix factorization–provably. In: ACM Symposium on Theory of Computing, pp. 145–162 (2012)

work page 2012
[7]

available online, february 3 (2015)

Bader, B.W., Kolda, T.G., et al.: MATLAB tensor toolbox version 2.6. available online, february 3 (2015)

work page 2015
[8]

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing11(3), 701–712 (2018)

Bai, X., Xu, F., Zhou, L., Xing, Y., Bai, L., Zhou, J.: Nonlocal similarity based nonnegative tucker decomposition for hyperspectral image denoising. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing11(3), 701–712 (2018)

work page 2018

Show all 56 references

[9]

Cambridge University Press, forthcoming (2025)

Ballard, G., Kolda, T.G.: Tensor Decompositions for Data Science. Cambridge University Press, forthcoming (2025)

2025
[10]

Mathematics of Operations Research26(2), 193–205 (2001)

Ben-Tal, A., Nemirovski, A.: On polyhedral approximations of the second-order cone. Mathematics of Operations Research26(2), 193–205 (2001)

2001
[11]

In: Proc

Boardman, J.W., Kruse, F.A., Green, R.O.: Mapping target signatures via partial unmixing of AVIRIS data. In: Proc. Summary JPL Airborne Earth Science Workshop, Pasadena, CA, pp. 23–26 (1995)

1995
[12]

John Wiley & Sons (2009)

Cichocki, A., Zdunek, R., Phan, A.H., Amari, S.i.: Nonnegative matrix and tensor factorizations: applications to exploratory multi-way data analysis and blind source separation. John Wiley & Sons (2009)

2009
[13]

IEEE Transactions on Geo- science and Remote Sensing32(3), 542–552 (1994)

Craig, M.D.: Minimum-volume transforms for remotely sensed data. IEEE Transactions on Geo- science and Remote Sensing32(3), 542–552 (1994)

1994
[14]

In: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp

Ding, C., Li, T., Peng, W., Park, H.: Orthogonal nonnegative matrix t-factorizations for clustering. In: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 126–135 (2006) 38

2006
[15]

Linear Algebra and its Applications513, 342–375 (2017)

Domanov, I., De Lathauwer, L.: Canonical polyadic decomposition of third-order tensors: Relaxed uniqueness conditions and algebraic algorithm. Linear Algebra and its Applications513, 342–375 (2017)

2017
[16]

1141–1148 (2004)

Donoho, D., Stodden, V.: When does non-negative matrix factorization give a correct decomposition into parts? In: Advances in Neural Information Processing Systems (NIPS), pp. 1141–1148 (2004)

2004
[17]

IEEE Signal Processing Letters25(3), 328–332 (2018)

Fu, X., Huang, K., Sidiropoulos, N.D.: On identifiability of nonnegative matrix factorization. IEEE Signal Processing Letters25(3), 328–332 (2018)

2018
[18]

IEEE Signal Processing Magazine 36(2), 59–80 (2019)

Fu, X., Huang, K., Sidiropoulos, N.D., Ma, W.K.: Nonnegative matrix factorization for signal and data analytics: Identifiability, algorithms, and applications. IEEE Signal Processing Magazine 36(2), 59–80 (2019)

2019
[19]

IEEE Transactions on Pattern Analysis and Machine Intelligence41(5), 1056–1071 (2019)

Fu, X., Huang, K., Sidiropoulos, N.D., Shi, Q., Hong, M.: Anchor-free correlated topic modeling. IEEE Transactions on Pattern Analysis and Machine Intelligence41(5), 1056–1071 (2019)

2019
[20]

IEEE Transactions on Signal Processing 64(23), 6254–6268 (2016)

Fu, X., Huang, K., Yang, B., Ma, W.K., Sidiropoulos, N.D.: Robust volume minimization-based matrix factorization for remote sensing and document clustering. IEEE Transactions on Signal Processing 64(23), 6254–6268 (2016)

2016
[21]

IEEE Transactions on Signal Processing63(9), 2306–2320 (2015)

Fu, X., Ma, W.K., Huang, K., Sidiropoulos, N.D.: Blind separation of quasi-stationary sources: Exploiting convex geometry in covariance domain. IEEE Transactions on Signal Processing63(9), 2306–2320 (2015)

2015
[22]

IEEE Signal Processing Magazine37(5), 78–94 (2020)

Fu, X., Vervliet, N., De Lathauwer, L., Huang, K., Gillis, N.: Computing large-scale matrix and tensor decomposition with structured factors: A unified nonconvex optimization perspective. IEEE Signal Processing Magazine37(5), 78–94 (2020)

2020
[23]

Society for Industrial and Applied Mathematics, Philadelphia, PA (2020)

Gillis, N.: Nonnegative Matrix Factorization. Society for Industrial and Applied Mathematics, Philadelphia, PA (2020)

2020
[24]

IEEE Signal Processing Letters31, 1610–1614 (2024)

Gillis, N., Luce, R.: Checking the sufficiently scattered condition using a global non-convex opti- mization software. IEEE Signal Processing Letters31, 1610–1614 (2024)

2024
[25]

European Journal of Remote Sensing49(1), 587–598 (2016)

Hassanzadeh, S., and, A.K.: Compression and noise reduction of hyperspectral images using non- negative tensor decomposition and compressed sensing. European Journal of Remote Sensing49(1), 587–598 (2016)

2016
[26]

Advances in Neural Information Processing Systems29 (2016)

Huang, K., Fu, X., Sidiropoulos, N.D.: Anchor-free correlated topic modeling: Identifiability and algorithm. Advances in Neural Information Processing Systems29 (2016)

2016
[27]

IEEE Transactions on Signal Processing64(19), 5052– 5065 (2016)

Huang, K., Sidiropoulos, N.D., Liavas, A.P.: A flexible and efficient algorithmic framework for constrained matrix and tensor factorization. IEEE Transactions on Signal Processing64(19), 5052– 5065 (2016)

2016
[28]

IEEE Transactions on Signal Processing62(1), 211– 224 (2013)

Huang, K., Sidiropoulos, N.D., Swami, A.: Non-negative matrix factorization revisited: Uniqueness and algorithm for symmetric decomposition. IEEE Transactions on Signal Processing62(1), 211– 224 (2013)

2013
[29]

In: ACM-SIAM Symposium on Discrete Algorithms, pp

Koiran, P.: An efficient uniqueness theorem for overcomplete tensor decomposition. In: ACM-SIAM Symposium on Discrete Algorithms, pp. 1909–1932 (2025)

2025
[30]

computational complexity32(2), 8 (2023)

Koiran, P., Saha, S.: Absolute reconstruction for sums of powers of linear forms: degree 3 and beyond. computational complexity32(2), 8 (2023)

2023
[31]

URLhttps://arxiv.org/abs/2403.00643

Koiran, P., Saha, S.: Undercomplete decomposition of symmetric tensors in linear time, and smoothed analysis of the condition number (2024). URLhttps://arxiv.org/abs/2403.00643

2024 arXiv
[32]

Theoretical Computer Science1037, 115,159 (2025)

Koiran, P., Saha, S.: Complete decomposition of symmetric tensors in linear time and polylogarith- mic precision. Theoretical Computer Science1037, 115,159 (2025)

2025
[33]

SIAM Rev.51, 455–500 (2009)

Kolda, T., Bader, B.: Tensor decompositions and applications. SIAM Rev.51, 455–500 (2009)

2009
[34]

In: IEEE International Conference on Acoustics, Speech and Signal Processing (2025)

Kolomvakis, C., Vandaele, A., Gillis, N.: Boolean matrix tri-factorization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (2025)

2025
[35]

Journal of Machine Learning Research20(26), 1–6 (2019) 39

Kossaifi, J., Panagakis, Y., Anandkumar, A., Pantic, M.: Tensorly: Tensor learning in python. Journal of Machine Learning Research20(26), 1–6 (2019) 39

2019
[36]

URL https://arxiv.org/abs/2411.14344

Kothari, P.K., Moitra, A., Wein, A.S.: Overcomplete tensor decomposition via Koszul-Young flat- tenings (2024). URL https://arxiv.org/abs/2411.14344

2024
[37]

Linear Algebra and its Applications18(2) (1977)

Kruskal, J.B.: Three-way arrays: rank and uniqueness of trilinear decompositions, with application to arithmetic complexity and statistics. Linear Algebra and its Applications18(2) (1977)

1977
[38]

IEEE Transactions on Signal Processing68, 3400–3410 (2020)

Leplat, V., Gillis, N., Ang, A.M.: Blind audio source separation with minimum-volume beta- divergence NMF. IEEE Transactions on Signal Processing68, 3400–3410 (2020)

2020
[39]

IEEE Transactions on Geoscience and Remote Sensing53(10), 5530–5546 (2015)

Lin, C.H., Ma, W.K., Li, W.C., Chi, C.Y., Ambikapathi, A.: Identifiability of the simplex volume minimization criterion for blind hyperspectral unmixing: The no-pure-pixel case. IEEE Transactions on Geoscience and Remote Sensing53(10), 5530–5546 (2015)

2015
[40]

Pattern Recognition148, 110,207 (2024)

Liu, Q., Lu, L., Chen, Z.: Non-negative tucker decomposition with graph regularization and smooth constraint for clustering. Pattern Recognition148, 110,207 (2024)

2024
[41]

URLhttps://arxiv.org/ abs/2104.08580

Marmoret, A., Cohen, J.E., Bertin, N., Bimbot, F.: Uncovering audio patterns in music with nonnegative Tucker decomposition for structural segmentation (2021). URLhttps://arxiv.org/ abs/2104.08580

2021 arXiv
[42]

Measurement Science Review20(3), 126–138 (2020)

Rošt’áková, Z., Rosipal, R., Seifpour, S., Trejo, L.J.: A comparison of non-negative tucker decom- position and parallel factor analysis for identification and measurement of human eeg rhythms. Measurement Science Review20(3), 126–138 (2020)

2020
[43]

Theses, Ecole normale supérieure de lyon - ENS LYON (2023)

Saha, S.: Algebraic and Numerical Algorithms for Symmetric Tensor Decompositions. Theses, Ecole normale supérieure de lyon - ENS LYON (2023)

2023
[44]

In preparation

Saha, S., Gillis, N.: Identifiability of nonnegative Tucker decompositions—Part II: Algorithms and applications (2025). In preparation

2025
[45]

IEEE Transactions on signal processing 65(13), 3551–3582 (2017)

Sidiropoulos, N.D., De Lathauwer, L., Fu, X., Huang, K., Papalexakis, E.E., Faloutsos, C.: Tensor decomposition for signal processing and machine learning. IEEE Transactions on signal processing 65(13), 3551–3582 (2017)

2017
[46]

In: International Society for Music Information Retrieval (ISMIR), pp

Smith, J.B., Kawasaki, Y., Goto, M.: Unmixer: An interface for extracting and remixing loops. In: International Society for Music Information Retrieval (ISMIR), pp. 824–831 (2019)

2019
[47]

In: IEEE International Conference on Acoustics, Speech and Signal Processing (2023)

Sun, Y., Huang, K.: Volume-regularized nonnegative Tucker decomposition with identifiability guarantees. In: IEEE International Conference on Acoustics, Speech and Signal Processing (2023)

2023
[48]

Van Loan, C.F., Pitsianis, N.: Approximation with Kronecker Products, pp. 293–314. Springer, Netherlands, Dordrecht (1993)

1993
[49]

Vervliet, N., Debals, O., Sorber, L., Van Barel, M., De Lathauwer, L.: Tensorlab 3.0 (2016)

2016
[50]

Vu Thanh, O.: Low-rank matrix factorizations with volume-based constraints and regularizations. Ph.D. thesis, University of Mons (2024)

2024
[51]

In: International Joint Conference on Artificial Intelligence (2011)

Wang, H., Nie, F., Huang, H., Makedon, F.: Fast nonnegative matrix tri-factorization for large-scale data co-clustering. In: International Joint Conference on Artificial Intelligence (2011)

2011
[52]

Advances in Neural Information Processing Systems32 (2019)

Wang, M., Zeng, Y.: Multiway clustering via tensor block models. Advances in Neural Information Processing Systems32 (2019)

2019
[53]

URLhttps://arxiv.org/abs/2101.06827

Yin, W., Qu, Y., Ma, Z., Liu, Q.: HyperNTF: A hypergraph regularized nonnegative tensor factor- ization for dimensionality reduction (2022). URLhttps://arxiv.org/abs/2101.06827

2022 arXiv
[54]

Remote Sensing13(15) (2021)

Zare, M., Helfroush, M.S., Kazemi, K., Scheunders, P.: Hyperspectral and multispectral image fusion using coupled non-negative tucker tensor decomposition. Remote Sensing13(15) (2021)

2021
[55]

In: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp

Zhang, Y., Yeung, D.Y.: Overlapping community detection via bounded nonnegative matrix tri- factorization. In: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 606–614 (2012)

2012
[56]

IEEE Transactions on Image Processing24(12), 4990–5003 (2015) 40

Zhou, G., Cichocki, A., Zhao, Q., Xie, S.: Efficient nonnegative Tucker decompositions: Algorithms and uniqueness. IEEE Transactions on Image Processing24(12), 4990–5003 (2015) 40

2015

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

Abdolali, M., Gillis, N.: Simplex-structured matrix factorization: Sparsity-based identifiability and provably correct algorithms. SIAM Journal on Mathematics of Data Science3(2), 593–623 (2021)

work page 2021

[2] [2]

Journal of the American Statistical Association pp

Agterberg, J., Zhang, A.R.: Estimating higher-order mixed memberships via theℓ2,inf tensor per- turbation bound. Journal of the American Statistical Association pp. 1–11 (2024)

work page 2024

[3] [3]

Advances in Neural Information Processing Systems21 (2008)

Airoldi, E.M., Blei, D., Fienberg, S., Xing, E.: Mixed membership stochastic blockmodels. Advances in Neural Information Processing Systems21 (2008)

work page 2008

[4] [4]

In: Advances in Neural Information Processing Systems, vol

Anandkumar, A., Hsu, D.J., Janzamin, M., Kakade, S.M.: When are overcomplete topic models identifiable? uniqueness of tensor Tucker decompositions with structured sparsity. In: Advances in Neural Information Processing Systems, vol. 26 (2013)

work page 2013

[5] [5]

In: International Conference on Machine Learning, vol

Arora, S., Ge, R., Halpern, Y., Mimno, D., Moitra, A., Sontag, D., Wu, Y., Zhu, M.: A practical algorithm for topic modeling with provable guarantees. In: International Conference on Machine Learning, vol. 28, pp. 280–288 (2013)

work page 2013

[6] [6]

In: ACM Symposium on Theory of Computing, pp

Arora, S., Ge, R., Kannan, R., Moitra, A.: Computing a nonnegative matrix factorization–provably. In: ACM Symposium on Theory of Computing, pp. 145–162 (2012)

work page 2012

[7] [7]

available online, february 3 (2015)

Bader, B.W., Kolda, T.G., et al.: MATLAB tensor toolbox version 2.6. available online, february 3 (2015)

work page 2015

[8] [8]

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing11(3), 701–712 (2018)

Bai, X., Xu, F., Zhou, L., Xing, Y., Bai, L., Zhou, J.: Nonlocal similarity based nonnegative tucker decomposition for hyperspectral image denoising. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing11(3), 701–712 (2018)

work page 2018