Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Random Tensors and their Normal Distributions

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A closed-form characteristic function for Gaussian random tensors, proved via per-mode covariance matrices.

desk verdict Correct after a change of convention, but as written the central theorem contradicts the paper's own definition; a useful notation exercise, not a reliable reference. read the letter →

arxiv 1908.01131 v3 pith:BEXR2BCA submitted 2019-08-03 math.ST stat.TH

Changqing Xu , Kaijie Xu This is my paper · ORCID
classification math.STstat.TH MSC 53A4515A69
keywords randomtensorGaussiannormaldistributionstandardcharacteristicfunctionmomentsmultilinearproduct
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces Gaussian random tensors as a higher-order version of matrix normal distributions. A standard normal tensor is one whose fibers are standard normal along every mode, and the paper proves this is equivalent to a tensor with independent standard normal entries and to a tensor whose vectorization is standard normal. A general Gaussian tensor is defined as an affine transformation of a standard normal tensor, with one covariance matrix per mode. The central result gives the characteristic function as $\varphi_X(T) = \exp\{i\langle T, \mu\rangle - \tfrac{1}{2}\langle T^{[2]}, \Sigma\rangle\}$, where $\Sigma$ is the tensor product of the mode covariance matrices. This is the tensor analogue of the matrix normal characteristic function and would let statisticians work with high-order normal arrays without flattening them.

What carries the argument

The carrying object is the mode-wise tensor product action, written $A \times_1 U_1 \times_2 \cdots \times_m U_m$, together with the square tensor $T^{[2]}$ and the inner product identity $\langle T, ZU \rangle = \langle UT, Z\rangle$ that lets covariance tensors be pulled out one mode at a time. Lemma 4.10 identifies the mode-wise action with multiplication by the tensor product $U = U_1 \times \cdots \times U_m$, and Lemma 4.11 supplies the adjoint move used in the characteristic-function proof. The final object that carries the formula is $\Sigma = \Sigma_1 \times \cdots \times \Sigma_m$, a $2m$-order tensor built by multiplying one covariance factor per mode.

What would settle it

A concrete check: for a $2 \times 2 \times 2$ random tensor, test whether the fiber-wise standard-normal conditions (all mode-1, mode-2, and mode-3 fibers standard normal) force the eight entries to be independent standard normals. Any joint distribution that satisfies the fiber conditions but is not iid, or any affine Gaussian tensor whose second-moment tensor is not $\Sigma_1 \times \Sigma_2 \times \Sigma_3$, would invalidate the equivalence behind Theorem 4.12.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Gaussian distribution on an $m$-order tensor is specified by a mean tensor plus one covariance matrix per mode, and that the resulting characteristic function has the tensor-product form $\varphi_X(T) = \exp\{i\langle T, \mu\rangle - \tfrac{1}{2}\langle T^{[2]}, \Sigma_1 \times \cdots \times \Sigma_m\rangle\}$. The proof treats a Gaussian tensor as $X = \mu + ZU$ for a standard normal tensor $Z$ and a mode-wise product of matrices $U_k$ with $\Sigma_k = U_k^\top U_k$; Lemma 4.10 and Lemma 4.11 move the covariance tensors across the inner product, reducing the calculation to the known characteristic function of a standard normal tensor. Along the way the paper gives equivalent characterizations of standard normal tensors and derives tensor-form moments for Gaussian matrices, including $m_2[X] = \Sigma_1 \times \Sigma_2$.

Load-bearing premise

The formula is proved for tensors that are affine images of a standard normal tensor, but the paper assumes without a general-order proof that this class coincides with tensors whose fibers are Gaussian along every mode.

Editorial extensions

If this is right

  • If the formula holds, every Gaussian tensor is described by $m$ covariance matrices of sizes $n_k \times n_k$ rather than one huge covariance matrix, making high-order models parsimonious.
  • Moments of Gaussian tensors can be computed by differentiating the characteristic function in tensor form, avoiding the index ambiguity of vectorized derivatives.
  • Simulation of a Gaussian tensor reduces to drawing an iid standard normal array and applying mode-wise linear transformations, one per mode.
  • The standard-normal equivalences in Theorem 4.2 give a simple route to checking normality: a tensor is standard normal exactly when every fiber in every mode is standard normal.

Reading between the lines

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

  • If the fiber-wise and affine classes differ, the characteristic-function formula still defines a coherent 'affine Gaussian tensor' class; the paper's results would remain valid with that narrower definition.
  • The same adjoint inner-product trick could be applied to any tensor distribution built from an affine transformation of a base law, yielding characteristic functions for tensor elliptically contoured or skew-normal families.
  • The moment formulas suggest a tensor analogue of the covariance test: in a Gaussian tensor model, the empirical second-moment tensor should be close to a tensor product, so a statistically significant deviation is a direct lack-of-fit test.
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.

Referee Report

4 major / 5 minor

Summary. The manuscript develops a tensor-valued analog of the matrix normal distribution. It introduces a standard normal tensor, defines a Gaussian tensor as an affine image X = µ + Z ×1 U1 ... ×m Um of a standard normal tensor, and states several equivalent characterizations in terms of fiber-wise normality, mode-k flattenings, and characteristic functions. The main result, Theorem 4.12, claims that the characteristic function of a Gaussian tensor is exp{i⟨T, µ⟩ − 1/2⟨T^[2], Σ⟩} with Σ = Σ1 × ... × Σm. Sections 2 and 3 set up tensor calculus and matrix-normal facts that are then extended to tensors.

Significance. If carried out rigorously, the tensor-product form of the characteristic function would be a clean coordinate-free summary of the tensor normal distribution and a useful reference for high-order extensions of matrix normal theory. The elementary parts, such as the characterization of SND tensors in Theorem 4.2 and the computation of the SND characteristic function in Lemma 4.8, are mostly correct and clearly presented. However, the paper's positive contribution is currently limited: the key equivalence statements are either unproved or asserted, and the main characteristic-function theorem is entangled with an inconsistent covariance convention. The paper therefore does not yet provide a reliable foundation for the claimed results.

major comments (4)
  1. [§4, Eq. (4.5), Theorem 4.5, Theorem 4.12 proof] Definition (4.5) requires U_k U_k^⊤ = Σ_k, but Theorem 4.5 states Σ_k = U_k^⊤ U_k, and the proof of Theorem 4.12 explicitly uses 'Σ_s = U_s^⊤ U_s' in the final identity. For nonsymmetric U_k these are different factorizations: under (4.5) the second-order structure of X is (U_1 U_1^⊤) × ... × (U_m U_m^⊤), while the algebra in the proof of Theorem 4.12 computes (U_1^⊤ U_1) × ... × (U_m^⊤ U_m). Consequently the displayed characteristic function does not follow from the paper's own definition as written. The authors must reconcile the convention, either by changing (4.5) and recomputing the covariance of the affine construction or by correcting the proof to use U_k U_k^⊤ consistently.
  2. [§4, paragraph before Theorem 4.4] The fiber-wise definition of a Gaussian tensor uses an undefined scalar λ_k and writes A^(k)(:, j) ∼ N_{n_k, m_k}(M^(k)(:, j), λ_k Σ_k); the notation N_{n_k, m_k} with a single covariance matrix is not defined for matrix normal distributions. Without fixing these objects, the equivalence between the fiber-wise and affine definitions (Theorems 4.4 and 4.9) has no precise content.
  3. [Theorems 4.6 and 4.9] Theorem 4.6 is stated as an 'iff' result, but it essentially restates the definition of a Gaussian tensor, and the converse direction is not proved. Theorem 4.9, which is supposed to supply the converse from the unfolded matrices, is dismissed with 'can be proved by the same technique as in the case of order three.' Since Theorem 4.12 is proved only for the affine definition, the paper needs an actual proof that the fiber-wise or unfolded class coincides with the affine class before the characteristic-function result can be claimed for all Gaussian tensors.
  4. [Theorem 4.7] Theorem 4.7 claims an equivalence between tensor normality and the normality of every mode-k unfolding, but its proof is a one-line appeal to induction with no details. Because this equivalence is load-bearing for Theorem 4.9 and for the interpretation of the main definition, a real proof or a complete sketch should be supplied.
minor comments (5)
  1. [Theorem 3.10 and surrounding text] There is an unresolved equation reference '(??)' in the proof of Theorem 3.10, and several references to 'Lemma 3' should be 'Lemma 3.1' or 'Lemma 3.2' as appropriate.
  2. [Abstract and body] The text contains numerous typos, including 'initialised', 'dist ri-', 'matric es', and 'eq uiv-'; a careful proofreading pass is needed.
  3. [Theorem 4.4] Theorem 4.4 refers to Definition 4.1 for a general m-order tensor, but Definition 4.1 is stated only for third-order tensors; the general definition should be stated before the theorem.
  4. [Proposition 2.1] Item (1) of Proposition 2.1, as written, appears dimensionally inconsistent; the intended product notation should be clarified.
  5. [References] Reference [30] is incomplete, lacking volume and page numbers, and should be completed for the reader's convenience.

Circularity Check

1 steps flagged · score 1.0 of 10

No significant circularity: the central characteristic-function formula is derived from the affine definition plus the standard SND characteristic function, with only a tautological restatement (Theorem 4.6) and a non-circular covariance-convention gap in the proof.

  1. self definitional [Section 4, Definition (4.5) and Theorem 4.6]
    "Definition (4.5): X = µ + Z ×1 U1 ×2 U2 ×3 . . . ×m Um ... where Uk ∈ R^{nk×nk} satisfies UkU_k^T = Σ_k. Theorem 4.6: A random tensor Y ∈ T_{m;n} follows a normal distribution Y ∼ N_{m;n}(M, Σ_1, . . . , Σ_m) iff there exist some matrices Uk (k ∈ [m]) such that Y = X ×1 U1 ×2 U2 . . . ×m Um and X obeys a standard Gaussian distribution."

    The theorem's 'if and only if' reproduces the defining property almost verbatim: the definition already declares that a Gaussian tensor is exactly an affine multilinear transformation of an SND tensor. No new condition, independent characterization, or proof is supplied; the theorem is the definition restated. Any later appeal to Theorem 4.6 as an equivalence result would therefore be an appeal to the definition itself. The defect is minor because the characteristic-function derivation in Theorem 4.12 uses Definition (4.5) directly rather than Theorem 4.6.

full rationale

The paper's central derivation is self-contained against its own definition. Starting from Definition (4.5) (an affine multilinear transform of an SND tensor) and Lemma 4.8 (the characteristic function of an SND tensor), the proof of Theorem 4.12 computes φ_X(T) = exp{i<T,µ>} φ_Z(UT) and reduces <UT,UT> to <T^{[2]}, Σ> using the tensor-product identities of Lemmas 4.10 and 4.11. No parameter is fitted to data and no external benchmark is needed; the calculation is a direct application of the definition. The only genuine circular step is Theorem 4.6, which restates Definition (4.5) as an iff theorem without adding content; this is a mild self-definitional redundancy, not a load-bearing prediction. The citation to the authors' own [30] for the commutation tensor is parameter-free and elementary, and it is not used in the main Gaussian-tensor argument. Two non-circular weaknesses are noted: Theorem 4.9 is asserted without proof ('can be proved by the same technique as in the case of order three'), and the proof of Theorem 4.12 switches from UkU_k^T = Σ_k in Definition (4.5) to U_k^T U_k = Σ_k in the final identity. The latter makes the theorem unsupported as written for non-symmetric Uk, but this is a mathematical gap, not a circular reduction of the claimed result to its own inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 1 invented entities

All parameters in the paper are distributional parameters (mu, Sigma_k) inherited from the matrix normal literature; none are fitted to data. The main new object, the 'Gaussian tensor', duplicates the tensor normal distribution of [19] and [23], so it is not an independently evidenced invention.

assumptions (4)
  • standard math Background results on matrix normal distributions and their moments from Kollo and von Rosen [16] and Bilodeau and Brenner [2] are accepted as true.
    Invoked in Lemma 3.5 (proof from [2]) and Corollary 3.11 (comparison to Theorem 2.2.7 in [16]).
  • standard math For any positive semidefinite Sigma_k, there exists a square-root factor U_k with U_k^T U_k = Sigma_k.
    Used in Definition (4.5) and the proof of Theorem 4.12; this is standard PSD square-root existence.
  • domain assumption The distribution of a random tensor is identified with the joint distribution of its entries, and the fiber-wise conditions in Definition 4.1 and the paragraph before Theorem 4.4 characterize the standard and general Gaussian tensor.
    The equivalence of this fiber-wise definition with the affine definition (4.5) is assumed and only sketched via Theorems 4.7 and 4.9.
  • standard math The commutation tensor K_{m,n} = I_m x_{(2,3)} I_n and its algebraic properties from Xu, He, and Lin [30] are accepted.
    Used in Section 2 for transposes and derivatives; [30] is a self-citation of one of the authors but the commutation tensor is standard.
invented entities (1)
  • Gaussian tensor (random tensor with normal distribution)
    purpose: To extend the matrix normal distribution to higher-order tensors using tensor products instead of Kronecker products.
    The distribution is not new: it is the tensor normal or multilinear normal distribution already defined in [6], [19], and [23]. The paper provides no falsifiable prediction or new probabilistic object; it re-expresses the existing distribution in tensor notation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Random Tensors and their Normal Distributions." pith.science (2026). https://pith.science/paper/BEXR2BCA

@misc{pith2026190801131,
  author       = {Pith},
  title        = {Pith review of: Random Tensors and their Normal Distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEXR2BCA}},
  note         = {Machine review of arXiv:1908.01131}
}
read the original abstract

The main purpose of this paper is to introduce the random tensor with normal distribution, which promotes the matrix normal distribution to a higher order case. Some basic knowledge on tensors are introduced before we focus on the random tensors whose entries follow normal distribution. The random tensor with standard normal distribution(SND) is introduced as an extension of random normal matrices. As a random multi-array deduced from an affine transformation on a SND tensor, the general normal random tensor is initialised in the paper. We then investigate some equivalent definitions of a normal tensor and present the description of the density function, characteristic function, moments, and some other functions related to a random matrix. A general form of an even-order multi-variance tensor is also introduced to tackle a random tensor. Finally some equivalent definitions for the tensor normal distribution are described.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

  1. [2]

    Bilodeau and D

    M. Bilodeau and D. Brenner, Theory of Multivariate Statistics , Springer-Verlag, New York, 1999

    work page 1999

  2. [23]

    Ohlson, M

    M. Ohlson, M. R. Ahmad, D. von Rosen, The multilinear normal distri- bution: Introduction and some basic properties , Journal of Multivariate Analysis, 113(2013) 37-47

    work page 2013

  3. [19]

    A. M. Manceura, and P. Dutilleul, Maximum likelihood estimation for the tensor normal distribution: Algorithm, minimum sample size, and empirical bias and dispersion , Journal of Computational and Applied Math., 239(2013) 37-49

    work page 2013

  4. [6]

    P. J. Basser and S. Pajevic, A normal distribution for tensor-valued random variables: applications to diffusion tensor MRI , IEEE Trans- actions on Medical Imaging, 22(2003) 785-794

    work page 2003

  5. [1]

    Gleich, and Lek-Heng Lim, The spacey ran- dom walk: a stochastic process for higher-order data , SIAM Review, 59(2017), pp

    Austin Benson, David F. Gleich, and Lek-Heng Lim, The spacey ran- dom walk: a stochastic process for higher-order data , SIAM Review, 59(2017), pp. 321-345

    work page 2017

  6. [3]

    Chatterjee and N

    U. Chatterjee and N. Chatterjee, Vector & Tensor Analysis , Academic Publishers, NY, 2010

    work page 2010

  7. [4]

    Dustin Cartwright, Bernd Sturmfels, The number of eigenvalues of a tensor, Linear Algebra Appl., 438(2013), pp. 942-952

    work page 2013

  8. [5]

    J. D. Carroll and J. Chang, Analysis of individual differences in mul- tidimensional scaling via an n-way generalisation of ’Ecka rtYoung’ de- composition, Psychometrika, 35(1970), pp. 283-319

    work page 1970

Show all 30 references
  1. [7]

    Comon, G

    P. Comon, G. Golub, L.-H. Lim, and B. Mourrain, Symmetric tensors and symmetric tensor rank , SIAM. J. Matrix Analysis and Applications, 30(2008) 1254-1279

    2008

  2. [8]

    R. D. Cook, B. Li and F. Charomonte, Envelope models for parsi- monious and efficient multivariate linear regression , Statistica Sinica, 20(2010) 927-1010

    2010

  3. [9]

    R. D. Cook, I. S. Helland and Z. Su, Envelope and partial least squares regression, J. R. Stat. Soc. Ser. B. Stat. Methodol., 75(2013) 851-877

    2013

  4. [10]

    Couillet and M

    R. Couillet and M. Debbah, Random matrix methods for wireless com- munications, Cambridge Univ Press, 2011

    2011

  5. [11]

    P. J. Forrester, Log-gases and random matrices , Number 34, Princeton Univ Press, 2010

    2010

  6. [12]

    S. Hu, Z. Huang, C. Ling and L. Qi, On determinants and eigenvalue theory of tensors , Journal of Symbolic Computation, 50(2013) 508-531

    2013

  7. [13]

    Jiang, L.-H

    X. Jiang, L.-H. Lim, Y. Yao and Y. Ye, Statistical ranking and combi- natorial Hodge theory , Math. Program., Ser. B 127(2011) 203-244

    2011

  8. [14]

    N. L. Johnson, S. Kotz and N. Balakrishnan, Discrete Multivariate Distributions, Wiley, New York, 1997. 27

    1997

  9. [15]

    T. G. Kolda, B. W. Bader, Tensor Decompositions and Applications , SIAM Review, 2009

    2009

  10. [16]

    Kollo and D

    T. Kollo and D. von Rosen, Advanced Multivariate Statistics with Ma- trices, Springer, New York, 2005

    2005

  11. [17]

    S. Kotz, N. Balakrishnan and N. L. Johnson, Continuous Multivariate Distributions, Wiley, New York, 2000

    2000

  12. [18]

    L. -H. Lim, Singular values and eigenvalues of tensors: A variational approach, Proc. of 1st IEEE Int’l Workshop on Computational Ad- vances in Multi-Sensor Adaptive Processing (CAMSAP), 2005 , pp. 129- 132

    2005

  13. [20]

    M. L. Mehta, A non-linear differential equation and a Fredholm deter- minant, J. de Phys. I France 2(1992): 1721-1729

    1992

  14. [21]

    M. L. Mehta, Random matrices, volume 142, Academic Press, 2004

    2004

  15. [22]

    Mezzadri and N

    F. Mezzadri and N. C. Snaith, Recent perspectives in random matrix theory and number theory , Cambridge Univ Press, 2005

    2005

  16. [24]

    Qi, Eigenvalues of a real supersymmetric tensor , Journal of Symbolic Computation, 40(2005): 1302-1324

    L. Qi, Eigenvalues of a real supersymmetric tensor , Journal of Symbolic Computation, 40(2005): 1302-1324

    2005

  17. [25]

    Qi, Symmetric nonnegative tensors and copositive tensors , Linear Algebra Appl., 439(2013): 228-238

    L. Qi, Symmetric nonnegative tensors and copositive tensors , Linear Algebra Appl., 439(2013): 228-238

    2013

  18. [26]

    Qi, Hankel tensors: Associated Hankel matrices and Vandermond e decomposition, Communications in Mathematical Sciences, 13 (2015): 113-125

    L. Qi, Hankel tensors: Associated Hankel matrices and Vandermond e decomposition, Communications in Mathematical Sciences, 13 (2015): 113-125

    2015

  19. [27]

    Qi and Z

    L. Qi and Z. Luo, Tensor Analysis: Spectral theory and Special tensors , SIAM press, Philadelpha, 2017

    2017

  20. [28]

    Tao, Topics in Random Matrix Theory , Graduate studies in mathe- matics, American Mathematical Society, 2012

    T. Tao, Topics in Random Matrix Theory , Graduate studies in mathe- matics, American Mathematical Society, 2012

    2012

  21. [29]

    Weinberg, Gravitation and Cosmology: Principles and Applications of the General Theory of Relativity , Wiley, New York, 1972

    S. Weinberg, Gravitation and Cosmology: Principles and Applications of the General Theory of Relativity , Wiley, New York, 1972. 28

    1972

  22. [30]

    Xu, L.He and Z

    C. Xu, L.He and Z. Lin, Commutation matrices and com- mutation tensors , Linear and Multilinear Algebra, DOI: 10.1080/03081087.2018.1556242. 29

    2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.