Pith. sign in

REVIEW 2 major objections 5 minor 71 references

Separability of a matrix covariance can be tested by whitening with the separable MLE and checking sphericity, with an angular version that resists heavy tails.

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

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

T0 review · grok-4.5

2026-07-10 08:15 UTC pith:HDUNP2SB

load-bearing objection Solid high-d separability test with clean finite-sample level and a useful angular fix; consistency is real but conditional on open pseudo-MLE rates the authors flag themselves. the 2 major comments →

arxiv 2607.08388 v1 pith:HDUNP2SB submitted 2026-07-09 stat.ME math.STstat.TH

Testing Covariance Separability in High Dimensions

Tomas Masak , Marcus Mayrhofer , Una Radoji\v{c}i\'c This is my paper
classification stat.ME math.STstat.TH MSC 62H1562H1262F0562E20
keywords matrix-variate dataseparable covarianceKronecker covariancesphericity testhigh-dimensional testingelliptical distributionsmatrix normal MLEangular statistics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

When each observation is a matrix, an unstructured covariance is a huge tensor that is often impossible to estimate or invert. Separability reduces that tensor to a product of two ordinary matrices, but imposing the structure blindly can bias later analysis. This paper shows how to test the structure without ever forming the full covariance: estimate the separable factors by maximum likelihood, whiten the sample, and test whether the whitened data are spherical. Monte Carlo under the whitened null gives exact finite-sample level when the calibration law is correct, and the authors prove high-dimensional consistency against dense departures from separability. Projecting each whitened observation onto the unit sphere yields an angular version that keeps nearly the same power while staying closer to nominal level under heavy-tailed elliptical data calibrated as Gaussian. The procedures are cubic in sample size and come with pre-tabulated quantiles, so the cost of checking separability stays comparable to fitting the separable model itself.

Core claim

Within the elliptical family, covariance separability is equivalent to sphericity after whitening by a unique matrix-affine-equivariant separable estimator (the matrix-normal MLE). The resulting Monte Carlo test based on the unified John/Nagao/Ledoit–Wolf statistic therefore has exact finite-sample level under the null, and is consistent in the high-dimensional regime N proportional to pq against dense alternatives measured in Frobenius distance. An angular refinement that radially normalizes the whitened observations inherits the same level property and consistency under slightly stronger conditioning assumptions, while empirically reducing sensitivity to radial misspecification.

What carries the argument

Matrix whitening by the flip-flop MLE of the separable factors, after which John's, Nagao's and Ledoit–Wolf's sphericity statistics coincide (up to a constant absorbed by Monte Carlo); the angular version replaces the sample covariance of the whitened data by the spatial-sign covariance obtained after unit-sphere projection.

Load-bearing premise

Consistency rests on the unproven claim that the separable pseudo-MLE still converges, at a high-dimensional rate, when the true covariance is not separable.

What would settle it

Construct a dense non-separable family (fixed positive Frobenius gap after scaling by 1/√(pq)) for which the separable MLE fails to converge in Frobenius or operator norm when N ~ pq, or show that the angular test's empirical level collapses under a non-elliptical law while the Gaussian Monte Carlo quantiles stay fixed.

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

If this is right

  • Separability can be checked at roughly the cost of one separable fit, without ever storing the full pq-by-pq covariance.
  • When N is proportional to pq, dense non-separability is asymptotically detectable after MMLE whitening.
  • The angular test can be used under unknown elliptical tails with only a Gaussian Monte Carlo table, retaining nearly full power.
  • Pre-tabulated quantiles for each (p,q,N) make repeated testing and software packaging immediate.
  • Downstream matrix-variate procedures that rely on separability gain a reliable pre-test that respects the same computational budget.

Where Pith is reading between the lines

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

  • The same whitening-plus-sphericity reduction is likely usable for other Kronecker-type structural hypotheses (e.g., multi-way separability or core-shrinkage models) once an equivariant estimator under the null is available.
  • Closing the open high-dimensional rate for the pseudo-MLE under alternatives would immediately upgrade both consistency theorems from conditional to fully rigorous.
  • Because the angular statistic discards only radial scale, it may remain approximately pivotal for mild departures from ellipticity that still preserve spherical directions after whitening—an empirical regime worth mapping.
  • The eigenvalue-dispersion view of the statistic suggests natural sparse-alternative competitors (largest eigenvalue after whitening) that the paper already contrasts and that could be hybridized with the angular projection.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes high-dimensional tests of covariance separability for matrix-variate data by whitening with the matrix-normal MMLE and testing sphericity of the whitened sample. After whitening, John’s, Nagao’s, and Ledoit–Wolf statistics coincide (Lemma 1); Monte Carlo calibration under a matrix-affine-equivariant estimator yields exact finite-sample level under the null (Theorem 1, Corollaries 1–2, Proposition 1). Consistency against dense Frobenius alternatives is claimed under Assumptions (A1)–(A5) for the elliptical statistic (Theorem 2) and under (B1)–(B5) for an angular (spatial-sign) version that projects whitened observations onto the unit sphere (Theorem 3, Proposition 3). Simulations and an acoustic phonetic application support power and improved robustness of the angular test under heavy-tailed elliptical laws with Gaussian calibration.

Significance. The problem is timely: high-dimensional matrix data make unstructured covariance estimation infeasible, so a computationally light, finite-sample valid test of separability is useful. The finite-sample level theory via matrix affine equivariance is clean and standard. Unifying classical sphericity statistics after MMLE whitening, the efficient inner-product form of the statistic, and the angular refinement for distributional robustness are genuine contributions. An R package with pre-tabulated quantiles is a practical strength. The consistency theorems would be a solid theoretical contribution if the load-bearing pseudo-MLE rates under alternatives were established or clearly demoted; as written they remain conditional.

major comments (2)
  1. [§3.2 Assumptions (A4)–(A5); §5 (B4)–(B5); Theorems 2–3; Discussion §7; Appendix A.1] Theorems 2 and 3 (and the separation arguments that follow Assumptions (A5)/(B5)) treat as given the high-dimensional rates (A4) 1/√(pq)‖Σ̂1⊗̃Σ̂2−Σ1⊗̃Σ2‖_F=o_p(1) and the stronger operator-norm (B4). The manuscript states these rates under non-separable alternatives are open (Discussion §7; comparison to Franks et al. 2026, which covers only the separable case) and supports them only by numerical box-plots in Appendix A.1. Without (A4)/(B4) the inheritance of dense separation after whitening can fail, so the consistency claims are conditional on an unproved estimation result that is load-bearing for the central theoretical contribution. Either prove the rates, replace them by weaker verifiable conditions, or restate Theorems 2–3 as conditional on (A4)/(B4) and move the unconditional claim to finite-sample level plus empirical power.
  2. [Abstract; §1; Theorems 2–3] The abstract and introduction state that the paper proves high-dimensional consistency under dense alternatives without flagging that the proofs rest on open rates for the pseudo-true separable factors. That overstates what is established. Soften the abstract/intro claims to match the conditional status of Theorems 2–3, or complete the rate analysis.
minor comments (5)
  1. [Abstract] Abstract is truncated mid-sentence (“of the prop”). Fix before resubmission.
  2. [§2 Preliminaries] Notation for the tensor product ˜b versus Kronecker product is introduced carefully but still dense; a short display equating vec((Σ1˜bΣ2)X)=(Σ2⊗Σ1)vec(X) early in §2 would help readers who work only with Kronecker notation.
  3. [§6.1.2; Figure 3] Figure 3 panels for df=5 and small m show residual size inflation for the angular test; the text acknowledges this but could quantify how large N must be before Gaussian calibration is reliable under matrix-t.
  4. [§4] Comparison to Sung and Hoff (2026) is useful; make explicit in the main text (not only §4) which of their statistics coincides with the elliptical test after the latest revision, to avoid reader confusion.
  5. [Proposition 1] Proposition 1’s sample-size condition N≥⌈p/q+q/p⌉+2 is cited from Drton et al. (2021); a one-sentence reminder that this is almost necessary would help practitioners.

Circularity Check

0 steps flagged

No circularity: level and consistency claims rest on equivariance, classical sphericity functionals after whitening, and external dense-alternative assumptions rather than self-referential definitions or load-bearing self-citations.

full rationale

The derivation chain is self-contained and non-circular. Under the null, finite-sample level (Corollary 2, Proposition 2) follows from matrix-affine equivariance of the MMLE (Proposition 1, citing the external uniqueness result of Drton et al. 2021) plus orthogonal invariance of the eigenvalue-based statistics (Theorem 1 and Corollary 1); Monte Carlo calibration is performed under a stated F0 and does not fit parameters to the observed data. Consistency (Theorems 2 and 3) is proved by showing that dense Frobenius departure (A5/B5) transfers to a positive separation of the whitened (or angular) statistic after the estimation error controlled by the open rates (A4/B4); those rates are explicitly flagged as unproved under alternatives and are only supported numerically (Appendix A.1, Discussion §7), so the theorems are conditional rather than circular. Self-citations (Masak et al. 2023 on separable components, Mayrhofer et al. 2024 on MMLE equivariance) supply background estimators and expansions used in simulations or auxiliary lemmas; they do not underwrite the target level or consistency statements. The concurrent Sung–Hoff overlap is disclosed as independent. No step reduces a claimed prediction or first-principles result to its own inputs by construction, fitted parameter, or self-citation chain.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 2 invented entities

The method is classical hypothesis testing under an elliptical matrix model plus equivariant separable estimation. Load-bearing unproved pieces are the high-d rates of the separable (pseudo-)MLE under alternatives and the high-d dense-alternative regime; free choices are MC size and the Gaussian default calibration law. No new physical entities; the 'angular statistic' is a constructed test functional, not an ontological invention.

free parameters (3)
  • M (Monte Carlo replicates) = 999
    Recommended M=999 for p-value calibration; finite-M level is floor(α(M+1))/(M+1), not continuous α.
  • Gaussian default F0 for MC under unknown law = matrix-normal reference
    When the sampling distribution is unknown, calibration uses matrix-normal draws; angular projection is intended to reduce but not eliminate dependence on this choice.
  • Sample-size multiplier m defining N=max(4,⌈m pq⌉) = grid in {0.1,0.25,0.5,1,2,4}
    Simulation design parameter controlling high-d aspect ratio; not part of the test definition but drives reported power curves.
axioms (7)
  • domain assumption Model 1: X_n = Σ^{1/2} Z_n with Z_n i.i.d. absolutely continuous, orthogonally invariant, mean 0, Cov = I (elliptical/spherical generator).
    Foundation for whitening–sphericity equivalence and for angular radial cancellation (§3, §5).
  • domain assumption MMLE exists uniquely a.s. for N ≥ ⌈p/q + q/p⌉ + 2 and is matrix affine equivariant.
    Cited from Drton et al. (2021); required for Theorem 1 / Corollary 2 exact level (Prop. 1).
  • domain assumption (A1)/(B1) high-d regime p/√N → γ1, q/√N → γ2 ∈ (0,∞) so pq ≍ N.
    Sets the asymptotic regime for consistency Theorems 2–3.
  • ad hoc to paper (A4)/(B4) separable (pseudo-)MLE converges to the KL-projection factors at o_p(1) Frobenius or operator rate under alternatives.
    Assumed without proof; authors call theoretical guarantees open (§3.2, §7) and only check rates numerically (App. A.1).
  • domain assumption (A5)/(B5) dense alternative: 1/√(pq) ‖Σ − Σ1⊗̃Σ2‖_F ≥ Δ0 > 0 for large N.
    Defines the alternative class the Frobenius/sphericity statistics can detect.
  • domain assumption Bounded operator norms and eigenvalues of Σ, Σ1, Σ2 away from 0 and ∞ uniformly in N ((A2)/(B2)).
    Used to transfer dense non-separability to dense non-sphericity after whitening.
  • standard math Finite eighth moments of entries of Z_n for elliptical consistency ((A3)); milder absolute continuity/orthogonal invariance for angular ((B3)).
    Moment conditions for mean-variance arguments in proofs of Theorems 2–3.
invented entities (2)
  • Matrix-whitened unified John/Nagao/Ledoit–Wolf separability statistic T_N independent evidence
    purpose: Single test statistic for H0: Σ = Σ1⊗̃Σ2 obtained by MMLE whitening then Frobenius sphericity.
    Construction from classical sphericity tests; not a new physical object. Independent evidence via reduction to known statistics after whitening (Lemma 1).
  • Angular separability statistic T_N^{(A)} (spatial-sign covariance after whitening) independent evidence
    purpose: Reduce sensitivity of MC level to elliptical radial/tail misspecification while preserving shape information for separability.
    New recommended procedure in this paper; falsifiable via size/power under non-Gaussian elliptical laws (shown in simulations).

pith-pipeline@v1.1.0-grok45 · 40854 in / 4154 out tokens · 49784 ms · 2026-07-10T08:15:58.829109+00:00 · methodology

0 comments
read the original abstract

Separability is an important structural assumption often placed on the covariance when working with matrix-variate data, because it greatly simplifies both interpretation and computation of subsequent covariance-based statistical tasks. Yet testing the separability assumption is difficult in the high-dimensional regime. We propose to test separability by recasting the problem as a sphericity test after whitening the data using the separable maximum likelihood estimate of the covariance. The test is calibrated by Monte Carlo simulation, yielding finite-sample level control. Furthermore, we prove the test's high-dimensional consistency under dense alternatives. To reduce its reliance on distributional assumptions, we introduce an angular version of the test based on radial normalization after whitening. We demonstrate the practical utility, empirical power, and computational efficiency of the prop

Figures

Figures reproduced from arXiv: 2607.08388 by Marcus Mayrhofer, Tomas Masak, Una Radoji\v{c}i\'c.

Figure 1
Figure 1. Figure 1: Replication of the results of Sung and Hoff (2026) where m “ γ ´1 , introducing the proposed test as well. Lambda_max refers to the test based on the largest eigenvalue and corresponds to ϕ1 of Sung and Hoff (2026) while Elliptical is equivalent to their ϕ3. 6.1.2 Setup 2: Rank-2 Separable Component Alternative Our main synthetic example studies the empirical performance of the proposed tests within a fami… view at source ↗
Figure 2
Figure 2. Figure 2: Empirical power vs. the non-separability index for the Gaussian distribution. [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical power vs. the non-separability index for the [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Convergence rates of MMLE to the pseudo-parameters. The box plots do not [PITH_FULL_IMAGE:figures/full_fig_p029_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Empirical power vs. the non-separability index for the Gaussian distribution. [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

71 extracted references · 71 canonical work pages · 3 internal anchors

  1. [1]

    , title =

    John, S. , title =. Biometrika , year =

    work page

  2. [2]

    Tyler and Daniel Vogel , keywords =

    Alexander Dürre and David E. Tyler and Daniel Vogel , keywords =. On the eigenvalues of the spatial sign covariance matrix in more than two dimensions , journal =. 2016 , issn =. doi:https://doi.org/10.1016/j.spl.2016.01.009 , url =

    work page doi:10.1016/j.spl.2016.01.009 2016

  3. [3]

    2001 , publisher =

    Analysis of Time Series Structure: SSA and Related Techniques , author =. 2001 , publisher =

    work page 2001

  4. [4]

    , title =

    Nagao, H. , title =. The Annals of Statistics , year =

    work page

  5. [5]

    and Wolf, M

    Ledoit, O. and Wolf, M. , title =. The Annals of Statistics , year =

    work page

  6. [6]

    Gupta, A. K. and Nagar, D. K. , title =. 2018 , publisher =

    work page 2018

  7. [7]

    , title =

    Weidmann, J. , title =. 2012 , publisher =

    work page 2012

  8. [8]

    Van Loan, C. F. , title =. Journal of Computational and Applied Mathematics , year =

    work page

  9. [9]

    Van Loan, C. F. and Pitsianis, N. , title =. Linear algebra for large scale and real-time applications , year =

    work page

  10. [10]

    Golub, G. H. and Van Loan, C. F. , title =. 2013 , publisher =

    work page 2013

  11. [11]

    Gupta, A. K. and Varga, T. , title =. Journal of the Italian Statistical Society , year =

    work page

  12. [12]

    Davison, A. C. and Hinkley, D. V. , title =. 1997 , series =

    work page 1997

  13. [13]

    and Yao, J

    Li, W. and Yao, J. , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

    work page

  14. [14]

    and Yao, J

    Wang, Q. and Yao, J. , title =. Electronic Journal of Statistics , year =

    work page

  15. [15]

    and Radojičić, U

    Mayrhofer, M. and Radojičić, U. and Filzmoser, P. , title =. 2024 , eprint =

    work page 2024

  16. [16]

    Mitchell, M. W. and Genton, M. G. and Gumpertz, M. L. , title =. Journal of Multivariate Analysis , year =

    work page

  17. [17]

    Aston, J. A. D. and Pigoli, D. and Tavakoli, S. , title =. The Annals of Statistics , year =

    work page

  18. [18]

    Journal of the American Statistical Association , volume=

    Random surface covariance estimation by shifted partial tracing , author=. Journal of the American Statistical Association , volume=. 2023 , publisher=

    work page 2023

  19. [19]

    , title =

    Fuentes, M. , title =. Journal of Statistical Planning and Inference , year =

    work page

  20. [20]

    and Zimmerman, D

    Lu, N. and Zimmerman, D. L. , title =. Statistics & Probability Letters , year =

    work page

  21. [21]

    and Klein, D

    Filipiak, K. and Klein, D. and Vojtkov. The properties of partial trace and block trace operators of partitioned matrices , journal =. 2018 , volume =

    work page 2018

  22. [22]

    , title =

    Bingham, C. , title =. The Annals of Statistics , year =

    work page

  23. [23]

    Rosenthal, H. P. , title =. Israel Journal of Mathematics , year =

    work page

  24. [24]

    , title =

    White, H. , title =. Econometrica , year =

    work page

  25. [25]

    Rosenthal-type inequalities for the maximum of partial sums of stationary processes and examples , journal =

    Merlev. Rosenthal-type inequalities for the maximum of partial sums of stationary processes and examples , journal =. 2013 , volume =

    work page 2013

  26. [26]

    and Kuriki, S

    Drton, M. and Kuriki, S. and Hoff, P. D. , title =. The Annals of Statistics , year =

    work page

  27. [27]

    Srivastava, M. S. and von Rosen, T. and von Rosen, D. , title =. Mathematical Methods of Statistics , year =

    work page

  28. [28]

    and Oliveira, R

    Franks, C. and Oliveira, R. and Ramachandran, A. and Walter, M. , title =. The Annals of Statistics , year =

    work page

  29. [29]

    and Sarkar, S

    Masak, T. and Sarkar, S. and Panaretos, V. M. , title =. Biometrika , year =

    work page

  30. [30]

    Testing Separability of High-Dimensional Covariance Matrices

    Sung, B. and Hoff, P. D. , title =. arXiv preprint arXiv:2506.17463 , year =. 2506.17463 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  31. [31]

    Separable Geodesic Lagrangian Monte Carlo for Inference in 2-Way Covariance Models

    Simonis, Q. and Wells, M. T. , title =. arXiv preprint arXiv:2501.04913 , year =. 2501.04913 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    and Dierickx, G

    Dette, H. and Dierickx, G. and Kutta, T. , title =. Journal of Time Series Analysis , year =

    work page

  33. [33]

    and Dette, H

    van Delft, A. and Dette, H. , title =. The Annals of Statistics , year =

    work page

  34. [34]

    and Trushin, D

    Soloveychik, I. and Trushin, D. , title =. Journal of Multivariate Analysis , year =

    work page

  35. [35]

    and Kotz, S

    Fang, K.-T. and Kotz, S. and Ng, K. W. , title =. 1990 , publisher =

    work page 1990

  36. [36]

    Thompson, G. Z. and Maitra, R. and Meeker, W. Q. and Bastawros, A. F. , title =. Journal of Computational and Graphical Statistics , year =

    work page

  37. [37]

    and Hadjipantelis, P

    Pigoli, D. and Hadjipantelis, P. Z. and Coleman, J. S. and Aston, J. A. D. , title =. Journal of the Royal Statistical Society Series C: Applied Statistics , year =

    work page

  38. [38]

    and Aston, J

    Pigoli, D. and Aston, J. A. D. and Dryden, I. L. and Secchi, P. , title =. Biometrika , year =

    work page

  39. [39]

    , title =

    Garcia, D. , title =. Computational Statistics & Data Analysis , year =

    work page

  40. [40]

    Tang, R. and M. Pairwise curve synchronization for functional data , journal =. 2008 , volume =

    work page 2008

  41. [41]

    and Krey, S

    Ligges, U. and Krey, S. and Mersmann, O. and Schnackenberg, S. , year =

    work page

  42. [42]

    , title =

    Sueur, J. , title =. 2018 , publisher =

    work page 2018

  43. [43]

    , title =

    Gneiting, T. , title =. Journal of the American Statistical Association , year =

    work page

  44. [44]

    and Genton, M

    Gneiting, T. and Genton, M. G. and Guttorp, P. , title =. Statistical Methods for Spatio-Temporal Systems , editor =. 2007 , pages =

    work page 2007

  45. [45]

    A representation theorem for stochastic processes with separable covariance functions, and its implications for emulation

    Rougier, J. , title =. arXiv preprint arXiv:1702.05599 , year =. 1702.05599 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    and Dette, H

    Bagchi, P. and Dette, H. , title =. The Annals of Statistics , year =

    work page

  47. [47]

    and Kokoszka, P

    Constantinou, P. and Kokoszka, P. and Reimherr, M. , title =. Biometrika , year =

    work page

  48. [48]

    and Chiou, J.-M

    Wang, J.-L. and Chiou, J.-M. and M. Functional data analysis , journal =. 2016 , volume =

    work page 2016

  49. [49]

    Dawid, A. P. , title =. Biometrika , year =

    work page

  50. [50]

    Mardia, K. V. and Goodall, C. R. , title =. Multivariate Environmental Statistics , editor =. 1993 , pages =

    work page 1993

  51. [51]

    , title =

    Dutilleul, P. , title =. Journal of Statistical Computation and Simulation , year =

    work page

  52. [52]

    and Shen, W

    Zhang, Y. and Shen, W. and Kong, D. , title =. Journal of the American Statistical Association , year =

    work page

  53. [53]

    Hoff, P. D. and McCormack, A. and Zhang, A. R. , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

    work page

  54. [54]

    and Xie, J

    Yu, L. and Xie, J. and Zhou, W. , title =. Biometrika , year =

    work page

  55. [55]

    and Hoff, P

    McCormack, A. and Hoff, P. , title =. Bernoulli , year =

    work page

  56. [56]

    and Klein, D

    Filipiak, K. and Klein, D. and Roy, A. , title =. Journal of Multivariate Analysis , year =

    work page

  57. [57]

    and Zhou, S

    Greenewald, K. and Zhou, S. and Hero, III, A. , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

    work page

  58. [58]

    and Chen, K

    Lynch, B. and Chen, K. , title =. Biometrika , year =

    work page

  59. [59]

    Chen, S. X. and Zhang, L.-X. and Zhong, P.-S. , title =. Journal of the American Statistical Association , year =

    work page

  60. [60]

    and Jiang, D

    Bai, Z. and Jiang, D. and Yao, J.-F. and Zheng, S. , title =. The Annals of Statistics , year =

    work page

  61. [61]

    Genton, M. G. , title =. Environmetrics , year =

    work page

  62. [62]

    Yin, Y. Q. and Bai, Z. D. and Krishnaiah, P. R. , title =. Probability Theory and Related Fields , year =

    work page

  63. [63]

    Eaton, M. L. , title =

    work page

  64. [64]

    1999 , publisher=

    Theory of multivariate statistics , author=. 1999 , publisher=

    work page 1999

  65. [65]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Matrix variate regressions and envelope models , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2018 , publisher=

    work page 2018

  66. [66]

    The Annals of Statistics , number =

    Shuheng Zhou , title =. The Annals of Statistics , number =. 2014 , doi =

    work page 2014

  67. [67]

    Journal of the American Statistical Association , volume=

    Testing for nodal dependence in relational data matrices , author=. Journal of the American Statistical Association , volume=. 2015 , publisher=

    work page 2015

  68. [68]

    The annals of applied statistics , volume=

    Latent class modeling using matrix covariates with application to identifying early placebo responders based on EEG signals , author=. The annals of applied statistics , volume=

    work page

  69. [69]

    Statistica Sinica , pages=

    Multivariate functional principal component analysis: A normalization approach , author=. Statistica Sinica , pages=. 2014 , publisher=

    work page 2014

  70. [70]

    Journal of the Royal Statistical Society Series C: Applied Statistics , volume=

    The functional latent block model for the co-clustering of electricity consumption curves , author=. Journal of the Royal Statistical Society Series C: Applied Statistics , volume=. 2018 , publisher=

    work page 2018

  71. [71]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Sparse Kronecker product decomposition: a general framework of signal region detection in image regression , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2023 , publisher=

    work page 2023