On Testing Kronecker Product Structure in Tensor Factor Models

Clifford Lam; Zetai Cen

REVIEW 2 major objections 5 minor 22 references

On Testing Kronecker Product Structure in Tensor Factor Models

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a residual-comparison test for whether the factor loading matrix in a Tucker tensor factor model is a Kronecker product, and proves that the test's rejection probability is asymptotically controlled under the null.

desk verdict A useful new test with solid simulations, but the size-control theorem is unproven as stated. read the letter →

arxiv 2501.11208 v1 pith:6RYETRWC submitted 2025-01-20 math.ST stat.MEstat.TH

Zetai Cen , Clifford Lam This is my paper · ORCID

classification math.STstat.MEstat.TH MSC 62H2562M1062F03

keywords tensorfactormodelKroneckerproductstructureTuckerdecompositionreshapehigh-dimensionaltimeseriesweakfactorsfactor-structuredidiosyncraticerrorshypothesistesting

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 is trying to establish a practical, assumption-checking test for tensor factor models. Before applying a Tucker tensor factor model, one implicitly assumes that the vectorised factor loading matrix is a Kronecker product of smaller matrices, and this paper tests exactly that assumption. The test compares squared residual sums from a full tensor factor model with those from a factor model fitted to a reshaped, possibly vectorised, version of the data: under the null these residuals have the same asymptotic distribution, while under the alternative the reshaped residuals are inflated. The paper proves size control for the decision rule, allows weak factors, and provides a practical algorithm that identifies which modes of the tensor break the Kronecker structure.

What carries the argument

The central object is the Kronecker product structure set $\mathcal{K}_{b_1\times\cdots\times b_\kappa}$, the set of full-column-rank matrices that factor as $A_\kappa\otimes\cdots\otimes A_1$ with each small loading matrix of low rank and controlled factor strength; the null hypothesis is that the merged loading matrix $A_V$ lies in this set. The tensor reshape operator $\mathrm{Reshape}(\cdot,\cdot)$ merges selected modes into one, so a factor model can be fitted both on the original tensor and on the reshaped tensor. The test statistic compares empirical cumulative distribution functions of aggregated squared residuals from the two fits, using the set $\mathcal{R}$ of divisor combinations of the merged factor count to hedge against unknown factor numbers on the merged modes.

What would settle it

Simulate a null model in which the noise is generated without the sparse factor structure of Assumption (E1), for instance with a dense loading matrix or with innovations whose fourth moments are not finite, and record the empirical rejection rate of the decision rule (3.12) at $\alpha=0.01$; if the rejection rate stays above the nominal level as $T$ and the mode dimensions grow, the asymptotic size claim is false.

Watch

Extended reading notes

Core claim

The central claim is that Kronecker product structure in a tensor factor model can be tested by residual comparison, with the test's size controlled asymptotically. Theorem 1 links the structure to the Tucker decomposition through tensor reshape: a factor model on a reshaped tensor whose loading matrix is a Kronecker product is equivalent to a Tucker-decomposition tensor factor model on the original tensor. Theorem 2 shows that, under the null, the aggregated squared residual sums from the two fits converge to the same standard normal limits, and Theorem 3 then guarantees that the decision rule in (3.12) rejects with probability at most $\alpha$ as dimensions and sample size grow. A second reshape theorem allows a user to test each mode separately and locate the modes along which the Kronecker structure is lost.

Load-bearing premise

The test's null distribution relies on the noise tensor decomposing as a sparse factor structure plus idiosyncratic noise with bounded fourth moments; if the noise carries strong cross-sectional dependence, dense loadings, or heavy tails, the claimed size control may fail.

Editorial extensions

If this is right

Users of matrix and tensor factor models get a pre-test for the Kronecker loading assumption; rejecting the null indicates that a general vector factor model may be more appropriate than the structured tensor model.
The practical algorithm can identify which modes cause the loss of Kronecker structure, guiding modelling choices rather than only flagging misspecification.
Weak factors are covered by the theory, so the test is not restricted to pervasive-factor settings and its convergence rates are explicit.
Because the test works on reshaped versions of the tensor, unbalanced tensors with small mode dimensions can be tested with better accuracy than vectorising the whole tensor.
Real-data results indicate that Fama-French portfolio return matrices deviate from a matrix factor model, in cases where an earlier one-way versus two-way factor model test does not reject.

Reading between the lines

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

The same residual-comparison logic could be adapted to test other structured decompositions, such as CP or partially shared loading matrices, by replacing the Kronecker product structure set with the appropriate structure class.
Since Theorem 1 links reshaping to the identifiability of the Tucker model, the test doubles as a diagnostic for whether merging modes before estimation is legitimate.
A natural extension is to replace the 5% quantile aggregation in the decision rule with a smoother functional of the empirical distributions, which might improve power in small samples or under heavy-tailed noise.
The test answers a different question from existing covariance Kronecker tests: it targets the loading matrix rather than the covariance matrix, so the two approaches are complementary and could be combined in model diagnostics.

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

Here's my take: the paper is worth engaging with, but the headline theorem is not proven as stated.

What is genuinely new: the tensor reshape theorems (Theorem 1 and Theorem 4) are clean and independently useful, and this is the first direct test of Kronecker product structure in the loading matrix of a Tucker tensor factor model. It is clearly distinct from covariance-based Kronecker tests and from the one-way/two-way matrix tests of He et al. (2023). The simulations are thorough and honest, covering weak factors, heavy-tailed noise, misspecified ranks, and the practical algorithm for unknown mode sets. The real data applications are a nice bonus.

The soft spot is Theorem 3. The proof says that because qhat_y and qhat_x are asymptotically equal, the empirical exceedance probability at qhat_x is also bounded by alpha. That is not valid. For a fixed j, the exceedance probability at qhat_x differs from the one at qhat_y by O_P(T^{-1/2}), and it will exceed alpha with probability about 1/2, not with probability going to 1. The proof also never connects the single-j bound to the 5%-quantile-over-j statistic in (3.12). This is a logical gap, not a missing technical condition. Separately, Assumption (E1) is genuinely strong: the noise itself must follow a tensor factor model with sparse loadings plus a idiosyncratic component. Without empirical validation of that structure, the asymptotic null distribution can fail in exactly the settings where the test is most needed. There is also no theoretical power analysis under H1, only simulations.

That said, the test may still work in practice—the minimum over R and the quantile aggregation likely make it conservative. But as it stands, the theoretical guarantee is not established. I would send this to peer review because the question is important, the reshape theorems are solid, and the simulation evidence is substantial. A good referee should focus on the proof of Theorem 3 and ask for either a correct size bound for the actual decision rule or an explicitly conservative correction. Minor point: the R package link is not actually a URL.

For whom: statisticians and econometricians working on tensor time series factor models. Recommendation: engage, but require a revised Theorem 3 before publication.

Referee Report

2 major / 5 minor

Summary. The paper proposes a test for whether an order-K tensor time series follows a Tucker-decomposition tensor factor model, equivalently whether the factor loading matrix of the vectorized data has Kronecker product structure. The test compares squared residuals from a full tensor factor model with those from a factor model fitted to a reshaped version of the data, using empirical quantiles of aggregated residual squares. Theoretical results include tensor reshape theorems (Theorems 1 and 4), asymptotic normality of residual sums (Theorem 2), and a claimed size-control guarantee for the decision rule (3.12) under the null (Theorem 3). The authors provide simulations for K=2,3,4, two real-data applications (NYC taxi and Fama-French portfolios), and an R package. The main contribution is a practical testing procedure with non-trivial asymptotic analysis allowing weak factors and factor-structured noise.

Significance. If the central theorem were valid, the paper would provide a useful and timely tool for checking the Kronecker product structure assumption underlying Tucker tensor factor models, a question not directly addressed by earlier tests for Kronecker covariance structure. The tensor reshape theorems and the allowance for weak factors are genuine contributions, and the empirical study is reasonably extensive, including robustness checks and real data. The supplement contains detailed proofs and an implementation is provided. However, the main theoretical claim of asymptotic size control is not established: the proof of Theorem 3 contains a quantile-swap that is invalid, and the finite-sample simulations for K=2 actually show over-rejection relative to the nominal level. Because the size-control theorem is load-bearing for the proposed procedure, the paper in its current form does not deliver its central promise.

major comments (2)

[Section 3.3, Eq. (3.12) and Theorem 3] The aggregation step in the decision rule (3.12), which takes the 5% quantile over j of the per-j exceedance probabilities before minimizing over m, is not theoretically analyzed. Even if a corrected per-j bound were available, the paper gives no argument for how the empirical 5% quantile over the d/dk* values behaves under H0, nor how the level alpha relates to the fixed 5% aggregation constant. The statement of Theorem 3 concerns the per-j inequality (4.8), and the proof does not bridge that statement to the finite-j or growing-j quantile aggregation used in practice. This is a load-bearing gap because the size of the test is determined by the joint behavior of all j, not by a single j. A revised theorem must explicitly control the rejection event { 5% quantile over j of min_m p_{m,j} > alpha } under H0.
[Section 4.1, Assumption (E1)] Assumption (E1) imposes a strong and specific structure on the noise: it must decompose as a tensor factor model with approximately sparse loadings (||A_{e,k}||_1 = O(1)) plus an idiosyncratic component with i.i.d. elements and bounded fourth moments. This assumption is load-bearing because the CLT in Theorem 2 relies on the residual sums being dominated by the i.i.d. idiosyncratic terms after the factor-structured noise contributions are shown to be negligible. If real noise has stronger cross-sectional dependence, heavy tails, or non-sparse factor loadings, the asymptotic null distribution of the test statistics may fail. The simulations include heavy-tailed innovations (Setting Id/IId) and weak factors, and the empirical size remains reasonable, but there is no theoretical result covering these cases. The authors should state more precisely the scope of the theoretical results and discuss how Assumption (E1) might be relaxed or tested.

minor comments (5)

[Abstract] The abstract contains a typo: 'we demonstrate out tests' should read 'we demonstrate our tests'.
[Section 3.3 and Theorem 3] The symbol Phi is used for both the empirical cumulative distribution function in (3.11) and the standard normal CDF in Theorem 2; this dual use is confusing and should be disambiguated.
[Section 4.1, (F1) and (4.5)] Assumption (F1) defines X_{reshape,f,t} for the reshaped core factor, while (4.5) introduces X_{f,t} for the original core factor; the relationship between these two processes should be stated explicitly, as it is used in the proof of Lemma 2.
[Section 5.1, Tables 1-5] The tables report averages of hat(alpha) and hat(p) over 500 runs but no standard errors or confidence intervals; given that the K=2 results show non-negligible rejection rates, reporting the variability across runs would help assess whether the deviations from alpha are systematic.
[Section 5.2, Table 6] The real-data analysis interprets hat(q_alpha) values that are only slightly above alpha (e.g., 0.011 at alpha=0.01) as 'mild evidence' of no Kronecker structure, but given the finite-sample over-rejection observed in simulations for K=2, the strength of this evidence is unclear; a calibration study or more cautious wording would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the test statistic and null distribution are derived from first-principles asymptotics, with self-citations used only for auxiliary technical lemmas.

full rationale

The paper's central claim is an asymptotic size control theorem for a residual-comparison test. The test statistic compares squared residual sums from two fitted factor models (full tensor factor model versus reshaped/vector factor model), and no fitted constant or parameter is renamed as a prediction. The null distribution is obtained from a CLT on weighted sums of squared idiosyncratic noises, and the equivalence of the two residual sums under H0 follows from the algebra of tensor reshaping (Theorem 1) plus estimation-error rates (Lemmas 5-7). The Kronecker product structure set in Definition 1 is a formal characterization of the null hypothesis, not an input used to force the outcome. The self-citations to Cen and Lam (2024) supply technical bounds on weak dependence and factor estimation rates; they do not assume the existence of the proposed test or its size control. A proof gap in Theorem 3 (the swap of an estimated quantile for a limit quantile) is a genuine correctness concern, but it is a logical gap rather than circular reasoning: the theorem is not obtained by assuming its conclusion, and the empirical quantile is not fitted to reproduce alpha. Assumption (E1) is a strong model assumption but is a stated condition, not derived from the target result. Therefore no step in the derivation reduces to its own inputs by construction.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The central claim relies on a dense set of domain assumptions about factor strengths (L1, L2, S1), noise structure (E1, E2), and dimension-rate trade-offs (R1, R2). The strongest and most fragile is (E1), which imposes a factor-plus-sparse-idiosyncratic structure on the noise itself. The proofs also depend on propositions from the authors' prior preprint (Cen and Lam, 2024), which adds to the burden of verification.

assumptions (9)

domain assumption (L1): Mode-j loading matrices are full rank with normalized Gram matrices converging to positive definite matrices, with factor strengths 1/2 < delta <= 1.
Used in Section 4.1 to ensure PCA estimators are consistent and to set rates for factor strength.
domain assumption (L2): Under H0, each Ai in (3.5) is full rank with similar normalized Gram convergence and ordered factor strengths.
Specific to the null hypothesis and used in the proofs of Lemma 5 and Theorem 2.
domain assumption (F1): The core factor Freshape,t is a linear process of i.i.d. elements with absolutely summable coefficients.
Provides the sample moment convergence in Lemma 2 and the weak dependence needed for CLT.
domain assumption (E1): The noise Et decomposes as a tensor factor model plus idiosyncratic noise with sparse loadings and independent elements.
This is the load-bearing structure that yields the CLT on squared residuals in Theorem 2; Section 4.1.
domain assumption (E2): The noise components Fe,t and epsilon_t are linear processes with i.i.d. or independent elements and bounded fourth moments.
Used in Lemma 1 to bound noise correlations and in proofs of residual convergence.
domain assumption (R1) and (R2): Rate assumptions linking tensor dimensions, time length, and factor strengths.
Required for the convergence rates in Lemmas 4-7; R2 forces min mode dimension o(T) when testing all modes.
ad hoc to paper (S1): Factor strength is allocated equally across modes in terms of the Frobenius norm to resolve identification indeterminacy.
Introduced in the supplement to fix the non-uniqueness of Kronecker factor decompositions; not empirically verifiable.
domain assumption Propositions 1.1, 1.2, 1.3 from Cen and Lam (2024) on noise decorrelation and factor sample moments.
Cited as the basis for Lemmas 1 and 2; these results are from the authors' own preprint and are not reproduced here.
standard math Multivariate CLT for weighted sums from Ayvazyan and Ulyanov (2023).
Used in the proof of Theorem 2 to establish the asymptotic normality of the residual sums.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On Testing Kronecker Product Structure in Tensor Factor Models." pith.science (2026). https://pith.science/paper/6RYETRWC

@misc{pith2026250111208,
  author       = {Pith},
  title        = {Pith review of: On Testing Kronecker Product Structure in Tensor Factor Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6RYETRWC}},
  note         = {Machine review of arXiv:2501.11208}
}

abstract

We propose a test for testing the Kronecker product structure of a factor loading matrix implied by a tensor factor model with Tucker decomposition in the common component. Through defining a Kronecker product structure set, we define if a tensor time series response $\{\mathcal{Y}_t\}$ has a Kronecker product structure, equivalent to the ability to decompose $\{\mathcal{Y}_t\}$ according to a tensor factor model. Our test is built on analysing and comparing the residuals from fitting a full tensor factor model, and the residuals from fitting a (tensor) factor model on a reshaped version of the data. In the most extreme case, the reshaping is the vectorisation of the tensor data, and the factor loading matrix in such a case can be general if there is no Kronecker product structure present. Theoretical results are developed through asymptotic normality results on estimated residuals. Numerical experiments suggest that the size of the tests gets closer to the pre-set nominal value as the sample size or the order of the tensor gets larger, while the power increases with mode dimensions and the number of combined modes. We demonstrate out tests through a NYC taxi traffic data and a Fama-French matrix portfolio of returns.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

[1]

Ayvazyan, S. A. and Ulyanov, V. V. (2023). A multivariate clt for weighted sums with rate of convergence of order o(1/n). In Belomestny, D., Butucea, C., Mammen, E., Moulines, E., Rei , M., and Ulyanov, V. V., editors, Foundations of Modern Statistics , pages 225--257, Cham. Springer International Publishing

work page 2023
[3]

Tensor Time Series Imputation through Tensor Factor Modelling

Cen, Z. and Lam, C. (2024). Tensor time series imputation through tensor factor modelling. arXiv:2403.13153v2

work page Pith review arXiv 2024
[4]

Chang, J., He, J., Yang, L., and Yao, Q. (2023). Modelling matrix time series via a tensor CP-decomposition . Journal of the Royal Statistical Society Series B: Statistical Methodology , 85(1):127--148

work page 2023
[5]

Chen, E. Y. and Fan, J. (2023). Statistical inference for high-dimensional matrix-variate factor models. Journal of the American Statistical Association , 118(542):1038--1055

work page 2023
[6]

Y., Xia, D., Cai, C., and Fan, J

Chen, E. Y., Xia, D., Cai, C., and Fan, J. (2024). Semi-parametric tensor factor analysis by iteratively projected singular value decomposition. Journal of the Royal Statistical Society Series B: Statistical Methodology , 86(3):793--823

work page 2024
[7]

Chen, R., Yang, D., and Zhang, C.-H. (2022). Factor models for high-dimensional tensor time series. Journal of the American Statistical Association , 117(537):94--116

2022
[8]

Factor Strength Estimation in Vector and Matrix Time Series Factor Models

Chen, W. and Lam, C. (2024a). Factor strength estimation in vector and matrix time series factor models. arXiv:2405.07294v1

work page Pith review arXiv 2024
[9]

and Lam, C

Chen, W. and Lam, C. (2024b). Rank and factor loadings estimation in time series tensor factor model by pre-averaging. Annals of Statistics , 52(1):364--391

work page 2024

Show all 22 references

[10]

Guggenberger, P., Kleibergen, F., and Mavroeidis, S. (2023). A test for kronecker product structure covariance matrix. Journal of Econometrics , 233(1):88--112

2023
[11]

Han, Y., Chen, R., Yang, D., and Zhang, C.-H. (2024a). Tensor factor model estimation by iterative projection . The Annals of Statistics , 52(6):2641 -- 2667

2024
[12]

Han, Y., Yang, D., Zhang, C.-H., and Chen, R. (2024b). Cp factor model for dynamic tensors. Journal of the Royal Statistical Society Series B: Statistical Methodology , 86(5):1383--1413

2024
[13]

H., and Chen, R

Han, Y., Zhang, C. H., and Chen, R. (2022). Rank determination in tensor factor model. Electronic Journal of Statistic , 16:1726--1803

2022
[14]

He, Y., Kong, X., Trapani, L., and Yu, L. (2023). One-way or two-way factor model for matrix sequences? Journal of Econometrics , 235(2):1981--2004

2023
[15]

He, Y., Li, L., and Trapani, L. (2022a). Statistical inference for large-dimensional tensor factor model by weighted/unweighted projection. arXiv:2206.09800

2022 arXiv
[16]

He, Y., Wang, Y., Yu, L., Zhou, W., and Zhou, W.-X. (2022b). Matrix kendall's tau in high-dimensions: A robust statistic for matrix factor model. arXiv:2207.09633

2022
[17]

Kolda, T. G. and Bader, B. W. (2009). Tensor decompositions and applications. SIAM Review , 51(3):455--500

2009
[18]

and Yao, Q

Lam, C. and Yao, Q. (2012). Factor modeling for high-dimensional time series: Inference for the number of factors. The Annals of Statistics , 40(2):694--726

2012
[19]

Onatski, A. (2012). Asymptotics of the principal components estimator of large factor models with weakly influential factors. Journal of Econometrics , 168(2):244--258

2012
[20]

D., and Bajwa, W

Taki, B., Sarwate, A. D., and Bajwa, W. U. (2024). Low separation rank in tensor generalized linear models: An asymptotic analysis. In 2024 58th Annual Conference on Information Sciences and Systems (CISS) , pages 1--6

2024
[21]

Wang, D., Liu, X., and Chen, R. (2019). Factor models for matrix-valued high-dimensional time series. Journal of Econometrics , 208(1):231--248. Special Issue on Financial Engineering and Risk Management

2019
[22]

Yu, L., He, Y., Kong, X., and Zhang, X. (2022a). Projected estimation for large-dimensional matrix factor models. Journal of Econometrics , 229(1):201–217

2022
[23]

Yu, L., Xie, J., and Zhou, W. (2022b). Testing kronecker product covariance matrices for high-dimensional matrix-variate data. Biometrika , 110(3):799--814

2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Ayvazyan, S. A. and Ulyanov, V. V. (2023). A multivariate clt for weighted sums with rate of convergence of order o(1/n). In Belomestny, D., Butucea, C., Mammen, E., Moulines, E., Rei , M., and Ulyanov, V. V., editors, Foundations of Modern Statistics , pages 225--257, Cham. Springer International Publishing

work page 2023

[2] [3]

Tensor Time Series Imputation through Tensor Factor Modelling

Cen, Z. and Lam, C. (2024). Tensor time series imputation through tensor factor modelling. arXiv:2403.13153v2

work page Pith review arXiv 2024

[3] [4]

Chang, J., He, J., Yang, L., and Yao, Q. (2023). Modelling matrix time series via a tensor CP-decomposition . Journal of the Royal Statistical Society Series B: Statistical Methodology , 85(1):127--148

work page 2023

[4] [5]

Chen, E. Y. and Fan, J. (2023). Statistical inference for high-dimensional matrix-variate factor models. Journal of the American Statistical Association , 118(542):1038--1055

work page 2023

[5] [6]

Y., Xia, D., Cai, C., and Fan, J

Chen, E. Y., Xia, D., Cai, C., and Fan, J. (2024). Semi-parametric tensor factor analysis by iteratively projected singular value decomposition. Journal of the Royal Statistical Society Series B: Statistical Methodology , 86(3):793--823

work page 2024

[6] [7]

Chen, R., Yang, D., and Zhang, C.-H. (2022). Factor models for high-dimensional tensor time series. Journal of the American Statistical Association , 117(537):94--116

2022

[7] [8]

Factor Strength Estimation in Vector and Matrix Time Series Factor Models

Chen, W. and Lam, C. (2024a). Factor strength estimation in vector and matrix time series factor models. arXiv:2405.07294v1

work page Pith review arXiv 2024

[8] [9]

and Lam, C

Chen, W. and Lam, C. (2024b). Rank and factor loadings estimation in time series tensor factor model by pre-averaging. Annals of Statistics , 52(1):364--391

work page 2024