pith. sign in

arxiv: 2604.11300 · v1 · submitted 2026-04-13 · 🧮 math.ST · stat.ME· stat.TH

Detection and Mode-Identification of Multiple Change Points in Tensor Factor Models

Yuqi Zhang , Zetai Cen , Haeran Cho This is my paper

Pith reviewed 2026-05-10 14:52 UTC · model grok-4.3

classification 🧮 math.ST stat.MEstat.TH
keywords tensor time serieschange point detectionfactor modelsTucker decompositionmode identificationstructural breakshigh-dimensional statisticsmultiple change points
0
0 comments X

The pith

Tensor factor models allow consistent detection of multiple change points and identification of affected modes.

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

The paper develops algorithms to locate multiple structural breaks in high-dimensional tensor time series that obey a Tucker decomposition factor model, exploiting the low-rank structure for both statistical and computational gains. It further formalizes when each break affects only certain modes of the tensor and supplies a method to recover those modes after the breaks are found. This matters for multi-way data applications because correctly attributing changes to specific modes refines the estimated factor loadings once the series is segmented. The consistency of detection and mode recovery holds under a weak moment condition rather than stronger distributional assumptions.

Core claim

In a Tucker decomposition-based factor model for tensor time series, multiple change points can be detected by leveraging the low-rank structure, and each change can be associated with specific modes through a mode-identification algorithm, with both procedures achieving consistency when the changes are mode-identifiable and moments are weak.

What carries the argument

Mode-identifiability of structural changes within the Tucker factor model, which permits post-detection attribution of each break to the tensor modes that actually shift.

If this is right

  • Change point locations are estimated consistently even when several breaks are present.
  • Mode identification improves the accuracy of post-segmentation estimates of the mode-wise loading spaces.
  • The procedures apply directly to empirical tensor datasets such as daily taxi ridership across locations and time or multi-factor portfolio returns.
  • The weak moment condition extends the range of data for which the guarantees hold compared with stronger assumptions.

Where Pith is reading between the lines

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

  • Similar mode-identification logic could be adapted to other tensor factorizations if an analogous identifiability notion is defined.
  • The separation of detection from mode attribution may reduce error propagation in downstream forecasting after segmentation.
  • Applications to streaming tensor data would require checking whether the batch algorithms can be made online while preserving the consistency properties.

Load-bearing premise

The tensor data must follow a low-rank Tucker factor structure and each change must be mode-identifiable so that the algorithms can locate the breaks and correctly assign them to modes.

What would settle it

A simulated tensor series obeying the low-rank Tucker model and weak moment condition in which the detection step misses one or more true change points or the mode-identification step assigns a change to the wrong modes.

Figures

Figures reproduced from arXiv: 2604.11300 by Haeran Cho, Yuqi Zhang, Zetai Cen.

Figure 1
Figure 1. Figure 1: Barplots of the scaled change point estimators [PITH_FULL_IMAGE:figures/full_fig_p023_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Boxplots of mode-wise loading estimation errors with [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Daily total NYC Yellow Taxi trip volume with the change point estimators returned by [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Heatmap of the scaled mode-identification statistic [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
read the original abstract

We study the problems arising from modeling high-dimensional tensor-valued time series under a Tucker decomposition-based factor model with multiple structural change points. First, we propose an algorithm for detecting the multiple change points, which utilizes the low-rank structure of the data for statistical and computational efficiency. Also, the multi-dimensional array setting poses unique challenges, as some changes are associated with a subset of the modes, and the changes in different modes may interact with one another. Recognizing these, we investigate the problem of identifying each change with the tensor modes post-segmentation. To this end, we formalize the mode-identifiability of each change and propose an algorithm for detecting the modes at which the data are undergoing a mode-identifiable shift. We establish the consistency of both change point detection and mode-identification methods under a weak moment condition, and demonstrate their good performance on simulated datasets where, in particular, it is shown that the mode-identification step can improve the post-segmentation estimation of the mode-wise loading space. Additionally we analyze the datasets on New York City taxi usage and Fama--French portfolio returns using the proposed suite of methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript develops algorithms for detecting multiple change points in high-dimensional tensor time series under a Tucker-decomposition factor model, exploiting low-rank structure for efficiency, and for identifying which tensor modes are affected by each change after segmentation. It formalizes mode-identifiability of changes, claims consistency of both the detection and mode-identification procedures under a weak moment condition, reports favorable simulation results (including improved post-segmentation loading-space estimation), and applies the methods to NYC taxi usage and Fama-French portfolio returns data.

Significance. If the consistency results hold with the stated weak moment condition, the work would provide a useful extension of change-point methods to tensor-valued series, particularly by addressing mode-specific and interacting changes that are distinctive to the multi-way setting. The mode-identification step offers a practical benefit for downstream factor estimation, and the real-data applications illustrate relevance to finance and transportation.

major comments (2)
  1. The central consistency claim for change-point detection and mode identification is stated in the abstract but the manuscript provides no derivation details, explicit minimal-jump-size conditions, or handling of post-hoc mode choices; without these the claim cannot be verified from the available text.
  2. The abstract asserts that the mode-identification step improves post-segmentation estimation of the mode-wise loading space, yet no quantitative comparison (e.g., estimation error with vs. without mode identification) or discussion of error bars is supplied to support this improvement.
minor comments (2)
  1. Notation for the Tucker decomposition and the precise definition of mode-identifiable shifts should be introduced earlier and used consistently throughout.
  2. The simulation section would benefit from a table summarizing detection accuracy, mode-identification accuracy, and computational time across different tensor dimensions and signal strengths.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive evaluation of our work and for the constructive comments. We address each major comment below and describe the revisions we will incorporate to strengthen the manuscript.

read point-by-point responses

  1. Referee: The central consistency claim for change-point detection and mode identification is stated in the abstract but the manuscript provides no derivation details, explicit minimal-jump-size conditions, or handling of post-hoc mode choices; without these the claim cannot be verified from the available text.

    Authors: We agree that the main text presents the consistency results at a high level without full derivation details. The complete proofs appear in the supplementary material, but to improve verifiability we will add a new subsection in the main paper that outlines the key steps of the consistency arguments for both the change-point detection and mode-identification procedures. This subsection will explicitly state the minimal jump-size conditions, the weak moment assumptions, and the manner in which post-segmentation mode choices are accounted for in the theoretical analysis, including the relevant rate requirements. revision: yes

  2. Referee: The abstract asserts that the mode-identification step improves post-segmentation estimation of the mode-wise loading space, yet no quantitative comparison (e.g., estimation error with vs. without mode identification) or discussion of error bars is supplied to support this improvement.

    Authors: The simulation section reports that mode identification improves loading-space estimation, but we acknowledge that a direct quantitative comparison with variability measures is not currently presented. We will revise the simulation studies to include a table that reports the average estimation errors (Frobenius or subspace distance) for the mode-wise loading matrices both with and without the mode-identification step, across multiple simulation settings. Each entry will be accompanied by the standard deviation computed over the Monte Carlo replications to provide a clear measure of the improvement and its variability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper proposes new algorithms for multiple change-point detection and subsequent mode-identification in tensor-valued time series under a Tucker factor model, then establishes consistency of both procedures under a weak moment condition. No equations, assumptions, or cited results in the provided description reduce the central consistency claims or algorithmic outputs to quantities defined by the same fitted parameters or self-referential inputs. The low-rank structure is exploited as an external modeling assumption rather than derived from the detection procedure itself, leaving the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the Tucker decomposition assumption for the factor model and the existence of mode-identifiable shifts; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)
  • domain assumption Data follow a Tucker decomposition-based factor model with low-rank structure
    Invoked to enable statistical and computational efficiency in the detection algorithm
  • domain assumption Changes are mode-identifiable
    Required for the post-segmentation mode-identification step to be well-defined

pith-pipeline@v0.9.0 · 5503 in / 1290 out tokens · 39416 ms · 2026-05-10T14:52:28.964592+00:00 · methodology

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

15 extracted references · 15 canonical work pages · 1 internal anchor

  1. [1]

    and Cribben, I

    Anastasiou, A. and Cribben, I. (2025). Tensor time series change-point detection in cryptocurrency network data.Journal of the Royal Statistical Society Series A: Statistics in Society, page qnaf180. Babii, A., Ghysels, E., and Pan, J. (2025). Tensor PCA for factor models.Journal of Econometrics, page 106077. Bai, J. (2003). Inferential theory for factor ...

    work page 2025

  2. [2]

    Bai, J., Duan, J., and Han, X. (2024). The likelihood ratio test for structural changes in factor models.Journal of Econometrics, 238(2):105631. Bai, J. and Ng, S. (2013). Principal components estimation and identification of static factors. Journal of Econometrics, 176(1):18–29. Baltagi, B. H., Kao, C., and Wang, F. (2017). Identification and estimation ...

    work page 2024

  3. [3]

    Cho, H., Kirch, C., and Stoffregen, N. (2026). Multivariate data segmentation via multiscale moving sum procedure with localized pruning.Preprint. Cho, H., Kley, T., and Li, H. (2025). Detection and inference of changes in high-dimensional linear regression with nonsparse structures.Journal of the Royal Statistical Society Series B: Statistical Methodolog...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  4. [4]

    Peng, L., Zou, G., and Wu, J. (2025a). Detection of multiple structural changes in matrix factor 32 models.Journal of the Korean Statistical Society, 54(4):1296–1322. Peng, L., Zou, G., and Wu, J. (2025b). Estimation of high-dimensional matrix factor models with change points.Studies in Nonlinear Dynamics & Econometrics. Online first. Peˇ sta, M., Peˇ sto...

    work page 2025

  5. [5]

    Xie, Y., Huang, J., and Willett, R. (2012). Change-point detection for high-dimensional time series with missing data.IEEE Journal of Selected Topics in Signal Processing, 7(1):12–27. Yamamoto, Y. and Tanaka, S. (2015). Testing for factor loading structural change under common breaks.Journal of Econometrics, 189(1):187–206. Yu, L., He, Y., Kong, X., and Z...

    work page 2012

  6. [6]

    Iu 0u×u 0u×u 0u×u # , A 2,1 =

    =∅. This leads to an interesting re-formulation as Xt = ( (F ⊺ t ,G ⊺ t )⊺ ×1 (Z1,0 p1×u) +E t for 1≤t≤θ 1, (F ⊺ t ,G ⊺ t )⊺ ×1 (Z2,0 p1×u) +E t forθ 1 + 1≤t≤T, for some hypotheticalG t as an independent copy ofF t so that the core factor requirement As- sumption 1 can be satisfied. The core factor rank is thereforer 1 = 2uaccording to our formulation in ...

    work page 2024

  7. [7]

    Immediately, we also have eΛ-keΛ⊺ -k/p-k = Λ-kΛ⊺ -k/p-k +o P (1)

    That is, for any ℓ∈[K], we have 1 pℓ eΛℓeΛ⊺ ℓ = 1 pℓ (eΛℓ −Λ ℓeHℓ)eΛ⊺ ℓ + 1 pℓ ΛℓeHℓeΛ⊺ ℓ = 1 pℓ (eΛℓ −Λ ℓeHℓ)eΛ⊺ ℓ + 1 pℓ ΛℓeHℓ(eΛℓ −Λ ℓeHℓ)⊺ + 1 pℓ ΛℓeHℓeH ⊺ ℓ Λ⊺ ℓ = 1 pℓ ΛℓΛ⊺ ℓ +o P (1), (B.19) noting that Λ ℓΛ⊺ ℓ /pℓ has the firstr k eigenvalues being 1’s and the rest being zero. Immediately, we also have eΛ-keΛ⊺ -k/p-k = Λ-kΛ⊺ -k/p-k +o P (1). Hence...

    work page 2021

  8. [8]

    ×[r K], we have Φν,β(G)≤2C(A)(C ν(F))1/νcβ <∞

    β ∞X s=m δs,ν,u(G)≤2C(A)(C ν(F))1/νcβ <∞, 67 taking the maximum overu∈[r 1]×. . .×[r K], we have Φν,β(G)≤2C(A)(C ν(F))1/νcβ <∞. Finally we conclude Φν,β(C)≤2c K λ rKC(A)(C ν(F))1/νcβ <∞. Lemma B.14.Under the assumptions of Lemmas B.12 and B.13, the uniform d.a.n. of{X t} satisfies Φν,β(X)≤2c K λ rK C(A)(C ν(F))1/ν cβ +C(E) 2(Cν(Fe))1/νcβ + 2(Cν(ϵ))1/νcβ <...

    work page 2011

  9. [9]

    Assume analogous conditions for{X t}(with the sameνandβwithout loss of generality)

    β ∞X s=m |as| ≤c β. Assume analogous conditions for{X t}(with the sameνandβwithout loss of generality). Then{Y t}and{X t}are weaklyM-dependent inL 2ν, and{Y tXt}is weaklyM-dependent inL ν, all with the rate functionδ(m) =m −β. (iii) (Squared linear process). Consider{Y t}defined in (ii) above. Letβ >0. Then the process {Y 2 t }is also weaklyM-dependent in...

    work page 2011

  10. [10]

    By Assumption 2, J (4) k,τ,e − J (4) k,s,τ F ≤ 1 p2 k 1 e−τ eX t=τ+1 bΛk −Λ kbHk ⊺ΛkE Gk,tG⊺ k,t Λ⊺ kbΛk − 1 p2 k 1 τ−s τX t=s+1 bΛk −Λ kbHk ⊺ΛkE Gk,tG⊺ k,t Λ⊺ kbΛk F + 1 p2 k 1 e−τ eX t=τ+1 ΛkbHk ⊺ΛkE Gk,tG⊺ k,t Λ⊺ k bΛk −Λ kbHk − 1 p2 k 1 τ−s τX t=s+1 ΛkbHk ⊺ΛkE Gk,tG⊺ k,t Λ⊺ k bΛk −Λ kbHk F ≤2∥ bHk∥ 1√pk bΛk −Λ kbHk F · 1 e−τ eX t=τ+1 E Gk,tG⊺ k,t − 1 ...

    work page 2021

  11. [11]

    This shows that E     max ω−2 w κ2 T ≤ℓ≤θw−θw−1 q ω−2w κ2 T ℓ θwX t=θw−ℓ+1 Z⊺ k,t,i·Zk,t,j· −E(Z ⊺ k,t,i·Zk,t,j·)   2+ϵ  =O(1)

    Then for the same constantϵin Lemma B.22, we have for anyk∈[K]and 87 κT → ∞arbitrarily slowly, P    max j∈[q+1] max ω−2 j κ2 T ≤ℓ≤θj −θj−1 q ω−2 j κ2 T pkℓ θjX t=θj −ℓ+1 Zk,tZ⊺ k,t −E(Z k,tZ⊺ k,t) F ≥κ T    =o(1).(B.36) Proof of Lemma B.23.By Minkowski’s inequality, we have for anyk∈[K],w∈[q+ 1], and i, j∈[p k], E     max ω−2 w κ2 T ≤ℓ≤θw−θw−1 ...

    work page 2006

  12. [12]

    Next, we show that bθ−θ j ≤ 1 4∆◦.(B.46) If not, by Lemma B.27 and Lemma S3.5 in Cho et al. (2025), Vaℓ◦ ,θj ,bℓ◦ 2 − Vaℓ◦ ,bθ,bℓ◦ 2 ≥ s (θj −a ℓ◦)(bℓ◦ −θ j) bℓ◦ −a ℓ◦  1− vuut1− |bθ−θ j|/(θj −a ℓ◦) 1 +|bθ−θ j|/(bℓ◦ −θ j)   ωj ·(1 +O P (αT,p)) (Lemma B.27) ≥ 1− r 3 4 !s (θj −a ℓ◦)(bℓ◦ −θ j) bℓ◦ −a ℓ◦ ωj ·(1 +O P (αT,p))≳ √ ∆◦ωj(1 +o P (1)),(B.47) On t...

    work page 2025

  13. [13]

    From Lemma B.25, we know that |Vaℓ◦ ,τ,bℓ◦ |2 attains its maximum atτ=θ j within (aℓ◦, bℓ◦)

    We are now ready to derive the refined rate of estimation. From Lemma B.25, we know that |Vaℓ◦ ,τ,bℓ◦ |2 attains its maximum atτ=θ j within (aℓ◦, bℓ◦). By similar arguments,|(g ◦)⊺Vaℓ◦ ,τ,bℓ◦ | attains its maximum atτ=θ j. Therefore, we have (g ◦)⊺Vaℓ◦ ,θj ,bℓ◦ ≥(g ◦)⊺Vaℓ◦ ,bθ,bℓ◦ ≥0. Then together with Lemma B.26, we have bVaℓ◦ ,bθ,bℓ◦ 2 = (g◦)⊺bVaℓ◦ ,bθ...

    work page 2018

  14. [14]

    Firstly, forW 1, by Lemma S3.5 of Cho et al. (2025), and by (B.46), U11 = bℓ◦ −θ j√bℓ◦ −a ℓ◦ s bθ−a ℓ◦ bℓ◦ −bθ − s θj −a ℓ◦ bℓ◦ −θ j J (1) θj ,bℓ◦ 2 + J (2) θj ,bℓ◦ 2 ≤ bℓ◦ −θ j√bℓ◦ −a ℓ◦ s θj −a ℓ◦ bℓ◦ −θ j (θj −bθ)(bℓ◦ −a ℓ◦) (θj −a ℓ◦)(bℓ◦ −θ j) J (1) θj ,bℓ◦ 2 + J (2) θj ,bℓ◦ 2 ≤ s bℓ◦ −a ℓ◦ (θj −a ℓ◦)(bℓ◦ −θ j) · |θj −bθ| J (1) θj ,bℓ◦ 2 + J (2) θj ,...

    work page 2025

  15. [15]

    average±standard deviation

    are both positive with probability approaching one: B1 tr(B1) − C1 tr(C1) + B2 tr(B2) − C2 tr(C2) = B1 −C 1 tr(B1) +C 1 1 tr(B1) − 1 tr(C1) + B2 −C 2 tr(B2) +C 2 1 tr(B2) − 1 tr(C2) ≲P ∥B1 −C 1 +B 2 −C 2∥.(B.58) Next, fix anyk∈[K]. Notice that by (b)–(c), Definition 1, (2.3) and (3.2), we have, for all j∈[q+ 1],a j ∈Θ − j andb j ∈Θ + j , Γ(k) G,aj−1 ,bj =...

    work page 2000