acceptodds
Under review as a conference paper at ICLR 2027

Optimal detection of general moment changes: Simultaneous mean and covariance change detection and beyond

Abstract

We study multiple change-point detection in multivariate time series whose distributions change in a piecewise constant manner. Distributional changes can manifest across different moment orders, from shifts in the mean and covariance to changes in higher-order moments. Higher-order moments capture increasingly rich distributional features but become difficult to estimate in high dimensions. Our tensor representation unifies moments of different orders within a common linear algebraic framework, enabling a new method to detect changes in moments of all orders up to a prescribed fixed order . The resulting procedure accommodates temporal dependence and allows the dimension of the time series to grow with the sample size. Under suitable regularity conditions, the proposed procedure achieves a localization error rate that matches a newly developed minimax lower bound. We further derive limiting distributions under both nonvanishing and vanishing moment jumps and construct asymptotically valid confidence intervals in the vanishing-jump regime. Numerical experiments and real-data analyses demonstrate the method's effectiveness in detecting moment changes and a range of distributional shifts.

Primary area:
learning on time series and dynamical systems
Keywords:
Change-point inference, Optimal localization, Moment tensor, Time series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1dD5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses multiple change-point detection in multivariate time series, allowing simultaneous changes in moments up to a prescribed fixed order p, including mean, covariance, and higher-order moments. By augmenting the observations with a constant component, the authors embed these moments into a single order-p tensor.

The proposed method combines preliminary detection based on tensor CUSUM statistics over seeded intervals (algorithm 1) followed by a refinement procedure (15)-(16) estimating the most significant projection direction of each previously detected change point. Under sub-Gaussian and temporal mixing assumptions, the authors establish consistency and a localization rate result, supported by a matching minimax lower bound. They also derive limiting distributions and construct asymptotically valid confidence intervals for vanishing moment jumps.

The approach is evaluated on five synthetic scenarios involving mean, variance, and skewness changes, and on two Nasdaq-100 financial datasets. Comparisons with four existing methods demonstrate promising detection and localization performance.

Critical strengths

The augmented tensor representation provides an appealing way to combine information from different moment orders and exploit it within a unified change-point detection procedure. Its ability to detect higher-order changes, including those leaving the mean and covariance unchanged, is particularly valuable for general distributional change-point detection. The subsequent refinement procedure, which improves the localization of each preliminarily detected change point by estimating the most significant projection direction of the associated moment jump, also deserves attention.

The theoretical analysis is substantial, encompassing consistent multiple change-point detection, optimal localization rates, limiting distributions, and confidence intervals under temporal dependence.

Critical weaknesses

The computational complexity and practical feasibility are insufficiently addressed.

The proposed algorithm requires repeated evaluations of tensor operator norms and optimization of higher-order tensor projections. For p≥3, these generally involve challenging nonconvex optimization problems. The manuscript does not adequately explain how these problems are solved numerically, whether global optima can be obtained, how convergence to local optima may affect the results, or how approximation errors may impact the theoretical guarantees.

More importantly, the computational complexity of the proposed algorithm, in terms of both processing time and memory requirements, is not analyzed. These important aspects are largely overlooked throughout the manuscript. This omission is particularly concerning given the authors' methodological choice to exploit an extensive representation of the statistical information contained in the observations, potentially involving high-dimensional tensors whose size grows rapidly with both the observation dimension and the moment order. While this approach is mathematically appealing, it inevitably raises questions about computational tractability, memory requirements, and scalability.

A careful complexity analysis, together with an explanation of the numerical implementation and its computational costs, should have been an integral part of the study. Without such an analysis, it is difficult to assess not only the practical feasibility of the proposed approach but also its actual value relative to existing methods.

The experimental validation and comparison with existing methods remain limited.

Indeed, the experimental evaluation is restricted to D=100, and p in {2,3}, and five synthetic configurations. No systematic runtime or dimension-scaling comparisons are provided. Since high-dimensional applicability is an important motivation of the work, this omission makes it difficult to assess its practical significance.

Furthermore, the selection of competing methods appears somewhat restrictive. In particular, the authors could have considered additional kernel-based change-point detection methods. For instance, the approach presented in [1] shares several methodological features with the proposed algorithm, notably the use of a dyadic, multiscale strategy for change-point detection and localization. Including such methods in the experimental comparisons would provide a more comprehensive assessment of the relative merits of the proposed approach and help better establish its practical advantages over existing alternatives.

[1] Florian Kalinke and Shakeel Gavioli-Akilagun. Optimal online change detection via random Fourier features. In Advances in Neural Information Processing Systems, 2026.

Questions for authors

How are the tensor operator norms and maximizing projection directions computed in practice, particularly for p≥3? Are global optima guaranteed, or are approximate iterative methods employed? In the latter case, how do convergence to local optima and approximation errors affect the theoretical guarantees?

Could the authors provide a computational complexity analysis of the proposed algorithm, covering both processing time and memory requirements as functions of the sample size (n), dimension (D), and moment order (p)? How does the method scale in practice as these parameters increase, and how does its computational cost compare with existing approaches?

Could the authors extend the experiments to investigate the influence of dimension and moment order on detection performance and computational cost? Furthermore, could they justify their selection of competing methods and consider including additional kernel-based approaches, particularly [1]?

4 comments

Critical Weakness 1 and Questions 1 and 2: tensor computation and approximation guarantees

Algorithm 2 in Appendix H adapts the shifted power method of Kolda and Mayo (2011) to approximate tensor operator norms and maximizing directions. Tensor CUSUMs and moment jumps have the form A=∑i∈IwiYi⊗pA=\sum_{i\in I}w_iY_i^{\otimes p}, where Yi=(1,Xi⊤)⊤Y_i=(1,X_i^\top)^\top and II indexes the observations in the interval. We evaluate f(u)=∑i∈Iwi(Yi⊤u)pf(u)=\sum_{i\in I}w_i(Y_i^\top u)^p directly from the observations and seek the largest ∣f(u)∣|f(u)| over unit vectors.

Computation. For each random unit-vector initialization and each sign ξ∈−1,1\xi\in\\{-1,1\\}, the routine iterates the following two steps with a shift β>0\beta>0: first compute an unnormalized update vector gg from the current direction uu, then normalize gg to obtain the next direction:

\qquad u\leftarrow g/\\|g\\|_2.$$ Thus $g$ is an intermediate vector, recomputed at every iteration; its normalized value replaces $u$ in the next iteration. After the iterations, the routine evaluates $|f(u)|$ at the final direction from each start and sign, and retains the candidate with the largest value. This implementation avoids storing $(D+1)^p$ tensor entries. For $M$ starts and $J$ iterations per start, the computational cost is $O(MJm(D+p))$, where $m=|I|$. Processing the starts sequentially requires only $O(m(D+1))$ memory. These costs are per tensor and exclude the additional work required for scanning and validation. **Dimension-scaling experiment (Section I.1).** With $n=1600$, five affected coordinates, and fixed tuning with $p=3$, mean detector runtimes are: | Dimension $D$ | 5 | 20 | 50 | 100 | |---|---:|---:|---:|---:| | Time (seconds) | 5.90 | 8.67 | 13.53 | 21.34 | The averages weight the four change designs equally, with 100 repetitions per design and dimension. Timing excludes data generation and parameter selection. The descriptive fit $\widehat t(D)=5.30+0.1614D$ seconds ($R^2=0.9993$) closely describes the observed growth over this range. The experiment fixes $n$ and $p$, so it does not establish empirical scaling in those parameters. Moreover, the iteration and initialization budgets needed for a given approximation accuracy may depend on $p$ and $D$. **Comparison with RFF-MMD.** A separate benchmark uses 100 paired repetitions with $n=1600$, $D=50$, Moment Tensor fixed at $(p,\tau)=(2,3)$, and RFF-MMD's default tuning. Mean (sample SD) detector runtimes are 0.267 (0.026) seconds for Moment Tensor and 0.201 (0.029) seconds for RFF-MMD. Moment Tensor takes approximately 1.33 times as long, an additional 0.066 seconds per dataset. Both timings exclude data generation and parameter selection. This measures fixed-tuning detector cost and is separate from the adaptive detection comparisons in Section I.3. **Global convergence is conditional.** For $p\geq3$, finding the tensor operator norm and maximizing directions remain a nonconvex optimization problem. A finite collection of arbitrary starts does not guarantee a global optimum. Section H.3 proves geometric convergence to an empirical global maximizer when the population objective has the stated separation, the empirical perturbation is sufficiently small, the shift is prescribed as in the analysis, and an initialization lies sufficiently close to the population maximizing direction. The neighborhood radius is independent of $D$ for fixed $p$ and fixed separation and shift constants. This does not imply that a fixed number of random starts succeeds uniformly in $D$. **Approximation guarantees.** Section H.1 shows that computed statistics remain below the exact norms on no-change intervals. On suitable intervals isolating one change, it gives a factor-two norm approximation at the true change with high probability, under its signal, separation, shift, and initialization conditions. With sufficiently large threshold and signal constants, these controls preserve the preliminary localization bound. For refinement, Section H.2 gives, under its assumptions and sufficiently accurate preliminary estimates, $$d_\pm(u_k,\nu_k)\leq 2Cq^J+\frac{2p\epsilon_n}{c_0\kappa_k}, \qquad 0<q<1,$$ with high probability. Here $J$ is the iteration count, $d_\pm$ is Euclidean distance up to sign, $\epsilon_n$ bounds moment-estimation error, $\kappa_k$ is the jump magnitude, and $c_0$ is the separation constant. Choosing $J$ so that the computational term is no larger than the statistical term for every change preserves the accuracy required for the refined localization rate. Thus approximate optimization is sufficient under the stated conditions. Convergence to a poorer local optimum outside those conditions is not covered by these guarantees.

Critical Weakness 2 and Question 3: expanded experiments

The revision broadens the dimensions, moment orders, change types, and dependence levels, and adds the requested RFF-MMD comparator. All additional simulations use n=1600n=1600 and 100 repetitions per configuration.

Dimension and moment order (Section I.1). We study D∈5,20,50,100D\in\\{5,20,50,100\\}, with a change at 880 in the first five coordinates. Independent stationary Gaussian AR(1) latent series have coefficient 0.4 and continue across the change. The four designs change the mean from 0 to 0.5, variance from 1 to 1.5, skewness sign with unchanged mean and covariance, or all three moments (mean 0→0.20\to0.2, variance 1→1.441\to1.44, and reversed skewness). Moment order and threshold are selected from p∈1,2,3,4p\in\\{1,2,3,4\\} and τ∈1,2,3,4,5,6,8,10,15,20\tau\in\\{1,2,3,4,5,6,8,10,15,20\\}.

Mean Hausdorff errors, in observations, for Moment Tensor are:

Change typeD=5D=5D=20D=20D=50D=50D=100D=100
Mean7.067.576.857.06
Variance30.3279.5294.78142.42
Skewness18.0017.7918.5523.97
Combined12.8513.6520.0728.17

Moment Tensor has the lowest observed mean error in 12 of 16 configurations. Mean-change BS performs slightly better for mean changes at D=5,20D=5,20, and CPWZ performs better for variance changes at those dimensions. Variance-change localization deteriorates markedly as irrelevant coordinates are added, while mean-change errors remain stable and higher-order changes remain detectable.

For pure skewness changes, p=3p=3 is selected in 78%, 73%, 82%, and 98% of runs at the four dimensions. For mean changes, p=1p=1 is most frequently selected, in 71%–93% of runs. These results assess adaptive moment selection, rather than a controlled comparison fixing each order separately. The fixed-p=3p=3 dimension runtime study is described in Part 2.

Section I.2 also crosses six signal strengths s∈0.1,0.2,0.4,0.6,0.8,1s\in\\{0.1,0.2,0.4,0.6,0.8,1\\} with five latent AR coefficients ρ∈0,0.2,0.4,0.6,0.8\rho\in\\{0,0.2,0.4,0.6,0.8\\} at D=50D=50, reporting all 30 configurations. It documents improved localization with stronger signals and deterioration under stronger dependence.

Why these competitors? Mean-change BS targets mean shifts, CPWZ combines mean and covariance statistics, changeAUC uses classification to detect distributional changes, and MNSBS combines density estimation with seeded segmentation under dependence. We now also include the suggested kernel method, RFF-MMD. Its online procedure is adapted to retrospective change-location comparison with default tuning. We compare estimated locations, rather than online alarm delays. CPWZ is omitted from multiple-change comparisons because the implementation used is restricted to one change.

Continuous and discrete multiple changes (Section I.3). Changes occur at 533 and 1067. Standardized, reflected, and rescaled quantile transformations generate three regimes with changing mean, variance, and skewness. The Beta(1,4.16) design has D=50D=50, 33 affected coordinates, and latent AR coefficient 0.4. The Binomial(5,0.12) design has D=100D=100, 68 affected coordinates, and coefficient 0.25, with no added jitter. All methods estimate the number and locations freely.

Cells below give mean (SD) Hausdorff error in observations / percentage estimating the correct count.

MethodBetaBinomial
Moment Tensor40.49 (112.02) / 96%5.08 (6.73) / 100%
RFF-MMD43.00 (48.72) / 99%23.01 (11.43) / 100%
changeAUC197.35 (140.70) / 24%132.45 (126.77) / 45%
Mean-change BS486.39 (7.55) / 0%85.21 (164.01) / 80%
MNSBS519.74 (52.24) / 7%563.82 (53.67) / 17%

Moment Tensor and RFF-MMD have comparable mean errors in the Beta setting, where RFF-MMD has the higher correct-count frequency. In the Binomial setting, Moment Tensor's mean error is 77.9% smaller, with both methods recovering the count in all runs.

Section J also compares the five multiple-change methods on four vibration recordings with D=3D=3 or 6, using documented activity or road-surface transitions for evaluation. Moment Tensor has the smallest Hausdorff error on three recordings, while MNSBS performs best on one the the recordings. All four records select p=2p=2. Together, the expanded experiments identify advantages and limitations across distinct change types and distributions.

I thank the authors for taking my comments into account and for comparing their method with the one I suggested through a bibliographic reference. The authors are right that this method has batch-type-like characteristics. However, I remain convinced that the baselines and competitors chosen by the authors are not the most representative or the most effective ones, even though the authors justify their choice by noting that each belongs to a different category of methods. For the applications considered, I believe that lightweight kernel methods such as MMD, of which the method I suggested is a more complex extension, perform just as well. In the suggested reference, MMD is described at length throughout the first part of the paper; it could have been tested. Other standard methods such as uLSIF, RuLSIF and KLIEP would also have been natural choices. Nevertheless, I acknowledge the authors' work, both on the proposed method and on the revision.

Point: Additional comparisons with MMD-based and density-ratio methods.

We thank the reviewer for suggesting stronger baselines. We extend our prior RFF-MMD comparison and add Gaussian-kernel MMD, uLSIF, RuLSIF, and KLIEP on the same single-change datasets.

Gaussian-kernel MMD (MMD-SW). Following Gretton et al. (2012) [1], we use the Gaussian kernel

kσ(x,y)=exp⁡(−∥x−y∥222σ2).k_\sigma(x,y)=\exp\left(-\frac{\|x-y\|_2^2}{2\sigma^2}\right).

We compute biased empirical MMD$^2$ (including diagonal terms) between adjacent 80-observation windows, scanning every four steps. Bandwidth uses the median heuristic on up to 512 observations. We report the maximum-scoring split if its statistic exceeds a calibrated threshold; otherwise, no change.

Online RFF-MMD. We adapt Kalinke and Gavioli-Akilagun, Optimal Online Change Detection via Random Fourier Features (NeurIPS 2025) [2]. Our detector uses 1,000 Gaussian Fourier frequencies and sequential dyadic MMD tests. At the first alarm, it reports the maximizing candidate split, not the alarm time. Bandwidth is selected from independent stationary reference data. RFF-MMD and MMD-SW differ in scanning and decision rules, not just kernel approximation. Both use separately calibrated thresholds from stationary simulations preserving temporal dependence (nominal false-alarm level 0.05).

Experimental settings. As in Section I.1, we use n=1600n=1600, a change at η=880\eta=880, D∈{5,20,50,100}D\in\{5,20,50,100\}, and 100 repetitions per setting. Only the first five coordinates change. Latent coordinates are independent, unit-variance stationary Gaussian AR(1) processes with coefficient ρ=0.4\rho=0.4, continuing across the change. Repetitions use fresh innovations; all methods use identical series and original seeds.

For Z∼N(0,1)Z\sim N(0,1), define the mean-zero, unit-variance transformation

Sa(Z)=Z+a∣Z∣−a2/π1+a2(1−2/π).S_a(Z)=\frac{Z+a|Z|-a\sqrt{2/\pi}}{\sqrt{1+a^2(1-2/\pi)}}.

Affected coordinates undergo one of four transitions: mean, Z→Z+0.5Z\to Z+0.5; variance, Z→1.5 ZZ\to\sqrt{1.5}\,Z; skewness, S−0.5(Z)→S0.5(Z)S_{-0.5}(Z)\to S_{0.5}(Z); combined, S−0.5(Z)→0.2+1.2S0.5(Z)S_{-0.5}(Z)\to 0.2+1.2S_{0.5}(Z). The skewness-only case preserves mean and covariance; the combined case changes all three moments.

Table R1. Mean Hausdorff error over 100 repetitions (smaller is better). Bold entries indicate the best mean among the six methods.

ChangeDDTensorRFF-MMD [2]MMD-SWuLSIFRuLSIFKLIEP
Mean57.0640.96362.92391.44398.52470.76
Mean207.5737.12558.20604.80567.48695.36
Mean506.8580.96632.24689.12674.08693.60
Mean1007.06136.64678.32704.00678.96685.28
Variance530.32197.44712.28693.80696.64678.80
Variance2079.52709.44718.08720.00720.00714.92
Variance5094.78720.32707.84701.88706.52714.32
Variance100142.42718.08710.08710.60689.84701.92
Skewness518.00151.04714.84716.48720.00713.92
Skewness2017.79712.96704.60698.44698.44708.76
Skewness5018.55716.16699.52696.48700.80707.88
Skewness10023.97719.68720.00701.04710.80704.32
Combined512.85127.68720.00705.68709.96692.40
Combined2013.65579.20716.16712.60712.32698.92
Combined5020.07709.60711.92699.20699.20707.60
Combined10028.17700.96715.56703.92696.12694.88

Moment Tensor has the lowest mean error in all 16 settings. RFF-MMD is stronger than MMD-SW for mean changes but has mean errors of 37.12–136.64 versus 6.85–7.57 for Tensor. A missed detection receives finite Hausdorff error min⁡(880,1600−880)=720\min(880,1600-880)=720; values near 720 largely indicate missed detections, and 720 is not an upper bound. These results compare the implemented procedures, not the entire MMD family. We will add SDs, detection frequencies, and tuning details.

References: [1] Gretton et al. (2012), A Kernel Two-Sample Test, JMLR 13:723–773. [2] Kalinke and Gavioli-Akilagun (2025), Optimal Online Change Detection via Random Fourier Features, NeurIPS 38.

Reviewer_qXwUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a unified tensor-based framework for detecting multiple change points in multivariate time series by monitoring joint moments up to a fixed order p. A seeded tensor-CUSUM procedure first identifies preliminary changes, followed by a local refinement step based on an estimated leading jump direction. Under sub-Gaussian tails and temporal mixing, the authors establish localization guarantees, a matching minimax lower bound, limiting distributions, and confidence intervals in the vanishing-jump regime. Simulations cover mean, variance, and skewness changes, while a Nasdaq stock-return analysis illustrates the method on real data.

Critical strengths
  1. The framework unifies detection of mean, covariance, and higher-order moment changes, including under temporal dependence and growing dimension, which is a meaningful extension beyond methods focused on a single moment type.
  2. The theoretical contribution is substantial: the paper provides a localization rate, a matching minimax lower bound, and change-point inference results. Simulations show strong localization and change-count recovery in the tested settings.
Critical weaknesses
  1. The empirical evaluation is narrow relative to the broad theoretical claims: simulations use a small set of designed scenarios, while the real-data study covers only two stock panels. More varied dimensions, dependence strengths, jump structures, and comparisons would better establish practical robustness.
  2. The real-data analysis lacks ground-truth change points, and the method is configured to return one change in each period. Consequently, the financial-event alignment is suggestive rather than strong evidence of superior detection or calibration.
  3. The method relies on estimating and optimizing higher-order tensor operator norms and projection directions. The paper does not sufficiently characterize computational cost and numerical stability as dimension or moment order increases, which is important for practical use.
  4. The data-driven selection of moment order and threshold is described through odd–even validation and a moment-kernel score, but its sensitivity and stability across dependence patterns and weak signals are not fully established.
Questions for authors
  1. How does runtime and numerical accuracy scale with dimension D and moment order p, and what optimization procedure is used for the tensor operator norm in practice?
  2. How sensitive are the selected (p, τ) and localization results to the validation split, candidate grid, and strength of temporal dependence?
  3. Can the authors provide evidence on real or semi-synthetic data with known change points, or additional real-data analyses that do not fix the number of changes in advance?
4 comments

Critical Weakness 1: broader empirical evaluation

We provide additional simulations in Section I of the revised manuscript. Every setting uses n=1600n=1600 and 100 random repetitions. We report all settings of the stated configurations with mean Hausdorff distance between the true and estimated change-point sets, standard deviations (SDs), and change-count frequencies.

1. Change types and dimension (Section I.1).

We generate independent, unit-variance stationary Gaussian AR(1) latent coordinates with ρ=0.4\rho=0.4. To achieve independence across repetitions, we generate fresh Gaussian initial values and innovations in each repetition. The true change point is set to be 880. We vary the dimension over D∈5,20,50,100D\in\\{5,20,50,100\\} while restricting changes to the first five coordinates.

For a standard Gaussian random variable ZZ, define the standardized transformation Sa(z)=[z+a∣z∣−a2/π]/1+a2(1−2/π)S_a(z)=[z+a|z|-a\sqrt{2/\pi}]/\sqrt{1+a^2(1-2/\pi)}, so that Sa(Z)S_a(Z) has mean zero and variance one. Write mk=E[Xk]m_k=\mathbb{E}[X^k] for the kkth raw moment of an affected coordinate and let γ=2/π[3/2+(4/π−1)/8]/[1+(1−2/π)/4]3/2\gamma=\sqrt{2/\pi}[3/2+(4/\pi-1)/8]/[1+(1-2/\pi)/4]^{3/2} denote the skewness of S0.5(Z)S_{0.5}(Z). We consider four pre-change to post-change transitions in each affected coordinate:

  • Mean change: Z→Z+0.5Z\to Z+0.5. The mean increases from 00 to 0.50.5; variance and skewness remain 11 and 00, respectively. Raw moments: (m1,m2,m3):(0,1,0)→(0.5,1.25,1.625)(m_1,m_2,m_3):(0,1,0)\to(0.5,1.25,1.625).
  • Variance change: Z→1.5ZZ\to\sqrt{1.5}Z. The variance increases from 11 to 1.51.5; mean and skewness remain 00. Raw moments: (m1,m2,m3):(0,1,0)→(0,1.5,0)(m_1,m_2,m_3):(0,1,0)\to(0,1.5,0).
  • Skewness change: S−0.5(Z)→S0.5(Z)S_{-0.5}(Z)\to S_{0.5}(Z). Skewness and the third central moment change from −γ-\gamma to γ\gamma; mean and variance remain 00 and 11, respectively. Raw moments: (m1,m2,m3):(0,1,−γ)→(0,1,γ)(m_1,m_2,m_3):(0,1,-\gamma)\to(0,1,\gamma).
  • Combined change: S−0.5(Z)→0.2+1.2S0.5(Z)S_{-0.5}(Z)\to0.2+1.2S_{0.5}(Z). Mean, variance, and skewness change from (0,1,−γ)(0,1,-\gamma) to (0.2,1.44,γ)(0.2,1.44,\gamma); the third central moment changes from −γ-\gamma to 1.728γ1.728\gamma. Raw moments: (m1,m2,m3):(0,1,−γ)→(0.2,1.48,0.872+1.728γ)(m_1,m_2,m_3):(0,1,-\gamma)\to(0.2,1.48,0.872+1.728\gamma).

Across the four change types and four dimensions, Moment Tensor has the lowest mean Hausdorff error in 12 of the 16 settings compared with CPWZ, changeAUC, mean-change BS, and MNSBS. The results by change type are:

  • Mean changes: mean-change BS has the lowest error at D=5,20D=5,20, whereas Moment Tensor has the lowest at D=50,100D=50,100.
  • Variance changes: CPWZ has the lowest error at D=5,20D=5,20, whereas Moment Tensor has the lowest at D=50,100D=50,100.
  • Skewness changes: Moment Tensor has the lowest error at all four dimensions, despite the absence of a change in the mean vector or contemporaneous covariance matrix.
  • Combined changes: Moment Tensor has the lowest error at all four dimensions.

2. Signal strength and dependence (Section I.2).

We extend the combined-change design in Section I.1 at D=50D=50 by varying signal strength s∈{0.1,0.2,0.4,0.6,0.8,1}s\in\{0.1,0.2,0.4,0.6,0.8,1\} and temporal dependence ρ∈{0,0.2,0.4,0.6,0.8}\rho\in\{0,0.2,0.4,0.6,0.8\}. The affected coordinates follow S−0.5s(Z)→0.2s+1+0.44sS0.5s(Z)S_{-0.5s}(Z)\to0.2s+\sqrt{1+0.44s}S_{0.5s}(Z), recovering the original combined change at s=1s=1. The change location, affected coordinates, latent AR(1) construction, and odd-even tuning rule remain as in Section I.1. We use the evaluation conventions stated above; correct-count frequency is the percentage of repetitions returning exactly one change.

At ρ=0.4\rho=0.4, mean error decreases from 374.20 to 20.07 as ss increases from 0.1 to 1, while the correct-count frequency rises from 85% to 100%. For s≥0.4s\geq0.4, the correct-count frequency is at least 96% across all tested dependence levels. At s=1s=1, mean errors for ρ=0,0.2,0.4,0.6,0.8\rho=0,0.2,0.4,0.6,0.8 are 15.70, 15.00, 20.07, 38.10, and 96.00, respectively; correct-count frequencies are 100% except at ρ=0.8\rho=0.8, where the frequency is 99%. Thus stronger signals improve localization, whereas strong temporal dependence can substantially increase localization error even when the change count is recovered.

3. Continuous and discrete multiple changes (Section I.3).

For two changes at 533 and 1067, we use:

  • Beta(1,4.16): D=50D=50, 33 affected coordinates, and ρ=0.4\rho=0.4, with regime means (0,0.25,0.50) and variances (1,1.44,1.89).
  • Binomial(5,0.12): D=100D=100, 68 affected coordinates, and ρ=0.25\rho=0.25, with means (0,0.20,0.40) and variances (1,1.50,2.00).

We compare with BS, changeAUC, MNSBS and RFF-MMD (Kalinke and Gavioli-Akilagun, 2026). CPWZ is omitted because its implementation targets one change. Against the best competitor by mean error, Moment Tensor achieves 40.49 (SD 112.02) versus 43.00 (48.72) for RFF-MMD in Beta, and 5.08 (6.73) versus 23.01 (11.43) for RFF-MMD in Binomial. Its correct-count frequencies are 96% and 100%, respectively. RFF-MMD has a higher correct-count frequency of 99% in Beta, with the second best mean localization error.

Critical Weakness 2 and Question 3: real data with reference changes

In the revised manuscript, we added four multivariate smartphone vibration recordings with the true change points induced by the independently recorded activity or road-surface transitions (Section J).

Data and reference changes.

Recordings 03 and 04 follow walking-running-walking and walking-running-walking-running-walking, respectively. Both combine three-axis G-force with angular velocity (03) or linear acceleration (04). Motorbike recording 06 follows a stone-asphalt-stone route, and car recording 08 follows an irregular route over both surfaces. Both measure three-axis G-force.

The true changes for evaluation are the midpoints of independently recorded transition intervals. The table lists channel counts DD, sample sizes after preprocessing nn, and times in seconds from recording start.

RecordingDnAnalysis period (s)Reference changes (s)
036367812.05-195.9473, 133
046309019.55-174.0349, 80, 110, 140
063408211.74-215.8050, 180
083427192.70-306.23103.5, 198, 280

Preprocessing and evaluation.

We sort the irregular timestamps and resample on a 20 Hz grid starting at 0.05 s, using the most recent sensor vector at or before each grid time (last row for tied timestamps). Within the periods above, we subtract each channel's mean and divide by its population SD. All methods receive identical inputs.

Moment Tensor selects its moment order and threshold by odd-even validation, while competitors use implementation defaults. All methods estimate the unknown number and locations of changes. We retain every estimate and use reference changes only for evaluation.

Results. Entries give estimated count / Hausdorff distance in seconds. For nonempty sets, this is the maximum distance from an estimated or reference change to its nearest change in the other set.

Method03040608
Moment Tensor2/0.454/2.102/0.353/19.60
Mean-change BS1/59.504/21.750/38.2520/43.90
changeAUC5/38.058/20.208/63.6510/45.45
MNSBS3/16.104/1.307/62.558/37.90
RFF-MMD2/5.151/72.952/12.951/91.60

Moment Tensor selects the moment order p=2p=2 and recovers the reference count in every recording. It has the smallest distances on 03, 06, and 08, with all estimates on 03 and 06 within the observer intervals. MNSBS is most accurate on 04 (1.30 versus 2.10 s for Moment Tensor).

Critical Weakness 3 and Question 1: computation and numerical accuracy

Algorithm 2 in Appendix H adapts the shifted power method of Kolda and Mayo (2011) to approximate tensor operator norms and maximizing directions. Tensor CUSUMs and moment jumps have the form A=∑i∈IwiYi⊗pA=\sum_{i\in I}w_iY_i^{\otimes p}, where Yi=(1,Xi⊤)⊤Y_i=(1,X_i^\top)^\top and II indexes the observations in the interval. We evaluate f(u)=∑i∈Iwi(Yi⊤u)pf(u)=\sum_{i\in I}w_i(Y_i^\top u)^p directly from the observations and seek the largest ∣f(u)∣|f(u)| over unit vectors.

Computation. For each random unit-vector initialization and each sign ξ∈−1,1\xi\in\\{-1,1\\}, the routine iterates the following two steps with a shift β>0\beta>0: first compute an unnormalized update vector gg from the current direction uu, then normalize gg to obtain the next direction:

\qquad u\leftarrow g/\\|g\\|_2.$$ Thus $g$ is an intermediate vector, recomputed at every iteration; its normalized value replaces $u$ in the next iteration. After the iterations, the routine evaluates $|f(u)|$ at the final direction from each start and sign, and retains the candidate with the largest value. This implementation avoids storing $(D+1)^p$ tensor entries. For $M$ starts and $J$ iterations per start, the computational cost is $O(MJm(D+p))$, where $m=|I|$. Processing the starts sequentially requires only $O(m(D+1))$ memory. These costs are per tensor and exclude the additional work required for scanning and validation. **Approximation guarantees.** Under Appendix H's additional statistical assumptions and prescribed shift and initialization conditions, H.1 controls no-change statistics and provides a factor-two norm approximation at the true change on suitable single-change intervals, with high probability. H.3 establishes geometric convergence to an empirical global maximizer from a neighborhood of the population maximizing direction. Its radius is independent of $D$ for fixed $p$ and fixed separation and shift constants. H.2 gives, with high probability, $$d_\pm(u_k,\nu_k)\leq 2Cq^J+\frac{2p\epsilon_n}{c_0\kappa_k}, \qquad 0<q<1.$$ Here $d_\pm$ is Euclidean distance up to sign, $\epsilon_n$ bounds moment-estimation error, $\kappa_k$ is the jump magnitude, and $c_0$ is the separation constant. The terms bound computational and statistical error, respectively. Choosing $J$ so the first does not exceed the second for any $k$ preserves the refined localization rate. The constants depend on $p$, and a fixed random-restart budget need not satisfy the initialization condition uniformly in $D$. **Empirical runtime.** With $n=1600$, five active coordinates, and $p=3$, mean runtimes are 5.90, 8.67, 13.53, and 21.34 seconds for $D=5,20,50,100$, respectively. Over this range, the descriptive linear fit is $\widehat t(D)=5.30+0.1614D$ seconds ($R^2=0.9993$). This approximately linear growth in $D$ is in line with the theoretical per-tensor computational cost $O(MJm(D+p))$, which is linear in $D$ when $M$, $J$, $m$, and $p$ are fixed.

Critical Weakness 4 and Question 2: tuning and sensitivity

Appendix G.1 fits candidate segmentations on odd-indexed observations and validates them on even-indexed observations. The pair (p,τ)(p,\tau) with the lowest score is then applied to the full series.

Section I.1 and the supplementary mean-change study below use p∈1,2,3,4p\in\\{1,2,3,4\\} and τ∈1,2,3,4,5,6,8,10,15,20\tau\in\\{1,2,3,4,5,6,8,10,15,20\\}. Here pp sets the highest moment order monitored, while τ\tau controls detection sensitivity.

Selection across change types and dimensions (Section I.1). The table gives the most frequently selected pp in each setting, with selection percentages over 100 repetitions.

Change typeD=5D=20D=50D=100
Mean1 (93%)1 (87%)1 (81%)1 (71%)
Variance2 (67%)2 (50%)2 (45%)1 (44%)
Skewness3 (78%)3 (73%)3 (82%)3 (98%)
Combined1 (32%)3 (35%)2 (36%)3 (47%)

Order p=3p=3 is most frequently selected for skewness changes at every dimension. For variance changes at D=100D=100, p=1p=1 is most frequently selected despite an unchanged mean. In these baseline settings, the most frequently selected thresholds τ\tau range from 3 to 15.

Under combined changes, every candidate order contains a population signal through the changing mean. Variation in selected pp therefore need not imply poor localization. The mean Hausdorff errors, averaged over all runs, are 12.85, 13.65, 20.07, and 28.17 as DD increases.

Supplementary mean-change selection study. With n=1600n=1600, D=5D=5, and a change at 880, all five coordinate means shift from 0 to aa. The table crosses four signal strengths aa with four dependence levels ρ\rho, using the grid above and 100 repetitions per configuration. Cells show the most frequently selected pp (%) / τ\tau (%). Frequencies are calculated separately for each parameter. For tied thresholds, the percentage applies to each value.

ρ\rho / aa0.30.60.91.2
0.01 (68) / 5 (38)1 (80) / 10 (68)1 (90) / 15 (60)1 (86) / 20 (62)
0.31 (60) / 4 (46)1 (84) / 6 (50)1 (90) / 10 (76)1 (90) / 15 (50)
0.61 (62) / 3,4 (42 each)1 (82) / 5 (42)1 (94) / 8 (54)1 (98) / 10 (66)
0.91 (64) / 3,4 (40 each)1 (66) / 4 (44)1 (84) / 5 (44)1 (86) / 6 (50)

Order p=1p=1 is most frequently selected in all 16 configurations, with frequencies of 60–68% at a=0.3a=0.3 and 84–98% at a∈0.9,1.2a\in\\{0.9,1.2\\}. The most frequently selected thresholds vary across settings. These are parameter-selection frequencies, separate from the localization results below.

Localization and count sensitivity (Section I.2). The combined-change study uses n=1600n=1600, D=50D=50, five affected coordinates, and a change at 880. It crosses s∈0.1,0.2,0.4,0.6,0.8,1s\in\\{0.1,0.2,0.4,0.6,0.8,1\\} with AR(1) coefficients ρ∈0,0.2,0.4,0.6,0.8\rho\in\\{0,0.2,0.4,0.6,0.8\\}. Affected means change from 0 to 0.2s0.2s, variances from 1 to 1+0.44s1+0.44s, and skewness changes sign.

At ρ=0.4\rho=0.4, mean Hausdorff error falls from 374.20 to 20.07 as ss increases from 0.1 to 1, with correct-count frequency rising from 85% to 100%. At s=1s=1, errors along the increasing ρ\rho grid are 15.70, 15.00, 20.07, 38.10, and 96.00. Counts are correct in 100% of runs except at ρ=0.8\rho=0.8, where the frequency is 99%.

Reviewer_AMnuOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a method for multiple change-point detection for multiple times series. The method uses 2 stages: (1) tensor CUMSUM to rougly estimate the number of change points and their location; (2) refinement. The paper evaluates the method on synthetic dataset and a financial real data.

Critical strengths
  1. The problem settings seem interesting.

  2. The paper tries to use theoretical approach to address the problem.

Critical weaknesses
  1. The paper is hard to follow. For example, the method name Tensor is only introduced in the experiment sections. Furthermore, in line 418-419, the paper uses "Hausdorf distance" to evaluate the performance. The paper should clearly state, i.e., "distance between A and B". And what are A and B in this case?

  2. The paper contains a lengthy theoretical derivation in Appendix. But it lacks of clear discussion of how to convey and translate such results into the Algorithm 1.

  3. The experiments seems sparse without thorough baselines, ablation studies. Some standard data sets including well-log data, coal mine disaster data [1] are missing. The synthetic data seems to cherry pick with limited scenarios. Can the algorithm perform well under a stress test of randomly generated data?

References [1] Bayesian Online Changepoint Detection, 2007

Questions for authors
  1. What is the main reason to introduce additional dimension (1,Xt)∈RD+1(1, X_t) \in \mathbb{R}^{D+1}? What would be the effect of not having the additional dimension?
3 comments

Critical Weakness 1: Method name and Hausdorff distance

We use "Moment Tensor" consistently in the experiments and believe it clearly describes our method. Our method relies on the moment-tensor representation introduced in Section 1: both the detection and refinement procedures in Section 2 are built on this representation.

Appendix G.2 defines the Hausdorff distance between the true and estimated change-point sets,

B=\\{\widehat\eta_1,\ldots,\widehat\eta_{\widehat K}\\}.$$ Table 1 and the accompanying discussion link to this definition. We will also specify both sets in the caption to make the table self-contained. **Critical Weakness 2: Connection between theory and Algorithm 1** Theorem 1 in Section 2.1 analyzes Algorithm 1 exactly as stated: its preliminary estimates are defined as the algorithm's outputs. Under Assumptions 1 and 2 and the prescribed threshold, it guarantees the correct change count and localization error at most $3\Delta_{\min}/64$ with probability at least $1-n^{-5}$. Its proof in Appendix B follows the algorithm's thresholding, shortest-interval selection, and overlap removal. The discussion following Algorithm 1 connects its steps to the theoretical analysis: the central-half restriction controls the instability of moment estimates near interval boundaries; the threshold accounts for interval length, moment order, and temporal dependence; and shortest-interval selection with overlap removal isolates changes while avoiding repeated detections. These are the same steps analyzed in the proof of Theorem 1. The remark after Theorem 1 further connects Algorithm 1's guarantee to the refinement procedure in Section 2.2. The preliminary localization bound ensures a one-to-one correspondence between estimated and true changes and provides sufficient accuracy to select moment-fitting intervals lying within true segments and refinement windows containing individual changes. Theorem 2 then establishes the refined localization bound under additional signal and separation assumptions. **Question 1: Why add a constant coordinate?** Remark 1 in Section 1 illustrates the role of the constant coordinate with an explicit example for $p=3$: tensor entries involving two, one, or no constant coordinates recover first-, second-, or third-order moments, respectively. More generally, the augmentation embeds **all joint moments through order $p$** in a single order-$p$ tensor. For $Y_t=(1,X_t^\top)^\top$ and $\mathcal M_t^{(p)}=\mathbb E Y_t^{\otimes p}$, we have, for $1\leq r\leq p$, $$(\mathcal{M}_t^{(p)})_{1,\ldots,1,i_1+1,\ldots,i_r+1} =\mathbb{E}\left[\prod_{j=1}^{r}X_{t,i_j}\right].$$ Here the first $p-r$ indices are $1$, and $i_j\in\\{1,\ldots,D\\}$. Thus, a change at any moment order up to $p$ changes $\mathcal M_t^{(p)}$, allowing these moments to be monitored jointly. The constant-only entry cancels in the CUSUM. Without the constant coordinate, $\mathbb E X_t^{\otimes p}$ contains only degree-$p$ moments and can miss lower-order changes.

Critical Weakness 3: Additional simulations

We performed additional numerical comparisons, in Sections I.1 and I.3 of the revised manuscript, across moment change types, dimensions, and continuous/discrete distributions. Every setting uses n=1600n=1600 and 100 random repetitions. We report all settings of the stated configurations with mean Hausdorff distance between the true and estimated change-point sets, standard deviations (SDs), and change-count frequencies.

1. Change types and dimension (Section I.1).

We generate independent, unit-variance stationary Gaussian AR(1) latent coordinates with ρ=0.4\rho=0.4. To achieve independence across repetitions, we generate fresh Gaussian initial values and innovations in each repetition. The true change point is set to be 880. We vary the dimension over D∈5,20,50,100D\in\\{5,20,50,100\\} while restricting changes to the first five coordinates.

For a standard Gaussian latent variable ZZ, define the standardized transformation Sa(z)=[z+a∣z∣−a2/π]/1+a2(1−2/π)S_a(z)=[z+a|z|-a\sqrt{2/\pi}]/\sqrt{1+a^2(1-2/\pi)}, so that Sa(Z)S_a(Z) has mean zero and variance one. Write mk=E[Xk]m_k=\mathbb{E}[X^k] for the kkth raw moment of an affected coordinate and let γ=2/π[3/2+(4/π−1)/8]/[1+(1−2/π)/4]3/2\gamma=\sqrt{2/\pi}[3/2+(4/\pi-1)/8]/[1+(1-2/\pi)/4]^{3/2} denote the skewness of S0.5(Z)S_{0.5}(Z). We consider four pre-change to post-change transitions in each affected coordinate:

  • Mean change: Z→Z+0.5Z\to Z+0.5. The mean increases from 00 to 0.50.5; variance and skewness remain 11 and 00, respectively. Raw moments: (m1,m2,m3):(0,1,0)→(0.5,1.25,1.625)(m_1,m_2,m_3):(0,1,0)\to(0.5,1.25,1.625).
  • Variance change: Z→1.5ZZ\to\sqrt{1.5}Z. The variance increases from 11 to 1.51.5; mean and skewness remain 00. Raw moments: (m1,m2,m3):(0,1,0)→(0,1.5,0)(m_1,m_2,m_3):(0,1,0)\to(0,1.5,0).
  • Skewness change: S−0.5(Z)→S0.5(Z)S_{-0.5}(Z)\to S_{0.5}(Z). Skewness and the third central moment change from −γ-\gamma to γ\gamma; mean and variance remain 00 and 11, respectively. Raw moments: (m1,m2,m3):(0,1,−γ)→(0,1,γ)(m_1,m_2,m_3):(0,1,-\gamma)\to(0,1,\gamma).
  • Combined change: S−0.5(Z)→0.2+1.2S0.5(Z)S_{-0.5}(Z)\to0.2+1.2S_{0.5}(Z). Mean, variance, and skewness change from (0,1,−γ)(0,1,-\gamma) to (0.2,1.44,γ)(0.2,1.44,\gamma); the third central moment changes from −γ-\gamma to 1.728γ1.728\gamma. Raw moments: (m1,m2,m3):(0,1,−γ)→(0.2,1.48,0.872+1.728γ)(m_1,m_2,m_3):(0,1,-\gamma)\to(0.2,1.48,0.872+1.728\gamma).

Across the four change types and four dimensions, Moment Tensor has the lowest mean Hausdorff error in 12 of the 16 settings compared with CPWZ, changeAUC, mean-change BS, and MNSBS. The results by change type are:

  • Mean changes: mean-change BS has the lowest error at D=5,20D=5,20, whereas Moment Tensor has the lowest at D=50,100D=50,100.
  • Variance changes: CPWZ has the lowest error at D=5,20D=5,20, whereas Moment Tensor has the lowest at D=50,100D=50,100.
  • Skewness changes: Moment Tensor has the lowest error at all four dimensions, despite the absence of a change in the mean vector or contemporaneous covariance matrix.
  • Combined changes: Moment Tensor has the lowest error at all four dimensions.

Thus, Moment Tensor leads in all eight skewness and combined-change settings and in all eight settings with D=50,100D=50,100.

2. Continuous and discrete multiple changes (Section I.3).

For two changes at 533 and 1067, we use:

  • Beta(1,4.16): D=50D=50, 33 affected coordinates, and ρ=0.4\rho=0.4, with regime means (0,0.25,0.50) and variances (1,1.44,1.89).
  • Binomial(5,0.12): D=100D=100, 68 affected coordinates, and ρ=0.25\rho=0.25, with means (0,0.20,0.40) and variances (1,1.50,2.00).

We compare with BS, changeAUC, MNSBS and RFF-MMD (Kalinke and Gavioli-Akilagun, 2026). CPWZ is omitted because its implementation targets one change. Against the best competitor by mean error, Moment Tensor achieves 40.49 (SD 112.02) versus 43.00 (48.72) for RFF-MMD in Beta, and 5.08 (6.73) versus 23.01 (11.43) for RFF-MMD in Binomial. Its correct-count frequencies are 96% and 100%, respectively. RFF-MMD has a higher correct-count frequency of 99% in Beta, with the second best mean localization error.

Critical Weakness 3: Additional real data

The suggested well-log and coal-mine datasets (Adams and MacKay, 2007, Bayesian Online Changepoint Detection) are univariate benchmarks whose true change-point locations are unknown. We therefore added four multivariate smartphone vibration recordings with the true change points induced by the independently recorded activity or road-surface transitions (Section J).

Data and reference changes.

Recordings 03 and 04 follow walking-running-walking and walking-running-walking-running-walking, respectively. Both combine three-axis G-force with angular velocity (03) or linear acceleration (04). Motorbike recording 06 follows a stone-asphalt-stone route, and car recording 08 follows an irregular route over both surfaces. Both measure three-axis G-force.

The true changes for evaluation are the midpoints of independently recorded transition intervals. The table lists channel counts DD, sample sizes after preprocessing nn, and times in seconds from recording start.

RecordingDnAnalysis period (s)Reference changes (s)
036367812.05-195.9473, 133
046309019.55-174.0349, 80, 110, 140
063408211.74-215.8050, 180
083427192.70-306.23103.5, 198, 280

Preprocessing and evaluation.

We sort the irregular timestamps and resample on a 20 Hz grid starting at 0.05 s, using the most recent sensor vector at or before each grid time (last row for tied timestamps). Within the periods above, we subtract each channel's mean and divide by its population SD. All methods receive identical inputs.

Moment Tensor selects its moment order and threshold by odd-even validation, while competitors use implementation defaults. All methods estimate the unknown number and locations of changes. We retain every estimate and use reference changes only for evaluation.

Results. Entries give estimated count / Hausdorff distance in seconds. For nonempty sets, this is the maximum distance from an estimated or reference change to its nearest change in the other set.

Method03040608
Moment Tensor2/0.454/2.102/0.353/19.60
Mean-change BS1/59.504/21.750/38.2520/43.90
changeAUC5/38.058/20.208/63.6510/45.45
MNSBS3/16.104/1.307/62.558/37.90
RFF-MMD2/5.151/72.952/12.951/91.60

Moment Tensor selects the moment order p=2p=2 and recovers the reference count in every recording. It has the smallest distances on 03, 06, and 08, with all estimates on 03 and 06 within the observer intervals. MNSBS is most accurate on 04 (1.30 versus 2.10 s for Moment Tensor).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.