- 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 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 and 100 repetitions per configuration.
Dimension and moment order (Section I.1). We study , 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 , variance , and reversed skewness). Moment order and threshold are selected from and .
Mean Hausdorff errors, in observations, for Moment Tensor are:
| Change type | ||||
|---|---|---|---|---|
| Mean | 7.06 | 7.57 | 6.85 | 7.06 |
| Variance | 30.32 | 79.52 | 94.78 | 142.42 |
| Skewness | 18.00 | 17.79 | 18.55 | 23.97 |
| Combined | 12.85 | 13.65 | 20.07 | 28.17 |
Moment Tensor has the lowest observed mean error in 12 of 16 configurations. Mean-change BS performs slightly better for mean changes at , 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, is selected in 78%, 73%, 82%, and 98% of runs at the four dimensions. For mean changes, 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- dimension runtime study is described in Part 2.
Section I.2 also crosses six signal strengths with five latent AR coefficients at , 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 , 33 affected coordinates, and latent AR coefficient 0.4. The Binomial(5,0.12) design has , 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.
| Method | Beta | Binomial |
|---|---|---|
| Moment Tensor | 40.49 (112.02) / 96% | 5.08 (6.73) / 100% |
| RFF-MMD | 43.00 (48.72) / 99% | 23.01 (11.43) / 100% |
| changeAUC | 197.35 (140.70) / 24% | 132.45 (126.77) / 45% |
| Mean-change BS | 486.39 (7.55) / 0% | 85.21 (164.01) / 80% |
| MNSBS | 519.74 (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 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 . 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
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 , a change at , , and 100 repetitions per setting. Only the first five coordinates change. Latent coordinates are independent, unit-variance stationary Gaussian AR(1) processes with coefficient , continuing across the change. Repetitions use fresh innovations; all methods use identical series and original seeds.
For , define the mean-zero, unit-variance transformation
Affected coordinates undergo one of four transitions: mean, ; variance, ; skewness, ; combined, . 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.
| Change | Tensor | RFF-MMD [2] | MMD-SW | uLSIF | RuLSIF | KLIEP | |
|---|---|---|---|---|---|---|---|
| Mean | 5 | 7.06 | 40.96 | 362.92 | 391.44 | 398.52 | 470.76 |
| Mean | 20 | 7.57 | 37.12 | 558.20 | 604.80 | 567.48 | 695.36 |
| Mean | 50 | 6.85 | 80.96 | 632.24 | 689.12 | 674.08 | 693.60 |
| Mean | 100 | 7.06 | 136.64 | 678.32 | 704.00 | 678.96 | 685.28 |
| Variance | 5 | 30.32 | 197.44 | 712.28 | 693.80 | 696.64 | 678.80 |
| Variance | 20 | 79.52 | 709.44 | 718.08 | 720.00 | 720.00 | 714.92 |
| Variance | 50 | 94.78 | 720.32 | 707.84 | 701.88 | 706.52 | 714.32 |
| Variance | 100 | 142.42 | 718.08 | 710.08 | 710.60 | 689.84 | 701.92 |
| Skewness | 5 | 18.00 | 151.04 | 714.84 | 716.48 | 720.00 | 713.92 |
| Skewness | 20 | 17.79 | 712.96 | 704.60 | 698.44 | 698.44 | 708.76 |
| Skewness | 50 | 18.55 | 716.16 | 699.52 | 696.48 | 700.80 | 707.88 |
| Skewness | 100 | 23.97 | 719.68 | 720.00 | 701.04 | 710.80 | 704.32 |
| Combined | 5 | 12.85 | 127.68 | 720.00 | 705.68 | 709.96 | 692.40 |
| Combined | 20 | 13.65 | 579.20 | 716.16 | 712.60 | 712.32 | 698.92 |
| Combined | 50 | 20.07 | 709.60 | 711.92 | 699.20 | 699.20 | 707.60 |
| Combined | 100 | 28.17 | 700.96 | 715.56 | 703.92 | 696.12 | 694.88 |
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 ; 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.
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⊗p, where Yi=(1,Xi⊤)⊤ and I indexes the observations in the interval. We evaluate f(u)=∑i∈Iwi(Yi⊤u)p directly from the observations and seek the largest ∣f(u)∣ over unit vectors.
Computation. For each random unit-vector initialization and each sign ξ∈−1,1, the routine iterates the following two steps with a shift β>0: first compute an unnormalized update vector g from the current direction u, then normalize g 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.