acceptodds
Under review as a conference paper at ICLR 2027

Do We Really Need Complex Models? Revisiting PCA for Time-Series Anomaly Detection

15 views

Abstract

Time-series anomaly detection increasingly relies on complex deep or foundation-model architectures, while practical deployment demands methods that are accurate, efficient, and scalable. PCA offers an efficient and interpretable model of normal variation, but its standard formulation operates on static vector representations without explicitly accounting for temporal structure. We introduce TAP, a Time-series Adaptive PCA framework that addresses this limitation through three novel designs: Multi-Resolution PCA captures anomalies across different durations and temporal resolutions; Cross-Resolution Evidence Consensus identifies anomaly evidence supported by related representations; and Shape-Residual PCA detects deviations from recurrent normal patterns. We further develop efficient covariance compression and batched alignment to support scalable computation. Extensive experiments across 23 public dataset collections demonstrate state-of-the-art overall detection performance, improving upon the second-best method by 17.9% while running 10 times faster. On a single CPU core, TAP requires only 2.5 seconds per benchmark series on average and processes approximately 28,000 points per second, making it practical for accurate, large-scale monitoring and diagnostic workloads. The code is available at https://anonymous.4open.science/r/TAP-TSAD.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Anomaly Detection, Principal Component Analysis, Unsupervised Learning, Scalable Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_c7AHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Motivated by mixed performance of complex time‑series foundation models, this paper introduces TAP, a two‑branch adaptive PCA framework for unsupervised time‑series anomaly detection. It uses nested Toeplitz covariance compression for multi‑resolution modeling and cross‑resolution evidence consensus to filter spurious alerts. Its shape‑residual PCA branch improves detection of waveform‑shifting anomalies by phase‑aligning and removing recurrent normal patterns. Branch outputs are fused with normalized LogSumExp. Evaluated on 23 dataset collections against classical, deep and foundation‑model baselines, TAP achieves favorable accuracy‑speed Pareto‑optimality under CPU‑only execution. Limitations include its periodic‑pattern assumption and simple channel‑wise multivariate aggregation.

Critical strengths
  1. This work presents a well‑motivated investigation questioning the necessity of over‑complex deep and foundation models for time‑series anomaly detection, and revisits classical PCA by addressing its intrinsic temporal modeling limitations. The proposed TAP framework achieves strong Pareto‑optimal performance, delivering SOTA‑level detection accuracy with substantially higher inference speed on mainstream TSAD benchmarks.
  2. The technical contributions are non‑trivial. The designs including nested Toeplitz covariance compression, cross‑resolution evidence consensus via max‑bottleneck path, and shape‑residual PCA effectively mitigate two known flaws of vanilla subsequence PCA: sensitivity to fixed window sizes and insensitivity to waveform‑shaped anomalies. Two complementary branches capture distinct anomaly patterns.
  3. Empirical evaluations are comprehensive and rigorous. Extensive comparisons against classical, deep‑learning‑based and foundation‑model baselines are conducted across multiple large‑scale benchmarks. The paper includes thorough ablation studies, statistical significance tests, scalability analysis, and open‑sourced implementation, satisfying reproducibility requirements for top conferences.
Critical weaknesses
  1. The innovation is largely empirical and engineering‑driven. While building on existing building blocks (cross‑correlation, max‑bottleneck path, LogSumExp fusion), the paper lacks formal theoretical guarantees such as error bounds or convergence analysis for the proposed modules.
  2. Key assumptions and hyper‑parameters bring practical constraints. The shape‑residual PCA relies heavily on the assumption of dominant recurrent normal patterns; its single‑prototype mode may degrade under strongly non‑stationary time series. Several hyper‑parameters (e.g., CREC path length, preset window scales) require careful configuration without fully automatic adaptation.
  3. The multivariate extension is simplistic. It processes each channel independently and aggregates scores naively, without explicitly modeling cross‑channel dependencies. Therefore, performance gains for multivariate inputs are limited, leaving multi‑variable anomaly modeling as an unresolved gap.
Questions for authors
  1. Shape‑residual PCA assumes a dominant recurrent normal waveform. How does TAP perform on highly non‑stationary time series without obvious periodic components? Beyond the single case in the appendix, have you systematically measured the performance drop when this core assumption is violated?
  2. Many building blocks in TAP are well‑known techniques (cross‑correlation, max‑bottleneck path, LogSumExp fusion). Could you clearly articulate the unique novelty of nested Toeplitz covariance compression and the CREC consensus mechanism, and separate your original contributions from existing prior work?
2 comments

We sincerely thank the reviewer for recognizing TAP's motivation, technical design, accuracy--speed trade-off, and comprehensive evaluation. We clarify the contributions and practical limitations below.

W1 / Q2: contributions and theoretical guarantees.

We thank the reviewer for the opportunity to clarify our contributions. As stated in the Introduction and Sections 3.1--3.2, we do not claim cross-correlation, maximum-bottleneck paths, or LogSumExp as individually new operators. The original contributions are:

  • Nested window--scale PCA. Toeplitz covariance is established, but TAP derives all window lengths and within-window resolutions from one lag covariance through nested submatrices and exact covariance propagation, forming a coupled family for unknown anomaly durations and granularities.
  • CREC. The graph algorithm is established, but CREC introduces a consensus objective asking whether timestamp evidence persists across neighboring window/scale transformations. Direct aggregation gives 0.617 VUS-PR; CREC reaches 0.645.
  • Shape-residual PCA. Cross-correlation is the alignment operator; the contribution is removing aligned normal morphology before PCA so that the score describes abnormal waveform residuals rather than phase variation.

Nested construction establishes the relationships used by CREC, while S-R addresses a complementary failure mode. Table 3 shows that an ablated branch can degrade the stronger branch, whereas the complete representations reach 0.659; LogSumExp alone does not explain the gain.

TAP is a deterministic unsupervised scorer rather than a learned iterative optimizer, so training convergence is not the relevant theoretical object. Nested propagation satisfies the exact identity Cov⁡(AW)=ACov⁡(W)A⊤\operatorname{Cov}(AW)=A\operatorname{Cov}(W)A^\top, and CREC exactly solves its finite maximum-bottleneck objective. Revised Appendix B.1 analyzes the inverse-variance score using the exact finite AR(1) eigensystem, quantifies its behavior at w=8w=8, validates it on real series, and characterizes the white-noise boundary. A distribution-free detection-error bound would require normal/anomaly distribution assumptions absent from this benchmark setting; PCA anomaly scoring itself is established (Aggarwal, 2016). These analyses match TAP's modules and claims.

W2 / Q1: Non-stationary and aperiodic series.

We thank the reviewer for raising this question. As stated in Section 3.2, the recurrent-pattern assumption applies only to S-R; M-R requires neither periodicity nor a recurrent prototype. Appendix A.3 and Figure 13 now analyze multiple normal regimes, where the multiple-shape extension suppresses responses to normal regime changes while preserving the anomaly.

Appendix C.8 further partitions all 350 TSB-AD-U series by stationarity and periodicity without using anomaly labels:

Series group#Sub-PCAM-RS-RTAP
Stationary + periodic1170.4480.5740.5810.619
Stationary + non-periodic270.5150.6230.6120.640
Non-stationary + periodic1440.5970.7140.6160.709
Non-stationary + non-periodic620.5970.6280.6030.627

S-R contributes most under stable periodic structure, while M-R becomes stronger under non-stationarity. At the strict non-stationary/non-periodic boundary, S-R remains near Sub-PCA (0.603 vs. 0.597), and TAP matches M-R (0.627 vs. 0.628). Figure 14 shows that S-R is within 0.10 of Sub-PCA on 82.0% of all series, with 53 substantial gains and only 10 losses. The strictly aperiodic experiments in Table 12 and Figure 15 agree: S-R approaches Sub-PCA-like behavior, M-R preserves robustness, and fusion achieves the best mean VUS-PR (0.613).

Controlled aperiodic experiments with drifting levels and variance and no periodic component further confirm this result: across spikes, mean shifts, and variance bursts, S-R remains above Sub-PCA (0.574 vs. 0.526), M-R reaches 0.589, and TAP performs best at 0.613 (Appendix C.8, Table 12).

The same preset window family and maximum scale level are used for all 350 series without dataset-specific tuning. Appendix C.6 reports path-length sensitivity over ℓ∈{2,3,4}\ell\in\{2,3,4\}; full-model VUS-PR remains approximately 0.659 throughout, showing little sensitivity to this choice.

W3: Multivariate extension.

We thank the reviewer for raising this point. As stated in Appendix A.2, TAP is primarily designed for univariate TSAD, and its direct channel-wise extension does not explicitly model cross-channel dependencies. Nevertheless, Appendix C.2 shows that this extension ranks first on five of six TSB-AD-M metrics, reaching 0.45 VUS-PR, versus 0.43 for the strongest baseline and 0.31 for PCA. Thus, the temporal representations transfer effectively even without a dedicated multivariate module. Appendix C.10 already identifies joint modeling, for example through lagged cross-channel covariance, as future work.

Dear Reviewer,

As the rebuttal period draws to a close, we would like to ask whether any questions or concerns remain unresolved. We would be happy to discuss them further, provide clarification, or run additional experiments where helpful, given the computational efficiency of our method.

Reviewer_ccQ9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This is a method for maximizing the performance of Sub-PCA, which is known to perform well on the TSB-AD-U benchmark for time-series anomaly detection. When windows are extracted with stride 1 for each time point t, the windows corresponding to time t are generated with various window sizes (8–512), and compression that averages two adjacent values within each window is performed for up to four levels. Here, the covariance is taken into account. Using the Toeplitz approximation, under which the covariance between two positions within a window depends only on their positional difference (lag), the covariances for all (w, m) are efficiently computed from a single autocovariance. Accordingly, a score is computed for each (w, m). After these scores are standardized, the main path jointly supported by adjacent resolutions is extracted with CREC, and the M-R score is computed by combining the path score with the remaining scores. Separately, the S-R approach computes a score by applying PCA to the residuals obtained after aligning and removing the recurring normal waveform (prototype). The two scores are fused to compute the final anomaly score. In addition, the paper analyzes the characteristics of Sub-PCA from a high-/low-frequency perspective (sensitive to spike-type anomalies but insensitive to waveform changes) and uses this as the rationale for introducing S-R. In experiments, it achieves SOTA performance on the TSB-AD-U benchmark. An ablation study and runtime results are also presented, and the method does not take long to run.

Critical strengths
  • It has a practical motivation as a method for maximizing the performance of Sub-PCA, which is known to perform well on the TSB-AD-U benchmark for time-series anomaly detection.

  • It shows strong performance on a well-known benchmark, TSB-AD-U, under a reliable metric, VUS-PR.

  • It is a very efficient method because it does not require a long training time.

  • It appropriately points out that treating time-series windows as fixed-length vectors can cause anomalies to be fragmented or diluted depending on the window size.

  • It adds a high- and low-frequency-based interpretation of the method (Appendix B), providing a time-series-oriented analysis.

Critical weaknesses
  • The method is somewhat complex. Although the paper emphasizes the strength of simple models, it presents a very complex model built on PCA. I am doubtful whether this method falls within the category of simple models that the submission argues for.

  • There is no literature review of, or comparison with, existing PCA-based methods for time series. Although this work does not propose a new form of PCA itself, it is a time-series anomaly detection method that uses PCA, so I think a literature review is necessary. For example, applying PCA to stride-1 windows and estimating the lag covariance as a Toeplitz matrix is the basic construction of SSA [1,2], which applies PCA to the lag-correlation structure of a time series. The Fourier interpretation in Appendix B is also a known property in SSA. In addition, there are Dynamic PCA [3], which uses time-lagged variables, and wavelet-based Multiscale PCA [4].

[1] Vautard, R., & Ghil, M. (1989). Singular spectrum analysis in nonlinear dynamics, with applications to paleoclimatic time series. Physica D, 35, 395–424. ucla

[2] Golyandina, N., Nekrutkin, V., & Zhigljavsky, A. (2001). Analysis of Time Series Structure: SSA and Related Techniques. Chapman & Hall/CRC.

[3] Ku, W., Storer, R. H., & Georgakis, C. (1995). Disturbance detection and isolation by dynamic principal component analysis. Chemometrics and Intelligent Laboratory Systems, 30(1), 179–196.

[4] Bakshi, B. R. (1998). Multiscale PCA with application to multivariate statistical process monitoring. AIChE Journal, 44(7), 1596–1610.

Questions for authors
  • Could you add more synthetic experiments related to high and low frequencies?

  • This work interprets the proposed method from a Fourier perspective, presenting it as a PCA method that exploits the characteristics of time series. It would be good to also conduct experimental and methodological comparisons with existing PCA-based anomaly detection methods that exploit the characteristics of time series (e.g., SSA-based detection, Dynamic PCA, Multiscale PCA).

5 comments

We sincerely thank the reviewer for the careful assessment and for recognizing TAP's practical motivation, efficiency, strong benchmark performance, and time-series-oriented interpretation of PCA. The revised manuscript addresses the concerns below.

W1: Model complexity.

We thank the reviewer for raising this point. In the title, “complex models” refers to highly parameterized learned architectures, such as deep neural networks and foundation models, rather than to the sophistication of a methodological design. TAP explores how far a classical PCA backbone can be extended when explicitly redesigned for TSAD. M-R constructs a coupled window--scale covariance family and extracts cross-resolution consensus, while S-R exposes waveform deviations through aligned prototype residualization. These operations introduce neither a deep architecture nor gradient-based training, and TAP processes approximately 28,000 points per second on one CPU core. The revised Introduction now explicitly defines this use of “complexity” and describes TAP as a computationally lightweight PCA-based framework.

Q1: Synthetic high- and low-frequency experiments.

We thank the reviewer for this suggestion. Appendix B.2 now reports controlled experiments using backgrounds ranging from white noise to smooth periodic signals, with energy-matched low-frequency waveform perturbations and abrupt spikes. Results are averaged over 10 seeds using VUS-PR.

Anomaly typeSub-PCAFull M-RS-RTAP
Low-frequency shape0.3930.8160.9660.962
Abrupt spike0.8420.9150.8380.911

Sub-PCA is substantially stronger on spikes than on smooth shape deviations, consistent with the frequency interpretation. S-R closes the low-frequency gap, while M-R is strongest on spikes; TAP remains strong on both. Appendix B.1 additionally provides finite-window AR(1) analysis, a white-noise boundary analysis, and real-data validation of the underlying frequency ordering.

W2 / Q2: Existing PCA-based methods.

We thank the reviewer for these references. Revised Section 2.1 now discusses SSA, Dynamic PCA, Multiscale PCA, Joint Sparse PCA, adaptive kernel eigenspace methods, and more recent temporal robust and multiscale PCA work, while clarifying their relationships to and differences from TAP.

Following the methods' original settings, we evaluate SSA on univariate TSB-AD-U and Dynamic PCA and Multiscale PCA on multivariate TSB-AD-M. SSA reconstructs a trajectory matrix under one embedding length; its reconstruction-residual detector achieves 0.39 VUS-PR, compared with 0.54 for Sub-PCA and 0.66 for TAP in Table 1. Dynamic PCA augments process variables with their lags, while Multiscale PCA performs scale-wise cross-variable PCA after wavelet decomposition. On TSB-AD-M, both are fitted on the predefined training prefix and scored on the full series using continuous QQ statistics:

MethodTSB-AD-M VUS-PR
PCA0.31
Dynamic PCA0.32
Multiscale PCA0.36
TAP0.45

These methods emphasize single-length reconstruction, multivariate process monitoring, residual-subspace statistics, low-rank--sparse decomposition, or online adaptation. TAP instead uses inverse-variance-weighted projections over all principal directions, constructs concurrent window-length and within-window resolutions, aggregates neighboring timestamp evidence through CREC, and complements them with shape residualization. The complete results and methodological distinctions now appear in Table 1, Section 2.1, and Appendix C.2.

Thank you for the detailed response and the revised manuscript. I appreciate the added comparisons with time-series PCA methods (SSA, Dynamic PCA, and Multiscale PCA), the controlled synthetic experiments on high- and low-frequency anomalies, and the explicit definition of "complex models" in the Introduction.

However, the limited methodological distinction from SSA remains a novelty concern. Estimating the window covariance as a Toeplitz matrix from a single autocovariance sequence is the lag-covariance estimator used in SSA (Vautard & Ghil, 1989), and the frequency-like behavior of the resulting eigenvectors is a known property of SSA. I will maintain my current positive rating.

We thank the reviewer for the clarification and for maintaining the positive rating. We agree that Toeplitz lag-covariance estimation and its frequency-like interpretation are established and do not claim either as novel.

As stated in the Introduction and Section 3.1, M-R instead derives a coupled window--scale family from one lag covariance through nested submatrices and exact covariance propagation, enabling efficient anomaly detection across different durations and granularities. CREC then aggregates evidence across related resolutions. The frequency-like behavior analysis only explains PCA's behavior and motivates the complementary S-R branch.

We sincerely appreciate the reviewer’s broader suggestions regarding the positioning of model complexity, PCA-related literature, and controlled frequency experiments, which have substantially improved the revised manuscript.

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

This paper proposes Time-series Adaptive PCA (TAP) for time-series anomaly detection. TAP comprises two branches: (1) Multi-Resolution PCA (M-R PCA) constructs a window-scale covariance family from a shared lag-covariance to capture multi-scale deviations, and introduces Cross-Resolution Evidence Consensus (CREC) to aggregate joint evidence; (2) Shape-Residual PCA (S-R PCA) aligns windows to a recurrent normal prototype and scores PCA residuals to capture waveform deviations. The two branches are fused via normalized LogSumExp. Evaluated primarily on 23 univariate TSB-AD collections, TAP achieves a state-of-the-art VUS-PR of 0.66 while running efficiently at ~28,000 points per second on a single CPU core.

Critical strengths

The paper presents an elegant and highly efficient methodology that challenges the necessity of complex deep models by adapting classical PCA. Its two-branch design (M-R PCA via Toeplitz compression, S-R PCA via FFT alignment) effectively captures multi-scale and shape anomalies while achieving exceptional efficiency (~28,000 points/sec on a single CPU core), offering a practical lightweight alternative to foundation models. Furthermore, the empirical evaluation is highly rigorous, validated across 23 dataset collections using range-aware metrics (VUS-PR/ROC) and substantiated by strict statistical significance tests (Friedman/Nemenyi).

Critical weaknesses
  1. Lack of fine-grained evidence for multi-scale effectiveness. The paper motivates TAP with the premise that existing PCA methods overlook multi-scale temporal structures, yet the evaluation rests entirely on dataset-level aggregate metrics (VUS-PR). To substantiate that the performance gains stem from accurate detection at actual anomalies rather than aggregate statistical artifacts, the authors should provide timestamp-level evidence (e.g., score separation margins, anomaly neighborhood curves) comparing TAP against vanilla Sub-PCA at labeled anomalies.

  2. Incomplete related work. Section 2.3 fails to discuss prior PCA variants designed to address fixed-window limitations. For instance, methods such as Joint Sparse PCA (Jiang et al., TKDE 2013) and Adaptive Kernel PCA (O'Reilly et al., TKDE 2015) are representative examples of adapting PCA for temporal or dynamic structures. The authors should incorporate a broader discussion of such time-series-aware PCA variants and explicitly contrast TAP's "fixed multi-scale window family + CREC" with them, particularly regarding their distinct approaches to scale construction and evidence aggregation.

[1] Jiang R, Fei H, Huan J. A family of joint sparse PCA algorithms for anomaly localization in network data streams[J]. IEEE Transactions on Knowledge and Data Engineering, 2012, 25(11): 2421-2433.

[2] Jiang R, Fei H, Huan J. A family of joint sparse PCA algorithms for anomaly localization in network data streams[J]. IEEE Transactions on Knowledge and Data Engineering, 2012, 25(11): 2421-2433.

  1. Unverified theoretical assumptions. Appendix B.1 relies on the asymptotic equivalence between Toeplitz and circulant matrices to explain the "frequency-like" PCA behavior. However, the eigenvectors of finite-window Toeplitz matrices (e.g., w=8w=8) are not strictly Fourier modes. The authors must quantify this deviation (e.g., via principal angles) and validate it on real data. If the deviation is substantial, the spike preference observed in Figure 4 may merely be a synthetic artifact. A failure case (e.g., extreme short windows) is required to clearly delineate the theoretical boundary.

  2. Missing cross-ablation. Table 2 only presents independent internal ablations for the M-R and S-R branches. A cross-ablation—simultaneously disabling the core mechanisms of both branches—is necessary to demonstrate their synergy. A substantial performance drop compared to the full model would more directly substantiate the necessity of fusing both representations.

Questions for authors
  1. Timestamp-level evidence for multi-scale effectiveness: Since the current evaluation relies on aggregate metrics, could the authors provide timestamp-level evidence (e.g., score separation margins or anomaly neighborhood curves) demonstrating that TAP yields significantly higher score separation at labeled anomalies compared to vanilla Sub-PCA?

  2. Discussion of related PCA variants: Could the authors clarify how TAP's "fixed multi-scale window family + CREC" fundamentally differs in scale construction and evidence aggregation from prior time-series-aware PCA variants, such as Joint Sparse PCA and Adaptive Kernel PCA?

  3. Finite-sample validity of the "frequency-like" assumption: For extremely short windows (e.g., w=8w=8), could the authors quantify the deviation (e.g., via principal angles) between the Toeplitz eigenvectors and the Fourier modes? If this deviation is significant, could the authors provide at least one failure case to clarify the boundary of the M-R mechanism (e.g., where M-R fails to detect a spike anomaly)?

  4. Cross-ablation experiment: Could the authors provide a cross-ablation experiment (simultaneously disabling the core mechanisms of both M-R and S-R branches) to explicitly demonstrate the synergy between the two branches? Does the performance drop more significantly compared to the independent ablations in Table 2?

2 comments

We sincerely thank the reviewer for the careful and constructive assessment, particularly for recognizing TAP's time-series-specific redesign of PCA, empirical rigor, and computational efficiency. The revised manuscript addresses the concerns as follows.

W1 / Q1: Timestamp-level evidence.

We thank the reviewer for this suggestion. We evaluate eight cases spanning seven domains, including shape deviations, spikes, and contextual anomalies. After min-max score normalization, the local separation margin is the mean score within labeled intervals minus that at nearby normal timestamps.

SeriesDomainSub-PCA marginTAP marginSub-PCA VUS-PRTAP VUS-PR
332_UCRFacility-0.1440.5560.0020.997
797_YAHOOSynthetic0.0870.4510.0211.000
364_UCRSensor-0.0550.3280.0110.723
544_SMAPSensor0.0630.3190.0650.974
238_SVDBMedical0.1240.2800.0470.794
139_WSDWeb service0.1380.1830.3480.788
421_UCRHuman activity0.1210.3130.4800.910
467_UCREnvironment0.2500.3600.0310.705

TAP improves both measures in all cases. Across all 350 series, VUS-PR gains above 0.20 occur in 20.9% (73/350), versus losses above 0.20 in 1.4% (5/350). Appendix C.9 now defines the margin precisely and provides the raw signals, labels, and complete score curves in Table 13 and Figure 16.

W2 / Q2: Related PCA variants.

We thank the reviewer for highlighting these directions. Section 2.3 was intended to motivate fixed-window limitations in subsequence TSAD through variable-length methods such as GDFlex and MADRID. The revised Section 2.1 now separately discusses PCA-based temporal methods, including the suggested references and more recent temporal robust and multiscale PCA work.

Joint Sparse PCA (Jiang et al., 2013) imposes shared sparsity on abnormal principal subspaces to localize which source/channel causes an event in multivariate network streams; its KLE extension uses temporal correlation to stabilize source localization. TAP instead localizes when anomalous subsequences occur in a univariate series. O'Reilly et al. (2015) address online non-stationarity by splitting/merging kernel eigenspaces and adapting the recent KPCA training window. That window controls streaming memory and model updates, rather than providing concurrent evidence from multiple subsequence resolutions. TAP jointly models multiple window lengths and within-window scales, with CREC aggregating timestamp evidence across neighboring resolutions. These methodological and application-level distinctions are now explicit in Section 2.1.

W3 / Q3: Finite-window validity.

We thank the reviewer for raising this point. Appendix B.1 already stated that finite Toeplitz matrices are not exactly diagonalized by the DFT; the revised Appendix B.1 now quantifies the finite-window behavior. For finite AR(1) covariance Rij=ρ∣i−j∣R_{ij}=\rho^{|i-j|}, the exact eigenvectors are phase-shifted sinusoidal components (Sherman, 2023), with

λk=1−ρ21−2ρcos⁡(ωk)+ρ2.\lambda_k=\frac{1-\rho^2}{1-2\rho\cos(\omega_k)+\rho^2}.

For positive autocorrelation, λk\lambda_k decreases with frequency. At w=8w=8, the maximum angle from corresponding discrete-sine modes is only 0.21∘0.21^\circ at ρ=0.01\rho=0.01 and 0.64∘0.64^\circ at ρ=0.03\rho=0.03. Across 30 finite-sample realizations, the mean Spearman correlation between PC rank and spectral centroid increases from 0.002 for white noise to 0.797 and 0.929, respectively. We further validate this on TSB-AD-U: among 330 series with positive lag-1 autocorrelation, the median correlation is 0.952 and 97.0% exhibit positive frequency ordering. At the white-noise boundary, the covariance is isotropic and the score reduces to standardized window energy; a sufficiently large spike remains detectable, while frequency-selective amplification disappears. Appendix B.1 and Figures 5--6 contain the complete analysis.

W4 / Q4: Cross-ablation.

We thank the reviewer for this suggestion. Table 3 now reports a 2×22\times2 cross-ablation with identical normalization and LSE fusion. Headings give standalone VUS-PR; cell parentheses give the drop from full TAP.

M-R constructionS-R w/o Res. (0.527)Full S-R (0.602)
Vanilla (0.521)0.529 (-0.130)0.580 (-0.079)
Full M-R (0.645)0.605 (-0.054)0.659

The double ablation causes the largest drop. Fusion alone is insufficient: Full M-R + S-R w/o residualization (0.605) underperforms Full M-R (0.645), while Vanilla + Full S-R (0.580) underperforms Full S-R (0.602). Only the two complete representations reach 0.659, demonstrating complementary evidence rather than a gain caused by LSE alone.

Dear Reviewer,

As the rebuttal period draws to a close, we would like to ask whether any questions or concerns remain unresolved. We would be happy to discuss them further, provide clarification, or run additional experiments where helpful, given the computational efficiency of our method.

Comments on the paper

1 comment

We sincerely thank all reviewers for their careful assessment and constructive feedback. We have incorporated the additional analyses and clarifications from the discussion into the revised manuscript. For ease of reference, we summarize the main revisions below.

1. Clarified the framing of model complexity and expanded the PCA-related literature.
The Introduction now clarifies that “complex models” refers to highly parameterized deep-learning architectures and time-series foundation models that often face scalability challenges, while TAP remains computationally lightweight despite its structured methodological design. Section 2.1 now discusses SSA, Dynamic PCA, Multiscale PCA, robust PCA for seasonal time series, Joint Sparse PCA, and adaptive kernel PCA, and explains how their objectives and representations differ from TAP’s multi-window, multi-scale, and timestamp-consensus formulation.

2. Added comparisons with existing time-series PCA methods.
Table 1 now includes SSA on TSB-AD-U, where it obtains 0.39 VUS-PR, compared with 0.54 for Sub-PCA and 0.66 for TAP. Appendix C.2 additionally reports Dynamic PCA and Multiscale PCA on TSB-AD-M, where they obtain 0.32 and 0.36 VUS-PR, respectively, compared with 0.45 for TAP.

3. Strengthened the finite-window and frequency-based analysis.
Appendix B.1 now analyzes the exact finite AR(1) eigensystem at the smallest window size w=8w=8, quantifies its deviation from discrete-sine modes through principal angles, and validates frequency ordering through controlled simulations and real TSB-AD-U series. It also characterizes the near-white-noise boundary, where frequency ordering disappears. Appendix B.2 adds controlled, energy-matched experiments on low-frequency waveform deviations and abrupt spikes, demonstrating the complementary behavior of M-R and S-R.

4. Added systematic evaluation under non-stationarity and aperiodicity.
Appendix C.8 partitions all 350 TSB-AD-U series by stationarity and periodicity without using anomaly labels. It further compares S-R with Sub-PCA across individual series and introduces strictly aperiodic, non-stationary synthetic experiments covering stochastic trends, piecewise drift, and varying variance. The results show that S-R contributes most when recurrent morphology is present and generally approaches Sub-PCA-like behavior when it is absent, while M-R maintains robustness without requiring a recurrent prototype.

5. Added cross-ablation and timestamp-level evidence.
Table 3 jointly ablates the M-R construction and S-R residualization. Removing both mechanisms causes the largest performance drop, from 0.659 to 0.529 VUS-PR, while incomplete branch fusion can underperform the stronger standalone branch. Appendix C.9 further reports timestamp-level score curves and local separation margins for eight cases spanning seven domains. TAP improves both the local margin and VUS-PR in all eight cases. Across all 350 series, VUS-PR gains above 0.20 occur in 73 cases, compared with five losses above 0.20.

We again thank the reviewers for their suggestions, which have substantially strengthened the positioning, analysis, empirical validation, and presentation of the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.