Pith. sign in

REVIEW 4 major objections 7 minor 14 references

Topology-driven identification of repetitions in multi-variate time series

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Persistent homology can recover recurrence times in noisy multi-variate time series, with stability guarantees for periodic, repetitive, and recurring signals.

desk verdict A useful framework and benchmark for recovering recurrence times, but the 'provably stable' claim outruns the proofs: only persistence diagrams are shown stable, not the thresholded selection. read the letter →

arxiv 2505.10004 v2 pith:SZTTEIBR submitted 2025-05-15 cs.CG eess.SPmath.ATstat.ML

Simon Schindler , Elias Steffen Reich , Saverio Messineo , Simon Hoher , Stefan Huber This is my paper · ORCID
classification cs.CGeess.SPmath.ATstat.ML MSC 55N3162M1068U05
keywords persistenthomologytimeseriesanalysisrecurrenceestimationcycledetectionperiodicitydelayembeddingindustrialautomationinjectionmoldingbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to recover the time points at which a multi-variate time series returns to a previous state, for three increasingly strict senses of repetition: recurring, repetitive, and periodic. It proposes a persistent-homology framework that turns the series into scalar-valued surrogate functions and reads recurrence times from significant local minima of their sublevel-set filtrations. The paper claims all three methods are stable under bounded perturbations of the input and computationally efficient, with $O(n \log n)$ runtime, and it validates them on real industrial data from an injection molding machine. If the framework holds, cycle boundaries can be recovered from unevenly sampled, noisy sensor data where classical Fourier or autocorrelation methods fail.

What carries the argument

The mechanism that carries the argument is the sublevel-set persistence diagram of a surrogate scalar function. For Method 1, $v_x(t)=\|x(t)-x(0)\|_2$ captures how far the trajectory has strayed from its starting point; for Method 2, a delay embedding $U_x(t)=(x(t),x(t+\Delta),\ldots,x(t+(d-1)\Delta))$ with $v_x(t)=\|U_x(t)-U_x(0)\|_p$ separates self-intersecting repetitions; for Method 3, $w_x(t_1,t_2)=\|x(t_1)-x(t_2)\|_2$ averaged over diagonals, $v_x(\Delta)=\operatorname{avg}_t w_x(t,t+\Delta)$, isolates constant period. In all three cases, zero-dimensional persistent homology tracks births and deaths of connected components in the sublevel filtration, and significant local minima\u2014points with birth $<\varepsilon$ and persistence $>\delta$\u2014are mapped back to time to yield recurrence times. The load-bearing identity is the bottleneck stability bound $d_B(D(v_x),D(v_{x'}))\le c\|x-x'\|_\infty$, which transfers input perturbations to diagram perturbations and underpins the 'provably stable' claim.

What would settle it

Run Method 1 on a synthetic recurring series with known cycle boundaries, adding bounded noise whose amplitude is swept from zero up to the assumed $\varepsilon$, and record whether the number of detected recurrence times changes before the perturbation reaches the bottleneck bound.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the problem of estimating recurrence times reduces to a stable question about zero-dimensional persistent homology. Given a time series $x$, each method constructs a nonnegative scalar function $v$ on the time interval\u2014distance to the starting point (Method 1), distance of a delay embedding to the embedded starting point (Method 2), or diagonal averages of the pairwise distance function (Method 3)\u2014and then computes the sublevel-set persistence diagram. The recurrence times $T_i$ are recovered by selecting diagram points with birth below $\varepsilon$ and persistence above $\delta$ and mapping the corresponding local minima back to the time domain. The stability argument shows that the composed map from input series to persistence diagram is Lipschitz, with bottleneck distance bounded by $2\|x-x'\|_\infty$ for Methods 1 and 3 and by $2\sqrt{d}\|x-x'\|_\infty$ for the delay-embedding method. The authors conclude that the stability guarantees plus experiments on their benchmark establish practical tools for periodic, repetitive, and recurring behavior.

Load-bearing premise

The load-bearing premise is that fixed thresholds $\varepsilon$ and $\delta$ can separate true recurrence points from noise for every series in the perturbation class; the stability proofs cover the whole persistence diagram, not the thresholded selection that actually produces the times.

Editorial extensions

If this is right

  • Recurrence boundaries can be estimated for unevenly sampled multi-variate series, a regime where FFT-based and autocorrelation-based period detection is undefined.
  • A single framework covers three levels of cyclic behavior, so practitioners can choose Method 1 for recurring signals, Method 2 for self-intersecting repetitive signals, and Method 3 for periodic signals.
  • The $O(n \log n)$ runtime makes the approach feasible for online monitoring and control tasks in industrial settings.
  • The stability bounds imply that small changes in the input series produce small changes in the persistence diagram, giving a theoretical foundation for reliability under measurement noise.
  • The published benchmark with annotated cycle boundaries provides a reusable evaluation protocol for future recurrence-estimation methods.

Reading between the lines

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

  • A natural extension is to make the thresholds $\varepsilon$ and $\delta$ adaptive rather than fixed, since the paper's stability proofs concern the full persistence diagram and do not directly guarantee that the selected point set\u2014and hence the recovered times\u2014is stable.
  • The evaluation reports errors only for sections where the correct number of cycles was found, so the published MAE/MARE numbers likely understate failures; a metric that counts missed or extra cycles would better expose where each method breaks.
  • The distance-to-start surrogate of Method 1 could be extended to quasi-periodic or chaotic series by tracking higher-dimensional persistence or by replacing the fixed start point with a sliding reference point.
  • A direct stress test would be to run Method 1 on synthetic recurring series with known cycle boundaries while sweeping the noise amplitude from zero to $\varepsilon$ and recording when the detected cycle count changes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes a persistent homology framework for estimating recurrence times in multi-variate time series, covering three increasingly constrained notions of cyclic behavior: recurring, repetitive, and periodic. Three methods are presented: Method 1 applies a sublevel-set filtration to the distance from the starting point; Method 2 first forms a delay embedding and then applies the same distance-based filtration; Method 3 filters the diagonal average of pairwise distances. The authors claim all three methods are provably stable under perturbations and validate them on a new injection molding benchmark dataset with annotated cycles, reporting MAE and MARE per dataset section. Appendix A derives Lipschitz bounds on the bottleneck distance between the persistence diagrams of the surrogate functions for the three methods.

Significance. If the claims were fully supported, the framework would be a useful contribution: it addresses a real gap in period/repetition estimation for unevenly sampled and multi-variate data, and the public benchmark dataset is a valuable asset for future comparison. The three-level taxonomy (periodic/repetitive/recurring) is a sensible structuring device, and the computational complexity (O(n log n)) is attractive. The stability proofs in Appendix A are correct as far as they go, and the experiments cover a realistic industrial scenario. However, the central 'provably stable' claim overreaches: the proofs bound only the persistence diagrams, not the thresholded selection procedure that actually produces the estimated recurrence times. In addition, the empirical evaluation omits exactly the cases where a method failed to count cycles correctly, so the reported accuracies are conditional on success. The underlying ideas remain defensible, and both issues are fixable in revision, but the current manuscript does not substantiate its headline claims.

major comments (4)
  1. [Section 3.1 and Appendix A, Eq. (2)] The paper's central 'provably stable' claim (abstract and Section 5) is not supported by Appendix A. The proofs in Eqs. (2)-(4) bound only the bottleneck distance between the full persistence diagrams, but the methods do not output a persistence diagram: they select points with birth < epsilon and persistence > delta (Section 3.1) and map those selected points back to time-domain recurrence times. This selection map is not continuous in the bottleneck distance; a point with birth just below epsilon or persistence just above delta can cross either threshold under an arbitrarily small L_infinity perturbation of the time series, so the selected set, its cardinality, and the resulting cycle boundaries can change discontinuously. The manuscript provides no margin or separation assumption that would make the thresholded output stable. Please either add a quantitative separation condition and prove stability of the thresholded selection, or revise the abstract and Section 5 to state that only the persistence diagrams, not the estimated recurrence times, are stable.
  2. [Section 4.3, Table 2, and Figure 3] The evaluation masks failures. The text states, 'For cases where a method failed to identify the correct number of cycles, results are omitted (marked with dashes in the table and excluded from the boxplots).' This makes every reported MAE and MARE conditional on the method having counted the cycles correctly. A method that frequently returns the wrong number of cycles can therefore receive an arbitrarily good conditional error. The conclusion that 'the experimental results validate their practical applicability' is not supported without reporting, at minimum, the number and rate of failures per section, and ideally an unconditional error measure that also penalizes mis-counted cycles.
  3. [Appendix A.3, Eq. (4)] The proof of Eq. (4) as written is invalid. The equality |v_x(Delta) - v_x'(Delta)| = | avg_t avg_t' (w_x(t,t+Delta) - w_x'(t',t'+Delta)) | does not hold; introducing a second independent averaging variable t' changes the quantity. The intended bound can be obtained directly by writing v_x(Delta) - v_x'(Delta) = avg_t (w_x(t,t+Delta) - w_x'(t,t+Delta)) and then applying the previously established sup-norm bound on w_x - w_x'. This is a fixable error, but as it stands the Appendix does not give a correct derivation of Method 3's stability.
  4. [Section 4.2 and Table 2] The threshold parameters (epsilon, delta, d, tau) are tuned on the same benchmark sections for which errors are reported; Table 2 lists one set of selected parameters per method, and the evaluation is in-sample. No train/test split, cross-validation, or parameter-sensitivity analysis is provided. The reported accuracies are therefore estimates of in-sample fit rather than evidence of generalization to new series or to other parameter settings. Please report a parameter sensitivity analysis or split the data into tuning and evaluation portions.
minor comments (7)
  1. [Section 2.1] The runtime complexity statement reads 'O(nα(n)' with a missing closing parenthesis; it should be O(n α(n)), and the reference to the inverse Ackermann function would benefit from a citation to a precise theorem.
  2. [Section 3, introductory paragraph] The sentence 'The strength of this approach lies in it's real-world applicability' contains a typo: 'it's' should be 'its'.
  3. [Figure 3 caption] The caption begins 'Distribtion of the absolute error'; this should be 'Distribution'.
  4. [Section 3.2 and Table 2] The delay parameter is denoted tau in Section 3.2 but tau is also used for cycle length in Section 1.2; Table 2 lists 'tau = 500' for Method 2. Please use a distinct symbol, e.g., Delta, to avoid ambiguity.
  5. [Section 1.2, Definition 5] The statement 'if x is epsilon-delta-approximately recurring then there is a recurring x-hat with ||x - x-hat||_inf < epsilon' is asserted without proof. A short construction or a reference would make the hierarchy of definitions clearer.
  6. [Section 4.2] No quantitative comparison against classical baselines (e.g., FFT or autocorrelation on the evenly sampled sections) is provided. A small baseline comparison would help calibrate the practical gain of the topological methods.
  7. [Section 1.2, Definition 1] Definition 1 defines periodicity by x(t+tau)=x(t) for all t in [0,T-tau], but the paper also allows time series with finitely many discontinuities. These two notions are not compatible in general; please clarify how the periodic definition is interpreted for discontinuous series.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: stability proofs are external and the thresholding gap is an unproven step, not a reduction to inputs.

full rationale

The derivation chain x -> v_x -> D(v_x) -> selected recurrence times is not circular. Appendix A proves only Lipschitz bounds for the diagram-valued maps (inequalities (2)-(4)), using the external Cohen-Steiner stability theorem [13]; those bounds are independent of the paper's own claims. Section 3.1's selection rule (b < epsilon and d-b > delta) is the literal translation of Definition 5, so the method matches the model class by design rather than deriving its output from its input. The abstract's 'provably stable' phrase overreaches, because the thresholding map from diagrams to recurrence times is not shown to be continuous in bottleneck distance; that is an omitted proof / correctness risk, not an equivalence. The experimental omission of failed counts ('For cases where a method failed to identify the correct number of cycles, results are omitted') weakens the empirical claim but is not a fitted parameter renamed as a prediction, since no tuning procedure is described. The only self-citation ([12]) is background on persistent homology and is not load-bearing. No specific equation reduces to another by construction, so no circular step can be exhibited.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The framework requires two hand-picked thresholds per method (epsilon, delta) plus embedding parameters for Method 2; the stability proofs rely on standard persistence stability and on continuity assumptions about the surrogate functions. No new theoretical entities are introduced.

free parameters (8)
  • epsilon (Method 1) = 0.3
    Threshold for the birth value of local minima in the distance-to-start function; hand-picked, no tuning procedure described (Table 2).
  • delta (Method 1) = 0.6
    Minimum persistence for a minimum to count as a recurrence; hand-picked (Table 2).
  • epsilon (Method 2) = 0.4
    Birth threshold for delay-embedded distance function; hand-picked (Table 2).
  • delta (Method 2) = 0.5
    Persistence threshold for delay-embedded distance function; hand-picked (Table 2).
  • d (Method 2) = 4
    Delay embedding dimension; chosen per method, no selection procedure (Table 2).
  • tau (Method 2) = 500
    Time delay for the embedding; chosen per method, no selection procedure (Table 2).
  • epsilon (Method 3) = 0.7
    Birth threshold for the diagonal-average recurrence function; hand-picked (Table 2).
  • delta (Method 3) = 0.1
    Persistence threshold for the diagonal-average recurrence function; hand-picked (Table 2).
assumptions (4)
  • standard math Persistence diagram stability theorem (Cohen-Steiner et al. 2005)
    Used in Appendix A to bound bottleneck distance between diagrams of perturbed surrogate functions.
  • domain assumption The time series can be modeled as a function I -> R^m with a finite number of discontinuities, and the surrogate functions are continuous for the sublevel set filtration
    Stated in Section 1.2 and assumed implicitly in the stability proofs; real sensor data may have more irregular sampling.
  • domain assumption The diagonal average in Method 3 is taken over a continuous domain with the same measure for the original and perturbed series
    Appendix A.3 uses avgt and avgt' to bound |vx(Delta)-vx'(Delta)|; for discrete non-uniform samples this comparison is not well-defined, and Section 3.3 notes 'careful implementation for non-uniform sampling'.
  • ad hoc to paper True recurrence events appear as sufficiently deep local minima of the surrogate function, separable by the chosen epsilon and delta
    This is the core modeling assumption behind all three methods (Section 3), and it is not proven; it is the premise the experimental evaluation tests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topology-driven identification of repetitions in multi-variate time series." pith.science (2026). https://pith.science/paper/SZTTEIBR

@misc{pith2026250510004,
  author       = {Pith},
  title        = {Pith review of: Topology-driven identification of repetitions in multi-variate time series},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SZTTEIBR}},
  note         = {Machine review of arXiv:2505.10004}
}
read the original abstract

Many multi-variate time series obtained in the natural sciences and engineering possess a repetitive behavior, as for instance state-space trajectories of industrial machines in discrete automation. Recovering the times of recurrence from such a multi-variate time series is of a fundamental importance for many monitoring and control tasks. For a periodic time series this is equivalent to determining its period length. In this work we present a persistent homology framework to estimate recurrence times in multi-variate time series with different generalizations of cyclic behavior (periodic, repetitive, and recurring). To this end, we provide three specialized methods within our framework that are provably stable and validate them using real-world data, including a new benchmark dataset from an injection molding machine.

Figures

Figures reproduced from arXiv: 2505.10004 by the authors.

Figure 1
Figure 1. Overview of the function spaces in our approach of first constructing [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. From left to right: recurrent, repetitive, and periodic time series, each [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Distribtion of the absolute error of predicted recurrence times relative to [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

  1. [1]

    Open Challenges in Time Series Anomaly Detection: An Industry Perspective

    Mueller A. Open Challenges in Time Series Anomaly Detection: An Indus- try Perspective. 2025. arXiv:2502.05392 [cs.LG]. url: https://arxiv. org/abs/2502.05392

    work page Pith review arXiv 2025

  2. [2]

    Transforms, Wavelets

    White P. Transforms, Wavelets. In: Encyclopedia of Vibration. Ed. by Braun S. Oxford: Elsevier, 2001:1419–35

    work page 2001

  3. [3]

    SW1PerS: Sliding windows and 1-persistence scoring; discovering periodicity in gene expression time series data

    Perea JA, Deckard A, Haase SB, and Harer J. SW1PerS: Sliding windows and 1-persistence scoring; discovering periodicity in gene expression time series data. BMC bioinformatics 2015;16. Publisher: Springer:1–12

    work page 2015

  4. [4]

    Sliding Windows and Persistence: An Application of Topological Methods to Signal Analysis

    Perea JA and Harer J. Sliding Windows and Persistence: An Application of Topological Methods to Signal Analysis. Foundations of Computational Mathematics 2015;15:799–838. 2 Dataset and code:https://github.com/JRC-ISIA/paper-2025-idsc-topology-d riven-identification-of-repetitions-in-multi-variate-time-series 14 S. Schindler et al

    work page 2015

  5. [5]

    Persistent homology of toroidal sliding window embeddings

    Perea JA. Persistent homology of toroidal sliding window embeddings. In: 2016 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2016:6435–9

    work page 2016

  6. [6]

    Topological time series analysis

    Perea JA. Topological time series analysis. Notices of the American Mathe- maticalSociety2019;66.Publisher:AmericanMathematicalSociety,AMS:686– 94

    work page

  7. [7]

    (Quasi) periodicity quantification in video data, using topology

    Tralie CJ and Perea JA. (Quasi) periodicity quantification in video data, using topology. SIAM Journal on Imaging Sciences 2018;11:1049–77

    work page 2018

  8. [8]

    Time-domain period detection in short- duration videos

    Yang J, Zhang H, and Peng G. Time-domain period detection in short- duration videos. Signal, Image and Video Processing 2016;10:695–702

    work page 2016

Show all 14 references
  1. [9]

    Topological phase estimation method for reparameterized periodic functions

    Bonis T, Chazal F, Michel B, and Reise W. Topological phase estimation method for reparameterized periodic functions. Advances in Computational Mathematics 2024;50:66

    2024

  2. [10]

    Cycling Signatures: Identi- fying Cycling Motions in Time Series using Algebraic Topology

    Bauer U, Hien D, Junge O, and Mischaikow K. Cycling Signatures: Identi- fying Cycling Motions in Time Series using Algebraic Topology. _eprint: 2312.04734. 2024. url: https://arxiv.org/abs/2312.04734

    2024 arXiv

  3. [11]

    Time series analysis using persistent homology of distance matrix

    Ichinomiya T. Time series analysis using persistent homology of distance matrix. Nonlinear Theory and Its Applications, IEICE 2023;14:79–91

    2023

  4. [12]

    Persistent Homology in Data Science

    Huber S. Persistent Homology in Data Science. In: Proc. 3rd Int. Data Science Conference (iDSC’20). Data Science – Analytics and Applications. Dornbirn, Austria (virtual), 2020.doi: 10.1007/978-3-658-32182-6_13

    2020 doi

  5. [13]

    Stability of persistence diagrams

    Cohen-Steiner D, Edelsbrunner H, and Harer J. Stability of persistence diagrams. In:Proceedings of the twenty-first annual symposium on Compu- tational geometry. 2005:263–71

    2005

  6. [14]

    Computational topology: an introduction

    Edelsbrunner H and Harer J. Computational topology: an introduction. American Mathematical Soc., 2010. Topology-driven identification of repetitions in multi-variate time series 15 A Stability Proofs In this section, we provide the stability proofs for the three methods introd...

    2010

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.