acceptodds
Under review as a conference paper at ICLR 2027

Extremal Geometry: Signal-Induced Supports for Time Series Representation

Abstract

Local patterns in time series can occur over different durations, yet many representations use fixed windows, scales, or resolutions. We study extremal geometry, which represents each local maximum or minimum by a signed magnitude and a two-sided temporal support determined by the surrounding signal. Spikelet is a concrete realization developed for variable-duration time-series analysis. We characterize the prior one-pass Spikelet construction and give a linear-time completion that recovers exact signal-induced supports. The completed supports are noncrossing separately for maxima and minima, inducing an inclusion hierarchy across temporal extents. Magnitude pruning preserves the observed value at every retained apex and creates no new strict extrema. For two-sided interior events, the absolute magnitude also equals the corresponding zero-dimensional persistence lifetime, providing a scalar connection to persistent homology. We further compare extremal geometry with fixed-scale Gaussian smoothing and fixed-resolution Fourier reconstruction: smoothing changes global extremal values, whereas reconstruction can create false extrema even with lower error. Controlled experiments audit these structural claims, and a real-data audit shows that magnitude-selected events retain widely varying signal-induced temporal supports.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series representation, variable-duration patterns, extremal geometry, signal-induced temporal support, local extrema, persistent homology

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X2ZTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the theoretical properties of the existing Spikelet representation for time series. It introduces a linear-time algorithm to complete signal-induced support intervals, proves several properties of the resulting structure, and establishes connections with persistent homology. The paper also compares extremal-structure preservation with Gaussian smoothing and Fourier reconstruction.

Critical strengths
  1. The paper is well organized, with clear definitions and detailed mathematical proofs.
  2. The analysis of support completion and the noncrossing property provides a useful characterization of the Spikelet representation.
Critical weaknesses
  1. The paper builds on the existing Spikelet representation. While the support completion algorithm and theoretical analysis are new, it is not clear how much they add to the original method. In particular, the practical benefit of using completed supports is not demonstrated.

  2. Some theoretical results are rather straightforward. For example, Theorem 1 follows from using linear interpolation between retained extrema and does not depend on the specific Spikelet construction. It is difficult to see these results as major contributions.

  3. The comparison with persistent homology could be more informative. The paper explains how the support intervals differ, but it remains unclear what useful information the proposed representation provides beyond existing persistence-based representations.

  4. Most experiments check the theoretical properties, while the real-data analysis is limited to three recordings. A comparison between the original and completed Spikelet supports on a simple task, such as motif matching or event localization, would help show why the new results matter.

Questions for authors
  1. What practical benefit does exact support completion provide over the original Spikelet representation?
  2. What information does the proposed support hierarchy retain that cannot be obtained from existing persistence-based representations?
  3. Can the authors provide some examples where the completed supports improve the results of a downstream task?
3 comments

Thank you for the careful review. Our separate General Response summarizes the overall revisions and their locations. Because the geometric distinctions are easier to see than to describe, we particularly invite the reviewer to examine Figs. 13–15 and Table 10 (Appendices I–K).

W1 / Q1 (What the results add to the original Spikelet). First, the theory characterizes what was already exact in the original method: the one-pass scan computes exact magnitudes and left endpoints (Proposition 2), while Theorem 1 guarantees preservation of retained apex values under magnitude-only pruning. Second, completion recovers each event's full signal-induced extent: the one-pass interval may stop at an early local low or high, whereas completion reaches the deepest trough or highest rebound within the full excursion. Fig. 14 (Appendix J) illustrates both cases on S&P 500 prices: a decline event's right endpoint moves from September 16 to October 12 (18 additional trading sessions), and a rebound event gains 19 sessions. Completion changes right endpoints for 20.8%–25.6% of selected events in three real recordings (Table 8). Contribution 1 now identifies Theorem 2's collective noncrossing and inclusion hierarchy as our main result.

W2 (Theoretical contributions). We agree that Theorem 1 is elementary and uses the Spikelet construction only through the retained apex set; the revision states this after Theorem 1. Its role is to establish a preservation baseline for Section 6. Proposition 6 (Appendix F) shows that a fixed rank-qq linear reconstruction cannot universally preserve q+1q+1 independently varying apex values; in contrast, Theorem 1 guarantees exact values at the apexes retained by adaptive pruning. Our main result is Theorem 2, while Proposition 4 and Corollaries 2–3 (Appendix F) give the most technically involved analysis, establishing conditions under which fixed-resolution reconstruction introduces false extrema.

W3 / Q2 (Information beyond persistent homology). The difference is the geometric object: Biswas et al. associate one window with each persistence pair of a maximum and minimum, whereas Spikelet assigns a two-sided support to each individual maximum or minimum (Fig. 13: two pair windows versus four supports; Table 10, Appendix I). Although absolute Spikelet magnitude equals 0D persistence lifetime for two-sided interior events (Proposition 1), it is only the smaller of the two apex-to-endpoint amplitude drops, aLa_L and aRa_R. Spikelet's event-centered representation additionally provides: (i) each event's temporal support—within lifetime bins, the support-length Q90/Q10 ratio reaches 235.22 (Table 9); (ii) access to both side drops aLa_L and aRa_R from the original samples, capturing asymmetry; and (iii) same-polarity support containment, organized into noncrossing inclusion forests by Theorem 2. These properties motivate the time-series structural queries in W4/Q3.

W4 / Q3 (Downstream impact). As a concrete event-localization illustration, Table 8 and Fig. 14 compare the original one-pass and exact completed supports: 20.8%–25.6% of selected events have provisional right endpoints, which completion corrects according to Definition 1. Beyond localization, we introduce time-series structural queries, retrieving variable-length events without prescribing a window length by combining shape, magnitude, duration, and asymmetry with relationships among internal Spikelets. Fig. 15 (Appendix K; Section 7) retrieves plateau-like, nested, sharp, and asymmetric events from a measured supersonic-jet acoustic record (2 million samples) for physical analysis. Together with the three original recordings and S&P 500 data, these illustrations span five records across physiology, animal behavior, human motion, finance, and acoustics. We do not claim improvements in downstream accuracy without controlled comparisons; benchmarking against existing query methods remains future work (Section 8). The present results establish two concrete capabilities: exact recovery of signal-induced event supports and structural retrieval of variable-length waveform events, including internal substructures, from measured data without prescribing fixed windows.

We would be glad to discuss any point further.

Reviewer_X2ZTOpenReview ↗

Response to Author's Rebuttal

Thank you for the response. The revisions clarify the main theoretical contribution and its distinction from persistence-based methods. The new examples also help illustrate the role of support completion.

My main concern remains whether these results offer meaningful advantages over existing methods. The new examples demonstrate the proposed representation's capabilities, but do not provide comparative evidence of its benefits. I will therefore keep my rating unchanged.

Thank you for reconsidering our revision and for acknowledging that the main theoretical contribution and its distinction from persistence-based methods are now clearer.

We understand the remaining concern. Our examples demonstrate concrete capabilities, but do not establish comparative downstream advantages. Our principal contribution is the mathematical guarantee for the collective extremum-centered support family (Theorem 2), together with a precise characterization of what the original Spikelet scan computes (Proposition 2). Beyond the new examples, the paper provides structural comparisons: positive smoothing changes global extremal values (Proposition 3); fixed-resolution reconstruction can introduce false extrema (Proposition 4; Corollaries 2–3); fixed-rank linear reconstruction cannot universally preserve independently varying apex values (Proposition 6); and certain compactly supported wavelets admit crossing supports (Proposition 8). Table 9 further shows large support-length variation within persistence-lifetime bins.

We agree that controlled downstream comparisons are needed to establish whether and when these properties translate into measurable advantages, and we identify this as future work (Section 8). We therefore position the paper as a contribution to the mathematical foundations of time-series event representation, with downstream benchmarking as the next step.

Reviewer_CwQ8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Building upon the existing Spikelet time-series representation method, this paper proposes the Extremal Geometry framework and investigates signal-adaptive temporal support representations through theoretical analyses of linear-time support interval completion, extremum-preserving properties, and non-crossing hierarchical structures. However, the novelty and practical applicability of the proposed framework remain insufficiently validated.

Critical strengths

Proposed the Extremal Geometry framework, which has certain theoretical analysis capabilities.

Critical weaknesses
  1. The contributions mainly lie in the theoretical characterization and algorithm completion of existing methods, with relatively limited innovation in the methods.
  2. There are no experiments for downstream tasks such as classification, anomaly detection or Motif Discovery, and thus it has not been proven that the proposed support interval completion and hierarchical structure can bring about actual performance improvement.
  3. Mainly compared Fourier and Spline methods with fixed resolution, but lacked a direct quantitative comparison with adaptive wavelets, EMD, and topological persistence methods.
Questions for authors

The main approach was to verify the theoretical properties using artificially synthesized signals. The real data only involved three sets of records, and the experimental scale and data diversity were insufficient to support a broad applicability conclusion.|

The paper lacks integration experiments with deep temporal representation learning models, which makes it unclear what the actual contribution of this theoretical framework is to current machine learning research.

The motivation is not clearly stated. Why is extremal geometry being considered? What are the benefits of considering it? If it is not considered, what problems will arise?

2 comments

Thank you for the careful review. Our separate General Response summarizes the overall revisions and their locations. Because the geometric distinctions are easier to see than to describe, we particularly invite the reviewer to examine Figs. 13–15 and Table 10 (Appendices I–K).

W1 / Q3 (Motivation and novelty). Our motivation is to understand the structural basis of Spikelet's success in variable-length pattern discovery. Rather than prescribing a window length, it identifies events from intrinsic, topology-related features of the signal. A nontrivial connection is that independently developed Spikelet and zero-dimensional persistent homology (PH) yield the same scalar—magnitude and lifetime, respectively—for two-sided interior events (Proposition 1).

Yet their geometric objects differ. Biswas et al. use PH to associate one window with a maximum–minimum persistence pair; Spikelet assigns its own two-sided support to each extremum (Fig. 13; Table 10). This provides directly retrievable variable-length waveform events.

Our main novelty, Theorem 2, proves separate noncrossing inclusion hierarchies for maxima and minima. It underpins time-series structural queries characterizing events by the subevents they contain (Appendix K). Fixed windows may truncate slow events or include unrelated context (Fig. 1); without the collective analysis of extremal supports, their hierarchical relationships would lack a mathematical guarantee.

W2 / Q1 (Practical value and evidence). Because our guarantees hold for all signals in the stated class, synthetic experiments audit their implementation rather than establish the mathematical claims (50/50 eligible signals, Table 3). Real-data evidence now spans five records from physiology, animal behavior, human motion, finance, and acoustics, up to 2 million samples (Section 7; Appendices J–K).

The practical value of Theorem 2 is a guaranteed same-polarity hierarchy for querying relations between events. Combined with shape, magnitude, duration, asymmetry, and internal subevents, it motivates time-series structural queries beyond prior motif discovery and anomaly detection. Fig. 15 (Appendix K) retrieves plateau-like, nested, sharp, and asymmetric events from a measured supersonic-jet acoustic signal (400 kHz), forms physically relevant to intermittency, shock-like steepening, and waveform asymmetry. Retrieved intervals can then undergo physical diagnostics without a fixed window length.

Support completion has a different practical value: retrieving each event's full signal-induced extent. Completion changes the right endpoint of 20.8–25.6% of selected events on three real datasets (Table 8); in S&P 500 declines and rebounds, it adds 18 and 19 trading sessions beyond the one-pass endpoints (Fig. 14, Appendix J).

These examples show retrieval possibilities, not improved average performance; we agree that broader evaluations are needed to establish downstream benefits. Rigorous assessment requires a query system and benchmarks covering diverse constraints, datasets, and applications, which we leave to future work (Section 8).

W3 (Comparison with existing representations). We agree that direct quantitative comparisons with adaptive wavelets, EMD, and persistence-based methods are absent. Since our objective is to establish Spikelet's mathematical foundations, we analyze two basic operations: smoothing and fixed-resolution reconstruction. Positive smoothing contracts global extremal values (Proposition 3), and Fourier and periodic spline reconstruction can create false extrema under specified conditions (Proposition 4; Corollaries 2–3), whereas magnitude pruning preserves every retained apex value exactly (Theorem 1). Thus, waveform accuracy and extremal-structure preservation are distinct.

Two further comparisons are included. For compactly supported orthonormal wavelets with support longer than one translation unit, supports at adjacent levels can cross (Proposition 8), unlike Spikelet's same-polarity support families (Theorem 2). For the PH connection, among events within persistence-lifetime bins, the Spikelet support-length Q90/Q10 ratio reaches 235.22 on cricket data (Table 9), complementing the geometric comparison in Appendix I (Fig. 13; Table 10). Adaptive wavelets and EMD require method-specific analysis; downstream comparisons with them are future work (Section 8).

Q2 (Integration with deep learning). We agree that integration with deep time-series models would clarify downstream benefits. Extremal geometry addresses a complementary question: which temporal interval constitutes an event, rather than how its features should be learned. Spikelet provides signal-induced, variable-length supports and same-polarity hierarchies that could serve as adaptive event tokens or structural features. Integration and downstream evaluation remain future work (Sections 2 and 8).

We would be glad to discuss any point further.

Reviewer_CwQ8OpenReview ↗

Response to Author's Rebuttal

Thank you for the detailed response. My main concern remains the limited quantitative evidence that support completion improves event retrieval over the original one-pass Spikelet or closely related persistence-based methods. I keep my original rating.

Reviewer_HaNtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores a way to analyze time-based data by focusing on natural peaks and dips rather than cutting the timeline into rigid, fixed blocks. Each peak or valley is assigned a prominence score (how much it stands out) and a specific timeframe marking where the event begins and ends. Because this approach builds directly on an existing technique called Spikelet, the foundational premise is not entirely new.

From a mathematical standpoint, the authors show that this prominence score connects directly to persistent homology—a formal way of measuring how much a peak or valley stands out relative to its surroundings. They also prove that these time windows organize into a clean hierarchy: smaller events can sit neatly inside larger ones, but two distinct peaks will never partially overlap or cross boundaries (and the same rule applies separately to valleys).

The authors found that while the original Spikelet algorithm correctly calculates event sizes and start times, it occasionally leaves the end boundaries undefined. To resolve this, they design a fast, efficient algorithm that accurately fills in those missing endpoints.

Another section looks at simplifying noisy data by stripping away minor bumps and dips. By connecting the remaining key peaks and valleys with simple straight lines, the original data values are preserved without introducing artificial new highs or lows. The authors contrast this with conventional smoothing or Fourier reconstruction, which often distort peak heights or introduce fake ripples into the signal.

The paper's experiments mostly serve to confirm these mathematical properties. Using real-world datasets, they also demonstrate that even after filtering out minor fluctuations, the remaining major events can still span wildly different lengths of time.

Overall, the mathematical reasoning appears sound, but it is debatable how much of the work is truly original. Several core ideas—such as linking peaks to persistent homology and organizing them hierarchically—have already been explored in earlier research. The most compelling contributions are the detailed audit of the original Spikelet algorithm and the formal proof that these specific intervals never cross. Beyond that, some findings feel relatively routine, and the authors do not clearly demonstrate whether these theoretical properties offer any practical benefits in real-world applications.

Critical strengths

The primary strength of the paper is its rigorous mathematical dive into an existing time-series technique, formalizing properties that were previously left vague or incomplete. Most notably, the authors provide a reliable, efficient method to determine exact event boundaries and prove that these intervals form a natural, nested hierarchy. This puts the Spikelet method on solid theoretical footing and clarifies exactly which characteristics of the original signal are preserved.

Critical weaknesses

Limited novelty compared to existing literature: While the mathematical analysis is thorough, the core concepts do not clearly distinguish themselves from established tools, and the related work section does not sufficiently highlight the differences:

o Connection to peak prominence: Describing peaks by their magnitude and left/right boundaries strongly mirrors the standard concept of "peak prominence." For signals without identical values, these exact quantities have been calculated by standard libraries like scipy.signal.peak_prominences since 2018. While they handle tied values differently, the authors do not discuss this close parallel.

o Overlap with persistent homology: Biswas et al. (2024) have already shown how peaks and valleys can be tied to time intervals using persistent homology, demonstrating that certain intervals form non-crossing hierarchies. While the specific intervals defined here differ slightly, the underlying logic is very similar.

o Standard algorithmic techniques: Glisse (2023) showed that 1D persistence can be computed in linear time, and the interval-completion algorithm proposed here relies on well-established methods. Beyond auditing Spikelet and proving non-crossing intervals, the theoretical advance feels relatively modest.

Unclear practical value: The authors do not demonstrate why these theoretical corrections matter in practice. In particular, it remains unproven whether fixing the right-hand boundaries leads to better performance than using the original Spikelet intervals or persistence-based alternatives. Because the experiments primarily validate mathematical theorems rather than testing downstream tasks (like pattern recognition or anomaly detection), the practical benefit is hard to evaluate.

Questions for authors
  1. Novelty compared to standard tools (clarification): For signals without tied values, the magnitude and temporal boundaries appear nearly identical to standard peak prominence and base intervals (as implemented in SciPy). Is this understanding accurate? If so, what are the primary differences, and what unique value does your representation offer over existing tools?

  2. Distinction from prior theoretical work (clarification): Biswas et al. (2024) already investigated geometric intervals and hierarchical structures for 1D persistence. Could you clarify where your non-crossing theorem diverges from their findings, and what mathematical guarantees are established here that were not previously known?

  3. Practical impact of completing intervals (clarification and evidence): The paper shows that original Spikelet already finds magnitudes correctly, while your algorithm fixes missing right boundaries. Does having complete boundaries actually improve performance on real-world tasks, such as motif discovery or anomaly detection? If this hasn't been benchmarked, could you describe a concrete scenario where this correction makes a noticeable difference?

  4. Data storage and footprint (clarification): The formal definition of an event tracks its location, magnitude, and boundary points, but the linear reconstruction also seems to require the raw signal values at each peak and valley. Are these raw values stored as part of the representation? If so, how does the overall data footprint compare to other standard time-series formats?

  5. Sensitivity to noise (limitation and stability): Because boundaries hinge on exact local extrema, small perturbations in the signal could significantly alter the resulting intervals. Have you tested the stability of these boundaries in the presence of noise, or is there a theoretical reason to expect stability under specific assumptions?

1 comment

Thank you for the careful review. Our separate General Response summarizes the overall revisions and their locations. Because the geometric distinctions are easier to see than to describe, we particularly invite the reviewer to examine Figs. 13–15 and Table 10 (Appendices I–K).

Q1 / W1, W2, W4 (Peak prominence and novelty). Yes. For signals with distinct sample values, positive Spikelet magnitudes and support endpoints coincide with peak prominence and its bases in scipy.signal.peak_prominences (Section 2; Appendix I). For tied values, Spikelet uses a strict left first passage, a non-strict right first passage, and a fixed rule for equal support minima, while SciPy treats equal-height peaks differently. These conventions define the extremum-centered supports for which Theorem 2 proves noncrossing: supports of maxima, and separately of minima, are nested or disjoint, yielding an inclusion forest for each polarity. The novelty is this guarantee for the collective family, not a new per-peak measure.

The original Spikelet algorithm already uses stack-based reductions (Imamura & Nakamura, 2021); related linear-time reductions also appear in 1D persistence (Glisse, 2023). Our contribution is to characterize precisely what the original Spikelet scan computes: exact magnitudes and left endpoints, but potentially provisional right endpoints, and to recover the exact two-sided supports through an O(N)O(N) completion (Proposition 2). Proposition 4 and Corollaries 2–3 (Appendix F) give our most technically involved analysis: sufficient conditions for false extrema under fixed-resolution reconstruction.

Q2 / W3 (Distinction from Biswas et al.). We agree that persistent homology (PH) also provides nested signal-domain windows. The difference is the geometric object: Biswas et al. associate a window with each persistence pair of a maximum and minimum, whereas Spikelet assigns its own two-sided support to each individual extremum. Fig. 13 shows two pair-indexed windows versus four extremum-indexed supports on the same waveform; Table 10 compares their guarantees (Appendix I). Their nesting results concern pair windows and panels; the noncrossing of the extremum-indexed support family, separately for each polarity (Theorem 2), is the guarantee established here. Although absolute Spikelet magnitude equals 0D PH lifetime for two-sided interior events (Proposition 1), this scalar does not determine event duration or left/right asymmetry: within lifetime bins, Spikelet support-length Q90/Q10 ratios reach 235.22 (Table 9). This event-centered geometry also provides the basis for time-series structural queries over variable-length events, combining shape, duration, asymmetry, and internal Spikelet structure (Fig. 15, Appendix K).

Q3 / W5 (Practical impact of completion). Exact completion fixes provisional right endpoints from the original Spikelet scan, while preserving its already-correct magnitudes and left endpoints (Proposition 2). Its practical meaning is that the original interval may stop at an early local low or high, whereas completion reaches the deepest trough or highest rebound within the full event excursion. Fig. 14 (Appendix J) illustrates both cases on S&P 500 prices: a decline event's right endpoint moves from September 16 to October 12 (18 additional trading sessions), and a rebound event gains 19 sessions. Completion changes right endpoints for 20.8%–25.6% of selected events in three real recordings (Table 8). We have not benchmarked downstream tasks; these are localization differences, not demonstrated accuracy gains.

Q4 (Storage). Yes, apex values are stored. For piecewise-linear approximation, KτK_\tau retained apexes and two signal endpoints require Kτ+2K_\tau+2 position–value pairs; in our real-data audit, magnitude selection retains 1.4%–7.4% of all extrema (Table 8). For event-based analysis, the original NN samples remain available; each record (p,ℓ(p),r(p),m(p))(p,\ell(p),r(p),m(p)) adds O(Kτ)O(K_\tau) storage, giving O(N+Kτ)O(N+K_\tau) total. Spikelet serves as an event index into the original signal, not as a lossless-compression method (Section 4.2).

Q5 (Noise sensitivity). We agree that small perturbations can shift endpoints, particularly near tied minima or first-passage boundaries. Prior Spikelet work has addressed noise sensitivity in applications using magnitude-based reduction, constant-segment extraction, and local support-dependent reduction (Appendix A.3). These task-dependent operations have been used in motif discovery and anomaly analysis; they do not prove stability of the raw supports. Persistence diagrams are bottleneck-stable under sup-norm perturbations (Cohen-Steiner et al., 2007), and absolute Spikelet magnitude equals 0D persistence lifetime for two-sided interior events (Proposition 1). However, diagram stability does not guarantee support-boundary stability. Formal guarantees remain future work (Section 8).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful and constructive reviews. We have uploaded a revised PDF and updated the OpenReview abstract accordingly. The main revisions address three concerns raised by multiple reviewers. Since our main contributions are mathematical, we have also added figures and tables to make the key distinctions easier to understand. Reviewer-specific questions and concerns are addressed in our individual responses.

  1. Distinction from zero-dimensional persistent homology (PH). Appendix I, Fig. 13, and Table 10 (new) compare pair-indexed PH windows with extremum-indexed Spikelet supports on the same waveform.

  2. Contributions beyond the original Spikelet method. Contribution 1 now leads with Theorem 2, establishing polarity-specific noncrossing and inclusion hierarchies for the collective support family. Section 2 clarifies the relationship to peak prominence and linear-time 1D persistence; Appendix I provides further comparisons. Appendix J and Fig. 14 (new) illustrate exact support completion on S&P 500 prices.

  3. Practical use of the theoretical results. Appendix K and Fig. 15 (new) demonstrate time-series structural queries on a measured supersonic-jet acoustic record of 2 million samples. Section 4.2 clarifies event storage and retrieval.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.