Online Change-Point Detection with Persistent Laplacian Features

Shiheng Nie; Yunguang Yue

Laplacian spectra catch regime shifts topological summaries miss

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-10 03:59 UTC pith:PHUPYOIF

load-bearing objection Genuine gap filled with correct theory; finite-support model is self-satisfying, not a real limitation the 3 major comments →

arxiv 2607.08635 v2 pith:PHUPYOIF submitted 2026-07-09 stat.ME

Online Change-Point Detection with Persistent Laplacian Features

Shiheng Nie , Yunguang Yue This is my paper

classification stat.ME

keywords persistentdetectiononlinebettiboundchallengeschange-pointcontrol

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes PL-CUSUM, an online change-point detector that feeds persistent Laplacian spectral features into a Page-CUSUM recursion. The central claim is that the positive eigenvalues of the persistent combinatorial Laplacian carry geometric information beyond what persistent homology (Betti coordinates) captures: there exist point clouds with identical Betti coordinates but distinct positive Laplacian spectra, so the spectrum can detect changes that homology counts provably cannot. Under a finite-support exponential-tilting model, the oracle CUSUM delay is bounded above by 2 log A / γ² and the local minimax lower bound matches at the same order, so the ridge-whitened separation γ_ρ governs detection speed. A plug-in version estimated from Phase I training data retains finite-horizon false-alarm control, with estimation error entering only through an effective separation γ_ρ − 2e.

Core claim

The paper identifies a concrete mechanism by which persistent Laplacian spectra extend the detection boundary beyond persistent homology. Proposition 1 constructs tree-family point clouds (path tree vs. star tree) that share identical persistent Betti coordinates at all scale pairs but differ in their order-zero positive Laplacian spectra, because the sum of squared vertex degrees distinguishes the trees. Corollary 2 then shows that if all feature support points share the same Betti coordinates, any stopping time based solely on Betti features has infinite post-change delay, while the positive-spectrum coordinates determine the same log A / γ² delay order as the full feature. The oracle CUSU

What carries the argument

The persistent combinatorial Laplacian L^{a,b}_q is a self-adjoint operator on chain spaces of a Vietoris–Rips filtration at scale pair (a, b). Its zero-eigenvalue multiplicity reproduces persistent Betti coordinates, but its positive eigenvalues encode within-scale connectivity that Betti numbers discard. The ridge-whitened separation γ_ρ = ||(Σ + ρI)^{-1/2}(μ₁ − μ₀)||₂ replaces KL information as the quantity governing detection delay in the local asymptotic regime. The Page-CUSUM recursion S_n = max{0, S_{n-1} + ŝ_n − ĉ} accumulates the whitened one-dimensional score, and a moving-block bootstrap calibrates the control limit to a finite-horizon false-alarm budget.

Load-bearing premise

The theoretical delay bounds are derived under a finite-support model where the feature vector takes values in a finite set and the post-change distribution is an exponential tilt of the pre-change distribution. Real point-cloud features from continuous data may not satisfy this finiteness, and the paper notes that extending the bounds requires replacing the assumption with moment-generating-function conditions that are not verified here.

What would settle it

Construct a change-point scenario where the pre-change and post-change point clouds have different Betti coordinates but identical positive Laplacian spectra across all scale pairs and dimensions. If such a scenario exists and the Laplacian-based detector fails while a Betti-based detector succeeds, the claimed superiority of spectral features would be falsified for that class of changes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If the positive Laplacian spectrum strictly dominates Betti coordinates for certain change types, then topological change-point detectors that rely only on persistence diagrams or persistence landscapes have a detectable blind spot that PL-CUSUM fills.
The log A / γ² delay scaling means that detection speed is controlled by a geometric separation measurable from Phase I training data, giving practitioners a concrete pre-deployment estimate of expected delay.
The Phase I/Phase II calibration procedure with block-bootstrap control limits provides a template for bringing other topological or nonlinear feature maps into sequential detection with finite-horizon false-alarm guarantees.
The tree-family construction in Proposition 1 can serve as a benchmark problem for evaluating whether any proposed topological feature captures within-scale connectivity or merely homological counts.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Genuine gap filled with correct theory; finite-support model is self-satisfying, not a real limitation

read the letter

The main thing to know: this paper builds an online change-point detector using persistent Laplacian (PL) spectra inside a Page-CUSUM recursion, and proves matching oracle upper and minimax lower bounds at order log(A)/γ². The gap it fills is real — prior topological change-point work uses persistent homology summaries (Betti numbers, persistence diagrams), and nobody had systematically brought PL spectra into a sequential detection framework with false-alarm and delay guarantees. Proposition 1 is a clean bonus: it constructs tree-based point clouds where all trees share identical Betti coordinates but differ in their PL positive spectra, so the spectral information is doing genuine work that homology counts cannot capture. Code and data are shipped for 38 real tasks, which is above average for this subfield. The proof chain is standard but correctly executed: exponential-tilt Taylor expansions feed into Wald's identity for the upper bound and a block change-of-measure for the lower bound. The stress-test note is right that the finite-support assumption (flagged as the weakest link by the reader) is not actually a restriction — Proposition 5 shows PL features take only finitely many values by construction, so the model is self-satisfying. That said, the reader's other two concerns do land. The gap between oracle theory (2 log A/γ²) and plug-in practice (O(σ² log(N/α)/γ_eff²) with γ_eff = (γ_ρ - 4e)₊) is real but transparently acknowledged — the abstract correctly scopes the delay bound to the oracle and the plug-in result to false-alarm control only. The g-dependence from overlapping windows is handled empirically via block bootstrap but not in the formal ARL bounds, which is a genuine but minor gap. The real-data results (Table 1: only method passing all 38 tasks under a common calibration protocol) are strong, though the shared bootstrap calibration pipeline could subtly favor the proposed method. This is a solid paper for researchers working at the intersection of TDA and sequential monitoring. The theory is correct, the construction is motivated, and the experiments are thorough. It deserves a serious referee who can check the change-of-measure argument in Proposition 4 and the concentration bounds in Proposition 9 carefully.

Referee Report

3 major / 9 minor

Summary. The paper proposes PL-CUSUM, an online change-point detection method that constructs a Page-CUSUM monitoring score from persistent Laplacian (PL) spectral features extracted from sliding-window point clouds. The methodological pipeline converts multivariate time series into delay-embedded point clouds, computes persistent Betti coordinates and positive PL spectra on Vietoris-Rips filtrations, and feeds a ridge-whitened one-dimensional projection into a recursive CUSUM statistic. The theory has two parts: (1) an oracle analysis under a finite-support exponential-tilting local model, where the oracle LLR-CUSUM delay upper bound and a local minimax lower bound both match at order log(A)/gamma^2 (Theorem 1, Propositions 3-4); and (2) a plug-in finite-horizon guarantee showing that estimation error from Phase I training enters only through an effective separation gamma_rho - 2e, preserving false-alarm control at level alpha + delta (Propositions 9-10, Corollary 5). A Phase I/Phase II workflow with moving-block bootstrap calibration is provided. Experiments on 38 public monitoring tasks compare PL-CUSUM against 12 baselines.

Significance. The paper makes a genuine contribution by bringing persistent Laplacian spectral features into a sequential detection framework with explicit false-alarm and delay criteria, an area the authors correctly note is underexplored. A notable strength is that the finite-support assumption underlying the oracle theory is not an external restriction: Proposition 5 shows that PL features take only finitely many values by construction, so the exponential-tilting local model is automatically applicable. The proof chain (Lemma 2 -> Corollary 4 -> Propositions 3-4) is standard and correctly executed, and the transparency about the gap between oracle delay bounds and plug-in false-alarm-only guarantees is appropriate. The empirical claim that PL-CUSUM is the only method among 13 to satisfy both Pre <= 0.05 and SDR >= 0.999 on all 38 tasks, if it holds up, is a strong practical result. Reproducible code and task configurations are provided via a GitHub repository.

major comments (3)

Section 4.4, Proposition 2 and Proposition 10: The plug-in delay bound is O(sigma^2 log(N/alpha) / gamma_eff^2), which differs from the oracle bound 2 log A / gamma^2 by the introduction of the sub-Gaussian variance parameter sigma^2 = R_{m,K,J}^2 / rho. This sigma^2 factor can be large when the feature range R_{m,K,J} is large or when rho is small, and the paper does not discuss whether the plug-in delay bound is tight or merely a proof artifact. Since the practical method is the plug-in version, the reader is left without a clear sense of how much conservatism the sigma^2 factor introduces. Figure 4(b) shows empirical convergence of plug-in delay to oracle delay as N_tr increases, but there is no analytical bridge between the two bounds. The authors should at minimum add a remark discussing this gap and whether the sigma^2 factor can be tightened, or whether it is inherent to the sub-G
Table 1 and Section 5.3: The claim that PL-CUSUM is the only method satisfying Pre <= 0.05 and SDR >= 0.999 on all 38 tasks is the central empirical selling point, but the evaluation protocol raises questions about fairness. All baselines use the same moving-block bootstrap calibration with M_cal = 2000, but methods like BOCPD (Adams and MacKay, 2007) and FOCuS-MD (Romano et al., 2023) have their own internal mechanisms for false-alarm control that may not align with a uniform bootstrap-calibrated control limit. BOCPD controls only 14/38 tasks on Pre, which could reflect a protocol mismatch rather than a fundamental limitation. The paper should clarify whether the calibration protocol was validated as appropriate for each baseline family, or at minimum add a caveat that the comparison reflects performance under a common calibration scheme rather than each method's native operating regime
Section 4.1 and the definition of gamma_rho in Section 3.3: The ridge-whitened separation gamma_rho = ||W_rho (mu_1 - mu_0)||_2 depends on the pooled covariance Sigma = (Sigma_0 + Sigma_1)/2, which in turn depends on the post-change distribution P_1. In the oracle theory, P_1 = P_{theta_n} is the exponentially tilted post-change distribution, but in the plug-in setting, the Phase I training segment must contain labeled post-change data to estimate mu_1 and Sigma_1. This means the method requires pre-change AND post-change training data, which is a strong assumption for online change-point detection. The paper should state this requirement more prominently (it is implied in Section 3.4 but not highlighted in the abstract or introduction), and discuss how the method would be deployed in settings where post-change data is unavailable before monitoring begins

minor comments (9)

Section 3.2: The definition of the persistent combinatorial Laplacian L^{a,b}_q uses the notation delta^{a,b}_{q+1} and (partial^a_q)^*, but the adjoint is taken with respect to 'the standard simplex-basis inner products on the corresponding chain spaces.' It would help the reader to specify whether these are the same inner products as in Memoli et al. (2022) or a variant, since different inner product choices change the spectrum
Figure 2: The horizontal axis spans tree orders from 4 to 9999, but only a few tick labels are shown (4, 16, 64, 1024, 9999). The curve appears to stabilize, but without finer tick marks it is difficult to assess the convergence behavior. Consider using a log-scale axis with more labels
Section 3.3, the correction term: The oracle correction c = (E_0 s_n + E_1 s_n)/2 is chosen as the midpoint, but other choices (e.g., the zero-drift point under P_0) are also standard in CUSUM literature. A brief remark on why the midpoint is preferred here would be helpful
Algorithm 1, line 4: 'Generate M_cal no-change score sequences of length N by Monte Carlo block resampling.' The block length l_cal is an input but the algorithm does not specify how it is chosen. A brief note on the selection rule for l_cal would improve reproducibility
Table 1 footnote: 'FOCuS-MD has 6 task-level run errors, which are not counted as Pre-controlled or SDR-qualified tasks.' This means FOCuS-MD is effectively evaluated on 32 tasks rather than 38, which affects the fairness of the method-level comparison. This should be noted more prominently in the text discussion, not just in a table footnote
Section 5.1: The SDR criterion is defined as P_nu(T > nu - 1, nu <= T <= N), but the threshold SDR >= 0.999 is quite stringent. A brief justification for this specific threshold (rather than, say, 0.95 or 0.99) would help the reader interpret the results
Appendix, Proposition 8: The sub-Gaussian parameter is derived from Hoeffding's lemma using the boundedness of the feature range R_{m,K,J}. This is a loose bound when the feature distribution is concentrated. A remark acknowledging this conservatism would be appropriate
The paper uses 'Page-CUSUM' and 'Page-CUSUM' interchangeably (hyphenated and en-dash). Standardize throughout
Section 2, paragraph on persistent Laplacian theory: The citation 'Memoli et al. (2022)' appears with an accented character (M'emoli). Ensure consistent encoding across the bibliography

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The central theoretical claim (Theorem 1) — that oracle LLR-CUSUM delay is O(log A / γ²) and matches the local minimax lower bound — is derived from standard sequential detection arguments (Wald's identity, Ville's inequality, change-of-measure on a low-alarm-probability block) applied to a finite-support exponential-tilting model. The ridge-whitened separation γ_ρ is a population-level Mahalanobis-type distance defined from P₀ and P_θ, not a fitted parameter renamed as a prediction. The finite-support assumption is not an externally imposed restriction that circularly forces the result: Proposition 5 shows that the PL feature map Φ takes only finitely many values by construction (determined by which distance-interval each pairwise distance falls into), so the finite-support model is automatically satisfied by the feature construction. The Taylor expansions in Lemma 2 and Corollary 4 are standard exponential-family local asymptotics, not definitions that embed the conclusion. The plug-in bounds (Proposition 2, Proposition 10) cleanly separate estimation error e from the effective separation γ_eff = (γ_ρ - 4e)₊, and the paper is transparent that the oracle bound and the plug-in bound have different forms. No self-citation chain is load-bearing for the main theorem. The derivation is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

8 free parameters · 3 axioms · 1 invented entities

The free parameters are standard for sliding-window TDA methods and are either user-specified or calibrated via bootstrap. The finite-support exponential-tilting axiom is the most restrictive: it is ad hoc to the paper's theory and acknowledged as a limitation in Section 6.

free parameters (8)

window length w
User-specified; sensitivity analyzed in Figure 6
stride h_stride
User-specified; affects g-dependence
lag order L
User-specified embedding parameter
scale grid ε_1,...,ε_K
Discretization of filtration scales
ridge parameter ρ
Regularization for whitening; sensitivity analyzed in Figure 6
spectral truncation level J
Number of positive eigenvalues retained
log-shift ε_log
Numerical stability for log-spectrum
control limit η = calibrated via block bootstrap
Selected to satisfy FARN ≤ α

axioms (3)

domain assumption Window point clouds are mutually independent (or g-dependent with g=⌈w/h⌉)
Stated in Section 3.1; used for the CUSUM recursion analysis. Overlap-induced dependence is handled by block-resampling but not formally in the ARL bounds.
domain assumption Feature increments are σ²-sub-Gaussian under both states
Assumed in Proposition 2 for finite-horizon bounds. Proposition 8 verifies this for bounded PL features with the ridge bound, but the connection to Proposition 2 is implicit.
ad hoc to paper Finite-support exponential-tilting model for local alternatives
Section 4.1; the post-change distribution P_θ is an exponential tilt of P_0 on a finite support. Drives all Taylor expansions and the γ² scaling.

invented entities (1)

Ridge-whitened separation γ_ρ independent evidence
purpose: Population-level measure of feature separation between pre/post-change states
Defined from P_0, P_θ, and Σ; not a fitted constant. The delay bound is expressed in terms of this quantity, and the lower bound matches it.

pith-pipeline@v1.1.0-glm · 33413 in / 2578 out tokens · 504078 ms · 2026-07-10T03:59:38.333618+00:00 · methodology

0 comments

read the original abstract

Online change-point detection in high-dimensional nonlinear time series faces two challenges. The underlying distributions are difficult to model, and state changes are difficult to characterize. We propose persistent Laplacian cumulative sum (PL-CUSUM) to address these challenges. PL-CUSUM maps delay-embedded sliding windows to point clouds. It extracts persistent Betti vectors and the positive spectra of persistent Laplacians from the same Vietoris-Rips filtration. A ridge-whitened projection converts these features into a scalar score. Page's CUSUM recursion then accumulates this score over time. The positive spectra capture within-scale connectivity and geometric information that persistent Betti vectors do not record. Under a finite-support local model, we prove that the oracle upper bound on detection delay and a local minimax lower bound have the same order. This common order is the logarithm of the average run length constraint divided by the squared ridge-whitened separation. We also establish finite-horizon false-alarm and expected-delay bounds for plug-in whitened scores under weak dependence within each state. The method has two phases. Phase I estimates the projection parameters, selects the feature configuration, and calibrates the control limit. Phase II updates the resulting CUSUM statistic online. Experiments on simulated and real monitoring data show stable false-alarm control and competitive detection performance.