acceptodds
Under review as a conference paper at ICLR 2027

Online Change-Point Detection in RKHS: A Non-Asymptotic Performance Analysis

Abstract

We study online change-point detection based on the empirical regularized Pearson divergence in a reproducing kernel Hilbert space (RKHS). The proposed procedure compares a left reference window, representing the nominal regime, with a right test window, containing the most recent observations. Although the resulting sliding-window statistic admits a tractable closed-form expression, the main contribution of this paper is a non-asymptotic analysis of its sequential detection performance. Under a bounded-kernel assumption, we derive finite-sample concentration inequalities under the null and obtain explicit threshold calibrations guaranteeing either average-run length control or uniform false-alarm control over an infinite horizon. Under the alternative, we establish non-asymptotic bounds on the per time miss-detection probability and on the tail of the detection delay. The analysis highlights the roles of partial contamination of the test window, regularization, and the window lengths. Experiments support the theoretical findings.

Primary area:
learning on time series and dynamical systems
Keywords:
Online change-point detection, kernel methods, nonparametric learning, density-ratio estimation, non-asymptotic analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RPxHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper analyzes an RKHS test statistic for online change detection based on the \chi-squared divergence. Given a sequence of observations x_1, x_2, \hdots with an unknown change time \tau, they consider a sliding window approach, where at time t the window x_{t-n,t},...x_{t} is compared with x_{t+1}, \hdots, x_{t+m} using a test statistic that approximates the \chi^2 divergence between the empirical measures defined by the windows. The analysis focuses on deriving explicit bounds on performance criteria for online change detection, including probability of false alarms, average run length to false alarm, and detection delays.

Critical strengths

Kernel based change detection has attracted significant interest in recent years. There are a wide variety of possible statistics and stopping rules that can be derived using RKHS-based approaches including MMD, and the various divergences, and it is of interest to undersand their trade offs. The authors point out that their non-asymptotic performance analyses of \chi^2 divergence for online detection is novel.

Another strength is that the authors consider many different performance criteria in the analysis including several different false alarm measures and multiple delay measures.

Critical weaknesses

Sliding window non-parametric change detection, where a block x[t−n:t]x_{[t-n:t]} is compared with x[t:t+m]x_{[t:t+m]} as t increases, has the limitation that if detection does not occur within the interval t∈[τ−m,τ+n]t \in [\tau-m, \tau+n], then the algorithm will fail (unless it detects due to random fluctuations.) This suggests that the approach in this paper is better suited for offline change detection, where the test statistic can be computed at all possible splits of the data, avoiding this problem.

A key question is: Do the operating regimes specified in the results (Theorem 3.1 through Theorem 4.2) lead to non-vacuous performance guarantees? Theorem 3.1 and Corollary 3.2 appear to have some limitations in terms of their operating regimes that should be explicitly stated. Namely, since the kernel is bounded by K and since the A^\hat{A} operator has $\rho >0 $, we can bound the statistic as χ^ρ2≤2Kρ\hat{\chi}_{\rho}^2 \leq \frac{2K}{\rho}. Therefore, if ν>2Kρ\nu > \frac{2K}{\rho} then Theorem 3.1 is vacuously true because detection stops altogether, which of course makes delay go to infinity. Likewise, in Corollary 3.2, for any α\alpha the νt\nu_t eventually get so large that detection stops altogether and delays go to infinity.

Questions for authors

Do the results imply non-trivial guarantees on false alarms and delays simultaneously (that is, under the same operating conditions for the algorithm in terms of \rho, n,m and \nu?)

Secondly, dependence on the constants in the the theoretical results are rather opaquely presented. Is it possible to provide an intuitive explanation for the right hand sides of (27), (25)? Additionally, what specifically does the sample-size condition implying about the size of the windows?

1 comment

We thank the reviewer for the careful assessment and for highlighting the importance of understanding the operating regimes and scope of the guarantees. These suggestions have led to substantial additions: explicit non-vacuity conditions, a joint-guarantee analysis, clearer interpretations of the miss-bound terms, and an offline extension. We address the four points below.

1. Sliding windows (Weakness 1). With fixed adjacent windows, a change at t0t_0 remains within the monitored data only for split indices t∈[t0−m+1,t0+n−1]t\in[t_0-m+1,t_0+n-1]. Our detection-delay tail bound concerns the initial transition, while the reference window is still nominal. Crucially, the proof uses set inclusion, not independence between successive overlapping tests. The revised sliding-window remark makes this limitation explicit. We also extend the pointwise analysis to offline samples Z1:TZ_{1:T}, testing each split with n=tn=t and m=T−tm=T-t; a union bound controls false alarms over the T−1T-1 splits. We do not claim an offline localization guarantee.

2. Non-vacuity (Weakness 2).

  • ARL thresholds. We agree that validity alone does not ensure a useful detector. The new statistic-range lemma establishes χ^ρ2<(K+ρ)2/(2ρ)\hat\chi^2_\rho<(\sqrt K+\sqrt\rho)^2/(2\rho) and, for positive-valued kernels, χ^ρ2≤K/ρ\hat\chi^2_\rho\le K/\rho. At n=80n=80, m=40m=40, K=1K=1, ρ=10−2\rho=10^{-2}, and target ARL γ=100\gamma=100, the sufficient analytical threshold is 104.3104.3, exceeding the bound 60.560.5: the resulting threshold is vacuous, despite its valid guarantee. The new operating-regimes table shows that, for n=2mn=2m and these K,ρ,γK,\rho,\gamma, non-vacuity first becomes possible at m=67m=67. Our experiments instead use direct nominal sliding-window ARL calibration, without claiming a distribution-free empirical guarantee.
  • Uniform false-alarm control. With fixed windows, our time-dependent analytical threshold eventually exceeds the statistic's range. A new finite-memory lemma shows that uniform infinite-horizon control requires summable per-time false-alarm probabilities along disjoint monitoring blocks, hence increasing conservativeness. The analysis also shows that allowing both window lengths to grow at least on the order of log⁡t\log t, with a sufficiently large constant, can keep the analytical thresholds non-vacuous.

3. Simultaneous guarantees (Q1). The false-alarm and miss bounds apply to the same (n,m,ρ,ν)(n,m,\rho,\nu), but a non-vacuous threshold alone does not ensure an informative miss bound. The new joint-guarantees appendix derives the necessary positive-margin condition Sϵt>2ν+4max⁡c0,d0S_{\epsilon_t}>\sqrt{2\nu}+4\max\\{c_0,d_0\\}. It also makes the limitations of our explicit constants transparent: for the derived miss bound even to be non-trivial, one needs both n>128log⁡8,K2(K+ρ)/ρ3≈266K2(K+ρ)/ρ3n>128\log 8\\,K^2(K+\rho)/\rho^3\approx266K^2(K+\rho)/\rho^3 and n>256log⁡8,K2K+ρ/ρ5/2n>256\log 8\\,K^2\sqrt{K+\rho}/\rho^{5/2}; the first is more restrictive when ρ≤K/3\rho\le K/3. These are necessary conditions for this particular upper bound to be informative, not requirements for the detector to operate or detect changes. The Neumann-series sample-size assumption is likewise specific to the miss-bound proof and is not used for false-alarm control. As a complementary practical insight, our new leading-order delay relationship is supported by 21 non-pilot predictions with a median absolute relative error of 10.8%10.8\% at n=120n=120; it is a tuning heuristic, not an additional guarantee.

4. Interpreting the operational miss bound (Q2). We have expanded the explanations of its two exponential terms and auxiliary assumptions:

  • Test-window term: fluctuations in the test-window embedding, with exponent proportional to mρat2/max⁡ϵt,1/2m\rho a_t^2/\max\\{\epsilon_t,1/2\\} for K=1K=1. Pre-change observations contribute variability but not post-change signal.
  • Reference-metric term: error in the estimated inverse operator. Sensitivity to perturbations of T^P\hat T_P can scale as 1/ρ21/\rho^2; its square produces the ρ4\rho^4 factor in the rate.
  • Residual-margin range condition: selects the quadratic branch of the Bernstein-type rate Ψρ\Psi_\rho; it is not a restriction on running the detector.
  • Neumann-series condition: controls ∥A−1(T^P−TP)∥op≤1/2\|A^{-1}(\hat T_P-T_P)\|_{\rm op}\le1/2 with high probability, enabling the inverse-operator expansion. It is unnecessary for evaluating the statistic or for the false-alarm theorem and corollaries.

Conclusion. We appreciate the reviewer's questions, which have considerably improved the transparency and practical interpretation of the work. The revised manuscript now delineates the strengths and limitations of the guarantees and provides quantitative guidance for their use, alongside the new notation table and offline extension.

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

This paper studies online change-point detection using a regularized Pearson divergence statistic in a reproducing kernel Hilbert space (RKHS). Its main contribution is a non-asymptotic analysis of the proposed detector, establishing finite-sample false-alarm guarantees and bounds on miss-detection probabilities and detection delays. Experiments on synthetic distributions and MNIST evaluate the effects of window lengths and regularization on detection performance.

Critical strengths

The paper addresses an important theoretical problem by establishing non-asymptotic performance guarantees for nonparametric online change-point detection.

Critical weaknesses

The detection statistic is closely related to existing Pearson divergence estimation and uLSIF methods. The main novelty lies in extending non-asymptotic performance analysis to the online setting. While this extension is potentially useful, the theoretical results largely confirm expected trade-offs involving estimation accuracy, window lengths, and regularization, rather than providing fundamentally new insights into sequential change-point detection.

Questions for authors

see above

1 comment

We thank the reviewer for recognizing the importance of non-asymptotic guarantees for nonparametric change-point detection and for prompting a clearer distinction from prior work. The revision separates the known Pearson estimator from our sequential theory and adds testable predictions and qualified comparisons.

1. Relation to prior work. The RKHS Pearson statistic is related to uLSIF (Section RKHS Pearson statistic; Appendix Pearson Divergence, uLSIF, and MMD); we do not claim a new estimator. Existing online density-ratio approaches (e.g., Liu et al., 2013; Ferrari et al., 2023) do not establish the finite-sample sequential guarantees derived here. Earlier kernel change-point results are asymptotic (Harchaoui et al., 2008), or concern MMD-based sequential detection (Kalinke & Gavioli-Akilagun, 2026), which uses a fixed metric. Regularized kernel two-sample tests (Hagrass et al., 2024) analyze individual tests, not sequential stopping times. To our knowledge, this is the first finite-sample analysis of joint reference-embedding and inverse-metric estimation in a sliding-window Pearson detector, covering ARL, infinite-horizon uniform false-alarm control, per-time misses, and delay tails.

2. Technical contributions.

  • Shared reference samples. The reference embedding and inverse metric use the same samples, without sample splitting. The Q1,…,Q4Q_1,\ldots,Q_4 decomposition and Neumann expansion separate test-embedding fluctuations, reference-embedding fluctuations, and inverse-metric perturbations. MMD lacks the last contribution, but not reference-embedding uncertainty.
  • Partial contamination. The effective signal (1−ϵt)∣Δ∣A−1(1-\epsilon_t)\\|\Delta\\|_{A^{-1}} accounts for pre-change observations remaining in the test window. Our miss bound distinguishes sampling noise, contamination, and errors in the reference-dependent metric, and yields a detection-delay tail bound.
  • Overlapping monitoring windows. ARL, uniform false-alarm, and delay-tail arguments use marginal concentration, union bounds, and set inclusion, without assuming successive statistics are independent.

3. Quantitative and practical insights.

  • Delay prediction and tuning. The delay analysis provides a leading-order heuristic for selecting mm. Before the reference window becomes contaminated, a delay D≤m−1D\le m-1 corresponds to D+1D+1 post-change observations in the test window, with effective signal (D+1)∣Δ∣A−1/m(D+1)\\|\Delta\\|_{A^{-1}}/m. Comparing this signal with the threshold and estimation floors suggests (D+1)/m∝1/n+1/m(D+1)/m\propto\sqrt{1/n+1/m} at a fixed false-alarm target. For n=120n=120, the coefficient of variation across mm of (D+1)/(m1/n+1/m)(D+1)/(m\sqrt{1/n+1/m}) is 6.96.9--10.410.4\\% for the four synthetic references (11.911.9\\% for MNIST), versus 22.522.5--26.026.0\\% (28.128.1\\%) for D/mD/m. From a pilot at m0=20m_0=20, the 21 non-pilot predictions have a median absolute relative error of 10.810.8\\% (maximum 24.424.4\\%); see Appendix Joint guarantees and the delay relationship.
  • Distribution-free limits. A new proposition establishes the necessary 1/ρ1/\rho scaling of distribution-free per-time false-alarm thresholds; not optimal Bernstein constants. We now quantify non-vacuity: for n=2mn=2m, K=1K=1, a positive-valued kernel, γ=100\gamma=100, and ρ=10−2\rho=10^{-2}, the analytical ARL threshold first becomes non-vacuous at m=67m=67. With fixed windows, the UFAP threshold eventually becomes vacuous as its logarithmic time factor increases. A finite-memory lemma explains why uniform infinite-horizon control requires increasingly conservative per-time decisions.
  • Direct ARL calibration. Thresholds are calibrated from first-alarm waiting times T−n+1T-n+1 on nominal sliding-window streams and validated on fresh nominal streams. This retains overlap-induced dependence and provides a practical operating point, without a distribution-free empirical guarantee or a fixed-reference geometric approximation.
  • Pearson versus MMD. For nonzero MMD, ρ/(ρ+K)≤ρχ^ρ2/MMD^2≤1\rho/(\rho+K)\le\rho\hat\chi^2_\rho/\hat{\mathrm{MMD}}^2\le1 characterizes the MMD-like large-ρ\rho regime. At a common target ARL of 100, with separately calibrated thresholds, Pearson yields lower miss probabilities on low-variance shifts: 0.2150.215 versus 0.5580.558 for MMD (anisotropic Gaussian, s=0.7s=0.7), and 0.4780.478 versus 0.7160.716 (anisotropic Laplace). The MNIST comparison at m=2m=2 also favors Pearson (0.0080.008 versus 0.2000.200). MMD remains competitive for ordinary shifts and has a slightly lower miss rate for the high-variance Gaussian shift. Realized ARLs are close, not identical; we make no claim of universal superiority.

Conclusion. We appreciate the reviewer's suggestions, which have helped us make the novelty and practical significance of the contribution much clearer. The revised paper now complements the new sequential guarantees with explicit non-vacuity conditions, a necessary regularization scaling, a delay-prediction study, and qualified detector comparisons, while making their limits explicit.

Reviewer_Zc2cOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies online change-point detection using a regularized Pearson divergence in an RKHS. It derives non-asymptotic guarantees for false alarms, miss detection, and detection delay, with experiments examining parameter trade-offs.

Critical strengths
  1. The paper addresses a relevant problem: characterizing finite-sample detection performance while controlling false alarms in sequential monitoring.

  2. The analysis explicitly accounts for partially contaminated test windows and errors in estimating the reference metric, distinguishing the roles of window lengths and regularization.

Critical weaknesses
  1. The practical insights from the bounds are unclear. The tuning guidelines largely reflect intuitive trade-offs, while the experiments rely on empirical calibration. No concrete procedure demonstrates how the theory improves parameter selection.
  2. The bounds appear highly conservative at the experimental scales. With the default ρ=0.01\rho=0.01, Eq. A.108 requires approximately 10610^6 reference samples for δ=0.05\delta=0.05, compared with tens to hundreds in the experiments. For the Table 1 configuration with target ARL 100, the analytical threshold is approximately 104.29, exceeding the statistic’s maximum possible value of 100 and preventing any detection. There are also no matching lower bounds or theory–experiment comparisons establishing sharpness.
  3. The experiments mainly study parameter sensitivity without direct comparisons against strong baselines. They do not establish when Pearson weighting improves detection over MMD or existing density-ratio methods.
Questions for authors
  1. Can the authors demonstrate actionable parameter-selection guidelines derived from the bounds, beyond empirical tuning?
  2. At what sample sizes do the guarantees become non-vacuous? Can the authors compare the bounds with empirical performance and justify the sharpness of the sample-size and regularization dependence?
  3. Can the authors compare against MMD-based and density-ratio detectors at matched realized ARL? Which types of distributional changes benefit from Pearson weighting?
1 comment

We thank the reviewer for the constructive questions, which helped us clarify the practical value and limits of our theory. The revised manuscript now distinguishes conservative distribution-free guarantees from empirical operating points, quantifies when the analytical bounds are non-vacuous, and adds an experimentally tested tuning procedure.

1. Practical parameter selection. Our bounds do not yield a closed-form optimum for (n,m,ρ,ν)(n,m,\rho,\nu), but clarify the roles of the parameters: increasing nn improves nominal embedding and inverse-metric estimation; increasing mm improves concentration but prolongs post-change contamination; and ρ\rho balances directional weighting against metric-estimation error.

The new delay analysis provides a leading-order heuristic for predicting delay when selecting mm. Before the reference window becomes contaminated, a delay D≤m−1D\le m-1 corresponds to D+1D+1 post-change observations in the test window, with effective signal (D+1)∥Δ∥A−1/m(D+1)\|\Delta\|_{A^{-1}}/m. Comparing this signal with the threshold and estimation floors suggests (D+1)/m∝1/n+1/m(D+1)/m\propto\sqrt{1/n+1/m} at a fixed false-alarm target. For n=120n=120, the coefficient of variation across mm of (D+1)/(m1/n+1/m)(D+1)/(m\sqrt{1/n+1/m}) is 6.96.9\\%--10.410.4\\% for the four synthetic references (11.911.9\\% for MNIST), versus 22.522.5\\%--26.026.0\\% (28.128.1\\%) for D/mD/m. From a pilot at m0=20m_0=20, the 21 non-pilot predictions have a median absolute relative error of 10.810.8\\%; see Appendix Joint guarantees and the delay relationship.

Our practical procedure is to (i) calibrate ν\nu directly on nominal sliding-window streams for a target ARL; (ii) use n>mn>m and select ρ\rho using the nominal operator spectrum; and (iii) run a representative post-change pilot to predict delays for candidate mm, choosing the shortest window with reliable in-window detection and acceptable delay. This pilot-based guideline does not purport to guarantee performance for an unspecified post-change distribution QQ.

2. Conservativeness and non-vacuity. We agree that the operating regimes should be explicit. For n=80n=80, m=40m=40, ρ=10−2\rho=10^{-2}, K=1K=1, and γ=100\gamma=100, the sufficient ARL threshold is about 104.3104.3, exceeding the statistic's upper bound of 60.560.5: the guarantee is valid but the resulting threshold is vacuous. A new range lemma, operating-regimes remark, and summary table quantify the window lengths needed for non-vacuous thresholds. Our experiments instead calibrate the actual sliding-window first-alarm time on independent nominal streams, preserving overlap-induced temporal dependence. This gives useful empirical operating points, without claiming a distribution-free guarantee.

A new proposition further shows why the 1/ρ1/\rho scaling cannot generally be removed from distribution-free per-time false-alarm thresholds. Nominal kernel features can be nearly orthogonal, producing a null statistic arbitrarily close to K/(2ρm)+K/[2(K+nρ)]K/(2\rho m)+K/[2(K+n\rho)]. The proposition establishes necessary scaling, not optimality of the remaining constants. Similarly, the large-nn condition for Neumann linearization is sufficient for our miss-bound proof, not a condition for implementing the detector.

3. MMD and density-ratio comparisons. We now distinguish the MMD corollary, which compares upper bounds, from empirical detector comparisons. For nonzero MMD, the operator inequality ρ/(ρ+K)≤ρχ^ρ2/MMD^2≤1\rho/(\rho+K)\le\rho\hat\chi^2_\rho/\hat{\mathrm{MMD}}^2\le1 characterizes the large-ρ\rho MMD-like regime; with K=1K=1 and ρ=3\rho=3, the ratio lies in [0.75,1][0.75,1].

The expanded regularization experiments reveal an informative contrast. Larger ρ\rho generally reduces misses for the four original references (Exponential: 0.3430.343 at 10−410^{-4} versus 0.0740.074 at 33), while small ρ\rho strongly favors some low-variance shifts (AG$_{0.2}$--LV, s=0.7s=0.7: 0.0030.003 at 10−410^{-4} versus 0.6040.604 at 33). In new comparisons calibrated separately to a common target ARL of 100100, Pearson achieves miss probabilities of 0.2150.215 versus 0.5580.558 for MMD on this Gaussian low-variance shift, and 0.0080.008 versus 0.2000.200 on MNIST with m=2m=2. Realized ARLs are close, though not identical. At m=4m=4, both detectors almost always detect the MNIST change, with Pearson alarming 0.80.8 sample earlier on average. Thus Pearson weighting can offer a substantial advantage for low-variance changes, but we do not claim uniform superiority over MMD.

Additional revisions. We have also added analyses of simultaneous guarantees and the finite-memory UFAP limitation, an offline extension, and a notation table. These changes clarify the novelty and limits of the theory while providing experimentally supported tuning guidance.

Comments on the paper

1 comment
AuthorsOpenReview ↗

GENERAL RESPONSE to all reviewers

We thank the three reviewers for their careful reading, thoughtful assessments, and constructive comments. We appreciate their recognition of the relevance and importance of finite-sample performance guarantees for nonparametric online change-point detection, as well as the novelty of the non-asymptotic analysis of the regularized Pearson χ2\chi^2 detector. In particular, our analysis covers partially contaminated test windows, estimation errors in the reference-dependent inverse metric, and complementary criteria for false alarms and detection delay. The reviewers' suggestions have helped us further clarify the originality, scope, and practical implications of our contributions. We address all concerns in the individual responses below. The corresponding changes have been incorporated into the revised manuscript, with the main revisions highlighted in blue. All references below are to the revised manuscript, identified by section, result, or table topic to avoid ambiguity.

Contribution. The estimator is known; the sequential theory is new. To our knowledge, this is the first finite-sample analysis of an online regularized Pearson detector whose inverse metric is learned from the same sliding reference window as the nominal mean embedding. The analysis establishes ARL, infinite-horizon uniform false-alarm, per-time miss-probability, and detection-delay tail guarantees without sample splitting or assuming independence between successive monitoring statistics. It requires independent observations within each regime and a bounded kernel, but no parametric distributional model. The revised manuscript makes this contribution explicit and clearly distinguishes the established estimator from the new theoretical results.

Main revisions

  1. Non-vacuity. The revised false-alarm section includes a statistic-range lemma, a remark, and a table quantifying the window lengths for which the analytical thresholds can trigger an alarm. At n=80n=80, m=40m=40, ρ=10−2\rho=10^{-2}, and target ARL γ=100\gamma=100, the sufficient analytical threshold is 104.3104.3, exceeding the statistic's upper bound of 60.560.5. This motivates empirical calibration without invalidating the analytical guarantee.
  2. Distribution-free price and joint guarantees. A new proposition establishes the necessary 1/ρ1/\rho scaling of thresholds providing distribution-free per-time false-alarm control. The joint-guarantees appendix identifies the additional signal-margin and sample-size requirements for informative miss bounds, clearly distinguishing sufficient analytical conditions from practical detector operation.
  3. Practical calibration. Appendix Empirical threshold calibration now calibrates the actual first-alarm waiting time T−n+1T-n+1 using independent nominal sliding-window streams, preserving the dependence between overlapping windows. Calibration is validated on fresh nominal streams, without relying on the former fixed-reference geometric approximation. Target ARLs from 100100 to 10001000 are reproduced with relative deviations generally within 1010\\% on the four synthetic references and below 1515\\% on MNIST.
  4. Delay relationship and tuning. The heuristic relationship (D+1)/m∝1/n+1/m(D+1)/m\propto\sqrt{1/n+1/m} motivates a practical pilot-based choice of mm. With n=120n=120, the proposed normalization substantially reduces variation across test-window lengths. Across 2121 non-pilot predictions, the median absolute relative error is 10.810.8\\%. These results support the usefulness of the relationship as an empirical approximation, rather than an additional theoretical guarantee.
  5. Pearson versus MMD. An operator inequality characterizes the MMD-like regime obtained for large ρ\rho. New comparisons, with thresholds calibrated separately to a common target ARL, demonstrate that Pearson can substantially improve detection of low-variance shifts for which MMD is less effective, while also showing that Pearson is not uniformly superior. The extended regularization experiments further distinguish low-variance changes from ordinary shifts and clarify when the inverse-operator weighting is beneficial.
  6. Scope and presentation. We added a comprehensive notation table, detailed explanations of the miss-bound terms, a discussion of the limitations imposed by finite-memory monitoring under uniform false-alarm control, and an extension to offline detection. Throughout the manuscript, we also clarified the distinction between rigorous performance guarantees, sufficient analytical conditions, and empirical operating points.

Taken together, these revisions substantially strengthen the paper. They make the novelty and scope of its theoretical contributions more explicit, characterize the limitations of the guarantees transparently, and complement the analysis with practical, experimentally validated guidelines for parameter selection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.