pith. sign in

arxiv: 2605.23419 · v1 · pith:2R33HZVTnew · submitted 2026-05-22 · 📊 stat.ME · eess.SP· math.ST· stat.TH

Generalized Stochastic Approximation of the Log-Likelihood Ratio for Robust Sequential Change-Point Detection

Serhii Zabolotnii This is my paper

Pith reviewed 2026-05-25 04:06 UTC · model grok-4.3

classification 📊 stat.ME eess.SPmath.STstat.TH
keywords change-point detectionlog-likelihood ratio approximationstochastic basisheavy-tailed distributionsCUSUMKunchenko boundnon-Gaussian processesmoment methods
0
0 comments X

The pith

Moment-based approximation of the log-likelihood ratio enables robust change-point detection in heavy-tailed non-Gaussian data.

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

The paper develops a unified framework to approximate the log-likelihood ratio on a generalized stochastic basis using only moments up to order 3s, without requiring the analytic form of the distribution. This approximation adapts classical procedures such as CUSUM, GRSh, and SRP to non-Gaussian settings by interpreting the convergence functional as the projection of the Kullback-Leibler divergence. It targets the regime of small relative change-points that alter distribution shape rather than scale. A threshold derived from Kunchenko's probability-error bound controls the false-alarm rate without tuning, and the approach is shown to function on extremely heavy-tailed data where classical methods produce complete failure.

Core claim

The log-likelihood ratio is approximated as the convergence functional J(s) = K^T Y, the projection of the Kullback-Leibler divergence onto the span of a generalized stochastic basis (polynomial, logarithmic, or fractional-power) using moments up to order 3s, which supplies both a formal order-selection criterion and the means to adapt CUSUM, GRSh, and SRP procedures with a robust threshold from Kunchenko's bound in the small relative change-point regime.

What carries the argument

Generalized stochastic approximation of the log-likelihood ratio on polynomial, logarithmic, or fractional-power bases using moments up to order 3s, interpreted as the projection of the Kullback-Leibler divergence onto the basis span.

If this is right

  • Classical CUSUM, GRSh, and SRP procedures adapt to non-Gaussian data without knowing the full density.
  • The false-alarm rate is controlled by Kunchenko's probability-error bound without empirical tuning.
  • The method remains operative on data with excess kurtosis greater than 20 where classical methods produce 100 percent false alarms.
  • Detection delay is reduced while the guaranteed false-alarm level is maintained.
  • Core theorems receive formal verification in Lean 4.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The moment-based projection approach could extend to other sequential decision tasks that rely on unknown distributions.
  • The order-selection criterion J(s) might be applied to approximate other information measures beyond the log-likelihood ratio.
  • The framework could support formal verification of additional statistical procedures in theorem provers.

Load-bearing premise

Moments up to order 3s on the chosen generalized basis provide a sufficient approximation to the true log-likelihood ratio for small relative change-points, and Kunchenko's bound directly controls the false-alarm rate without further assumptions.

What would settle it

Run the method and classical procedures on the nine public benchmarks with excess kurtosis greater than 20 and observe whether the proposed method maintains a low false-alarm rate and reduced detection delay while classical methods produce 100 percent false alarms.

Figures

Figures reproduced from arXiv: 2605.23419 by Serhii Zabolotnii.

Figure 1
Figure 1. Figure 1: ADD as a function of the target FAR ε for GSA-poly S=2 on the normal distribution (ntrials = 100). The empirical values are consistent with the theoretical asymptote ∝ 1/ √ ε (Theorem 6). Scenario C (ρ < 1) is the most challenging and most practically relevant: the change in signal energy is small, yet the shape of the distribution changes. This is precisely the regime in which the GSA detector with s ≥ 2 … view at source ↗
Figure 2
Figure 2. Figure 2: GSA robustness to contamination (H0 contaminated with Gaussian outliers). Left axis — ADD; right axis — FAR. The PE criterion keeps FAR low even at 10% contamination. ρ < 1, a significant fraction of the Kullback–Leibler divergence resides in moments of order ≥ 3. Confirmation of the Gaussian limit. For γ3 = 0 and s = 1, the GSA detector with the polynomial basis yields an ADD practically identical to that… view at source ↗
Figure 3
Figure 3. Figure 3: Gain from higher approximation order: ADD as a function of ε for s = 1 vs. s = 2 (normal distribution, ntrials = 100). At fixed FAR, the higher order yields a shorter detection delay. 4.4 Robustness Analysis for Weak Signals Of particular interest is the ability of the method to detect small parameter changes (“incipient faults”) in the most challenging scenario C (γ3 = 10). δ ADD (s=1) ADD (s=3) Improveme… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of GSA-CUSUM with baselines on the Laplace distribution (bar chart of ADD; ntrials as in the source experiment in reports/charts/). 4.7 Comparison of Decision Rules: CUSUM, GRSh, SRP To confirm the architectural modularity of the approach (§3.7), we compared three decision rules under a fixed GSA-LLR (Φpoly, s = 3, γ3 = 10, δ = 0.3). Decision rule ADD FAR DetRate Characteristic CUSUM 52.9 ± 2.2 … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of the poly / frac / log bases on the Pearson III distribution with skewness γ3 = 10. The fractional-power basis compresses the dynamic range, enabling lower ADD at the same FAR. 4.8 Summary of Monte Carlo Results The results of Sections 4.1–4.7, obtained on synthetic data with controlled parameters, support the following conclusions: 1. Benefit of higher approximation orders: Increasing s from … view at source ↗
Figure 6
Figure 6. Figure 6: NASA IMS 2nd_test, feature = vibration kurtosis. The true CP is at position 100 (black vertical line). Six GSA variants (green) trigger after the CP — true detections; classical Sign-CUSUM, MAD-CUSUM, EWMA (red) trigger before the CP — false alarms. γ series 4 = 70.16 for the full series, γ calib 4 = 6.5 for the calibration subsample. Source data: paper/shared/results_manifest.json → tier1_datasets.nasa_im… view at source ↗
Figure 7
Figure 7. Figure 7: NSL-KDD (15 traffic windows after log(1 + x)). FAR-DetRate space; the ideal corner is (0, 1). Four GSA variants (S=1 poly/frac/log and S=2 log) cluster at the ideal corner; Sign-CUSUM is nearly there (FAR=6.7%, DetRate=86.7%); MAD-CUSUM misses 73% of attacks; EWMA detects none. generate frequent false alarms. GSA with the PE threshold analytically accounts for the shape of the distribution, which allows it… view at source ↗
Figure 8
Figure 8. Figure 8: Ablation study of three detector parameters (exp8): winsorization, basis function clipping, and threshold scaling. Base configuration: Pearson III γ3=10, δ=0.3, GSA-poly S=2, ntrials=50. Labels show FAR / DR; green indicates FAR ≤ 5%, red indicates violation of the PE guarantee. This is the most influential parameter. sh = 1.5 yields 30% faster detection than the default sh = 2.0, at the same FAR= 0% and D… view at source ↗
read the original abstract

Sequential change-point detection in non-Gaussian stochastic processes is challenging because the underlying densities are rarely known in real time. Classical parametric procedures such as CUSUM lose optimality under distributional mismatch, whereas nonparametric alternatives often react slowly. We develop a unified framework that approximates the log-likelihood ratio (LLR) on a generalized stochastic basis -- polynomial, logarithmic, or fractional-power -- using only moments up to order 3s, with no analytic form of the distribution, and thereby adapts the classical CUSUM, GRSh, and SRP procedures to non-Gaussian data. The convergence functional J(s) = K^T Y is interpreted as the projection of the Kullback-Leibler divergence onto the basis span, yielding a formal criterion for selecting the approximation order. We target the regime of small relative change-points, where the signal energy changes little but the shape of the distribution -- tail structure and modality -- does. A robust threshold follows from Kunchenko's probability-error bound (KU-PE), which controls the false-alarm rate without empirical tuning. On nine public benchmarks across four domains, the method is, to our knowledge, the only one operative on extremely heavy-tailed data (excess kurtosis gamma_4 > 20), where classical methods produce 100% false alarms, while reducing the detection delay at a guaranteed false-alarm level. The core theorems are formally verified in Lean 4.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript develops a unified framework for sequential change-point detection in non-Gaussian processes by approximating the log-likelihood ratio via a generalized stochastic basis (polynomial, logarithmic, or fractional-power) using only moments up to order 3s. The approximation is formalized as the projection J(s) = K^T Y of the Kullback-Leibler divergence, which is then plugged into adapted CUSUM, GRSh, and SRP procedures; a threshold is obtained from Kunchenko's KU-PE probability-error bound to guarantee the false-alarm rate without tuning. The target regime is small relative change-points that alter distribution shape or tails rather than energy. Performance is reported on nine public benchmarks, with the claim that the method is the only one that remains operative for data with excess kurtosis gamma_4 > 20 (where classical methods yield 100% false alarms) while reducing detection delay; core theorems are formally verified in Lean 4.

Significance. If the projection error can be shown to be controlled so that the KU-PE bound remains valid, the work would constitute a meaningful advance in robust nonparametric sequential detection by supplying a moment-based procedure that retains formal false-alarm guarantees on heavy-tailed data where parametric and many nonparametric alternatives break down. The Lean 4 verification of the core theorems is a clear strength that supports reproducibility and rigor.

major comments (3)
  1. [Threshold derivation using KU-PE] The direct invocation of Kunchenko's KU-PE bound on the approximated statistic J(s) (described after the definition of the projection) supplies no explicit additive error term that accounts for the discrepancy between the finite-basis projection and the true LLR when the underlying distribution has gamma_4 > 20; because the bound is asserted to deliver a guaranteed false-alarm rate, this omission is load-bearing for the central robustness claim.
  2. [Moment approximation up to 3s] In the regime of small relative change-points (shape/tail shifts with little energy change), the manuscript states that moments up to order 3s suffice for the generalized basis approximation, yet provides no quantitative uniform bound on the resulting projection error that is independent of tail heaviness; without such control the crossing probabilities of the adapted CUSUM/GRSh/SRP procedures may exceed the KU-PE guarantee.
  3. [Lean verification of core theorems] The Lean-verified theorems address the exact (non-approximated) LLR case; the manuscript must clarify whether and how the finite-basis projection step is incorporated into the verified statements or bounded so that the practical procedure inherits the same guarantees.
minor comments (2)
  1. The nine benchmarks are referenced only by domain; adding a table that lists the specific datasets, their sample sizes, and measured gamma_4 values would improve reproducibility.
  2. Notation for the basis functions and the vector K should be introduced with an explicit definition before the first use of J(s) = K^T Y.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for the thorough review and valuable feedback on our manuscript. The comments have helped us identify areas where the presentation and theoretical guarantees can be strengthened. We provide point-by-point responses to the major comments below, indicating the revisions we plan to make.

read point-by-point responses

  1. Referee: [Threshold derivation using KU-PE] The direct invocation of Kunchenko's KU-PE bound on the approximated statistic J(s) (described after the definition of the projection) supplies no explicit additive error term that accounts for the discrepancy between the finite-basis projection and the true LLR when the underlying distribution has gamma_4 > 20; because the bound is asserted to deliver a guaranteed false-alarm rate, this omission is load-bearing for the central robustness claim.

    Authors: We thank the referee for highlighting this important aspect. The manuscript applies the KU-PE bound directly to J(s) as it is the observable statistic. To address the concern, we will revise the manuscript to include an analysis of the approximation error between J(s) and the true LLR, deriving an explicit bound in terms of the basis truncation and moment estimates. This will show that the error is bounded by a term that vanishes as s increases, allowing the false-alarm guarantee to hold with a modified threshold that accounts for the additive error. revision: yes

  2. Referee: [Moment approximation up to 3s] In the regime of small relative change-points (shape/tail shifts with little energy change), the manuscript states that moments up to order 3s suffice for the generalized basis approximation, yet provides no quantitative uniform bound on the resulting projection error that is independent of tail heaviness; without such control the crossing probabilities of the adapted CUSUM/GRSh/SRP procedures may exceed the KU-PE guarantee.

    Authors: The referee correctly notes the absence of a quantitative uniform bound independent of tail heaviness. Our framework is designed for small relative change-points where moments up to 3s capture the shape and tail changes. In the revision, we will add a quantitative bound on the projection error that depends on higher moments and demonstrate how increasing s controls the error to maintain the KU-PE guarantee. revision: yes

  3. Referee: [Lean verification of core theorems] The Lean-verified theorems address the exact (non-approximated) LLR case; the manuscript must clarify whether and how the finite-basis projection step is incorporated into the verified statements or bounded so that the practical procedure inherits the same guarantees.

    Authors: We will revise the manuscript to clarify that the Lean verification is for the exact LLR. We will incorporate a new result bounding the projection error and showing that the guarantees extend to the approximated statistic with high probability when s is chosen based on the data moments. This will be presented as an additional theorem. revision: yes

Circularity Check

0 steps flagged

No circularity: moment-based LLR projection and KU-PE bound are independent of target result

full rationale

The derivation approximates the LLR via finite moments on a chosen basis and interprets J(s)=K^T Y as its KL projection, then feeds the statistic into standard CUSUM/GRSh/SRP procedures whose false-alarm control is taken from the external Kunchenko KU-PE bound. The core theorems are stated to be Lean-verified. No equation reduces a prediction to a fitted parameter by construction, no self-citation chain bears the central claim, and the approximation order selection is presented as an explicit functional rather than a tautology. The supplied abstract and context contain no load-bearing step that collapses to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The framework relies on standard statistical assumptions about moments and the applicability of the cited bound; no new entities are postulated.

axioms (2)
  • domain assumption Moments up to order 3s are available and sufficient for the approximation
    The method relies on using only these moments without the analytic form of the distribution.
  • domain assumption Kunchenko's probability-error bound (KU-PE) provides a robust threshold controlling false-alarm rate
    Used for the threshold without empirical tuning.

pith-pipeline@v0.9.0 · 5786 in / 1397 out tokens · 29194 ms · 2026-05-25T04:06:43.772103+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 60 canonical work pages · 2 internal anchors

  1. [1]

    S. W. Zabolotnii and Z. L. Warsza. Semi-parametric polynomial modification of cusum algorithms for change-point detection of non-gaussian sequences. InProc. XXI IMEKO World Congress, pages 2088–2091, 2091

    work page 2088

  2. [2]

    Zabolotnii and Zygmunt L

    Serhii W. Zabolotnii and Zygmunt L. Warsza.Semi-parametric Estimation of the Change- Point of Parameters of Non-gaussian Sequences by Polynomial Maximization Method, page 903–919. Springer International Publishing, 2016

    work page 2016

  3. [3]

    Yu. P. Kunchenko.Polynomial Parameter Estimations of Close to Gaussian Random Vari- ables. Shaker Verlag, 2002

    work page 2002

  4. [4]

    Yu. P. Kunchenko.Stochastic Polynomials (in Russian). Naukova Dumka, Kyiv, 2006

    work page 2006

  5. [5]

    A. Wald. Sequential tests of statistical hypotheses.The Annals of Mathematical Statistics, 16(2):117–186, 1945

    work page 1945

  6. [6]

    Wald.Sequential Analysis

    A. Wald.Sequential Analysis. New York: Wiley, 1947

    work page 1947

  7. [7]

    E. S. PAGE. Continuous inspection schemes.Biometrika, 41(1-2):100–115, 1954

    work page 1954

  8. [8]

    A. N. Shiryaev. On optimum methods in quickest detection problems.Theory of Probability & Its Applications, 8(1):22–46, January 1963

    work page 1963

  9. [9]

    S. W. Roberts. A comparison of some control chart procedures.Technometrics, 8(3):411–430, August 1966. 58

    work page 1966

  10. [10]

    M. A. Girshick and Herman Rubin. A bayes approach to a quality control model.The Annals of Mathematical Statistics, 23(1):114–125, March 1952

    work page 1952

  11. [11]

    G. Lorden. Procedures for reacting to a change in distribution.The Annals of Mathematical Statistics, 42(6):1897–1908, December 1971

    work page 1908

  12. [12]

    Moustakides

    George V. Moustakides. Optimal stopping times for detecting changes in distributions.The Annals of Statistics, 14(4):1379–1387, December 1986

    work page 1986

  13. [13]

    Optimal detection of a change in distribution.The Annals of Statistics, 13(1):206–227, March 1985

    Moshe Pollak. Optimal detection of a change in distribution.The Annals of Statistics, 13(1):206–227, March 1985

    work page 1985

  14. [14]

    Basseville and I

    M. Basseville and I. V. Nikiforov.Detection of Abrupt Changes: Theory and Application. Prentice-Hall, 1993

    work page 1993

  15. [15]

    Chap- man and Hall/CRC, August 2014

    Alexander Tartakovsky, Igor Nikiforov, and Michele Basseville.Sequential Analysis. Chap- man and Hall/CRC, August 2014

    work page 2014

  16. [16]

    Vincent Poor and Olympia Hadjiliadis.Quickest Detection

    H. Vincent Poor and Olympia Hadjiliadis.Quickest Detection. Cambridge University Press, November 2008

    work page 2008

  17. [17]

    An efficient sequential nonparametric scheme for detecting a change of distribution.The Annals of Statistics, 22(2):763–804, 1994

    Louis Gordon and Moshe Pollak. An efficient sequential nonparametric scheme for detecting a change of distribution.The Annals of Statistics, 22(2):763–804, 1994

    work page 1994

  18. [18]

    G. J. Ross, D. K. Tasoulis, and N. M. Adams. Nonparametric monitoring of data streams for changes in location and scale.Technometrics, 53(4):379–389, 2011

    work page 2011

  19. [19]

    Hafiz Zafar Nazir, Muhammad Riaz, Ronald J. M. M. Does, and Nasir Abbas. Robust cusum control charting.Quality Engineering, 25(3):211–224, 2013

    work page 2013

  20. [20]

    Harchaoui, F

    Z. Harchaoui, F. R. Bach, and É. Moulines. Kernel change-point analysis. InNIPS 2008, pages 609–616, 2008

    work page 2008

  21. [21]

    Arlot, A

    S. Arlot, A. Celisse, and Z. Harchaoui. A kernel multiple change-point algorithm via model selection.JMLR, 20(162):1–56, 2019

    work page 2019

  22. [22]

    Desobry, M

    F. Desobry, M. Davy, and C. Doncarli. An online kernel change detection algorithm.IEEE Transactions on Signal Processing, 53(8):2961–2974, August 2005

    work page 2005

  23. [23]

    S. Li, Y. Xie, H. Dai, and L. Song. M-statistic for kernel change-point detection. InNIPS 2015, pages 3366–3374, 2015

    work page 2015

  24. [24]

    B. E. Brodsky and B. S. Darkhovsky.Nonparametric Methods in Change-Point Problems. Springer Netherlands, 1993

    work page 1993

  25. [25]

    R. P. Adams and D. J. C. MacKay. Bayesian online changepoint detection, 2007. arXiv:0710.3742

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  26. [26]

    Springer New York, 1998

    DouglasM.HawkinsandDavidH.Olwell.Cumulative Sum Charts and Charting for Quality Improvement. Springer New York, 1998

    work page 1998

  27. [27]

    Barndorff-Nielsen and D

    O. Barndorff-Nielsen and D. R. Cox. Edgeworth and saddle-point approximations with sta- tistical applications.Journal of the Royal Statistical Society Series B: Statistical Method- ology, 41(3):279–312, 1979

    work page 1979

  28. [28]

    H. E. Daniels. Saddlepoint approximations in statistics.The Annals of Mathematical Statistics, 25(4):631–650, December 1954. 59

    work page 1954

  29. [29]

    Information bounds and quick detection of parameter changes in stochastic systems.IEEE Transactions on Information Theory, 44(7):2917–2929, 1998

    Tze Leung Lai. Information bounds and quick detection of parameter changes in stochastic systems.IEEE Transactions on Information Theory, 44(7):2917–2929, 1998

    work page 1998

  30. [30]

    Sequential changepoint detection in quality control and dynamical systems

    Tze Leung Lai. Sequential changepoint detection in quality control and dynamical systems. Journal of the Royal Statistical Society Series B: Statistical Methodology, 57(4):613–658, November 1995

    work page 1995

  31. [31]

    Siegmund and E

    D. Siegmund and E. S. Venkatraman. Using the generalized likelihood ratio statistic for sequential detection of a change-point.The Annals of Statistics, 23(1):255–271, February 1995

    work page 1995

  32. [32]

    Willsky and H

    A. Willsky and H. Jones. A generalized likelihood ratio approach to the detection and estimation of jumps in linear systems.IEEE Transactions on Automatic Control, 21(1):108–112, February 1976

    work page 1976

  33. [33]

    Sprt and cusum in hidden markov models.The Annals of Statistics, 31(3):942–977, 2003

    Cheng-Der Fuh. Sprt and cusum in hidden markov models.The Annals of Statistics, 31(3):942–977, 2003

    work page 2003

  34. [34]

    Yu. P. Kunchenko.Approximation Polynomials in a Space with a Generating Element (in Russian). Naukova Dumka, Kyiv, 2003

    work page 2003

  35. [35]

    Yu. P. Kunchenko. A moment performance criteria of a decision-making for testing simple statistical hypothesis. InProc. IEEE International Symposium on Information Theory (ISIT), page 408, 1997

    work page 1997

  36. [36]

    Kunchenko's Polynomials for Template Matching

    O. Chertov and T. Slipets. Kunchenko’s polynomials for template matching. InProc. IEEE 6th International Conference on Intelligent Data Acquisition and Advanced Computing Sys- tems (IDAACS), pages 425–428, 2011. arXiv:1107.2085

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  37. [37]

    Signal detection in correlated non-gaussian noise using higher-order statistics.Circuits, Systems, and Signal Processing, 37(4):1704–1723, August 2017

    Elena Palahina, Mária Gamcová, Iveta Gladišová, Ján Gamec, and Volodymyr Palahin. Signal detection in correlated non-gaussian noise using higher-order statistics.Circuits, Systems, and Signal Processing, 37(4):1704–1723, August 2017

    work page 2017

  38. [38]

    Zabolotnii, S

    Serhii W. Zabolotnii, S. S. Martynenko, and S. V. Salypa. Method of verification of hy- pothesis about mean value on a basis of expansion in a space with generating element. Radioelectronics and Communications Systems, 61(5):222–229, May 2018

    work page 2018

  39. [39]

    Warsza.Application of the Poly- nomial Maximization Method for Estimation Parameters of Autoregressive Models with Asymmetric Innovations, page 380–390

    Serhii Zabolotnii, Oleksandr Tkachenko, and Zygmunt L. Warsza.Application of the Poly- nomial Maximization Method for Estimation Parameters of Autoregressive Models with Asymmetric Innovations, page 380–390. Springer International Publishing, 2022

    work page 2022

  40. [40]

    Springer Nature Switzerland, 2023

    Serhii Zabolotnii, Oleksandr Tkachenko, and Zygmunt Lech Warsza.Polynomial Maxi- mization Method for Estimation Parameters of Asymmetric Non-Gaussian Moving Average Models, page 223–231. Springer Nature Switzerland, 2023

    work page 2023

  41. [41]

    Springer International Publishing, 2018

    Zygmunt Lech Warsza and Serhii Zabolotnii.Estimation of Measurand Parameters for Data from Asymmetric Distributions by Polynomial Maximization Method, page 746–757. Springer International Publishing, 2018

    work page 2018

  42. [42]

    Serhii Zabolotnii, Oleksandr Tkachenko, Waldemar Nowakowski, and Zygmunt L. Warsza. Application of the Polynomial Maximization Method for Estimating Nonlinear Regression Parameters with Non-gaussian Asymmetric Errors, page 342–356. Springer Nature Switzer- land, 2024

    work page 2024

  43. [43]

    Selective review of offline change point detection methods.Signal Processing, 167:107299, February 2020

    Charles Truong, Laurent Oudre, and Nicolas Vayatis. Selective review of offline change point detection methods.Signal Processing, 167:107299, February 2020. 60

    work page 2020

  44. [44]

    Killick, P

    R. Killick, P. Fearnhead, and I. A. Eckley. Optimal detection of changepoints with a linear computational cost.Journal of the American Statistical Association, 107(500):1590–1598, October 2012

    work page 2012

  45. [45]

    A. J. Scott and M. Knott. A cluster analysis method for grouping means in the analysis of variance.Biometrics, 30(3):507–512, 1974

    work page 1974

  46. [46]

    G. J. J. van den Burg and C. K. I. Williams. An evaluation of change point detection algorithms, 2020. arXiv:2003.06222

    work page arXiv 2020

  47. [47]

    Evaluating real-time anomaly detection algorithms – the numenta anomaly benchmark

    Alexander Lavin and Subutai Ahmad. Evaluating real-time anomaly detection algorithms – the numenta anomaly benchmark. In2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA), page 38–44. IEEE, December 2015

    work page 2015

  48. [48]

    I. D. Katser and V. O. Kozitsin. Skoltech anomaly benchmark (skab). 2020

    work page 2020

  49. [49]

    Kats, 2022.github.com/facebookresearch/Kats

    Meta / Facebook Research. Kats, 2022.github.com/facebookresearch/Kats

    work page 2022

  50. [50]

    Montiel et al

    J. Montiel et al. River: machine learning for streaming data in python.JMLR, 22(110):1–8, 2021

    work page 2021

  51. [51]

    Wavelet filter-based weak signature detection method and its application on rolling element bearing prognostics.Journal of Sound and Vibration, 289(4-5):1066–1090, February 2006

    Hai Qiu, Jay Lee, Jing Lin, and Gang Yu. Wavelet filter-based weak signature detection method and its application on rolling element bearing prognostics.Journal of Sound and Vibration, 289(4-5):1066–1090, February 2006

    work page 2006

  52. [52]

    Dyer and R

    D. Dyer and R. M. Stewart. Detection of rolling element bearing damage by statistical vibration analysis.Journal of Mechanical Design, 100(2):229–235, April 1978

    work page 1978

  53. [53]

    Tartakovsky, B.L

    A.G. Tartakovsky, B.L. Rozovskii, R.B. Blazek, and Hongjoong Kim. A novel approach to detection of intrusions in computer networks via adaptive sequential and batch-sequential change-pointdetectionmethods.IEEE Transactions on Signal Processing, 54(9):3372–3382, 2006

    work page 2006

  54. [54]

    İlker Özçelik and Richard R. Brooks. Cusum - entropy: an efficient method for ddos attack detection. In2016 4th International Istanbul Smart Grid Congress and Fair (ICSG), page 1–5. IEEE, April 2016

    work page 2016

  55. [55]

    V. A. Siris and F. Papagalou. Application of anomaly detection algorithms for detecting syn flooding attacks.Computer Communications, 29(9):1433–1442, 2006

    work page 2006

  56. [56]

    Bai and P

    J. Bai and P. Perron. Estimating and testing linear models with multiple structural changes. Econometrica, 66(1):47–78, 1998

    work page 1998

  57. [57]

    Computation and analysis of multiple structural change models.Journal of Applied Econometrics, 18(1):1–22, 2003

    Jushan Bai and Pierre Perron. Computation and analysis of multiple structural change models.Journal of Applied Econometrics, 18(1):1–22, 2003

    work page 2003

  58. [58]

    Carla Inclán and George C. Tiao. Use of cumulative sums of squares for retrospec- tive detection of changes of variance.Journal of the American Statistical Association, 89(427):913–923, 1994

    work page 1994

  59. [59]

    Sepinav (sepsis icu navigator): A data-driven software tool for sepsis monitoring and intervention using bayesian online change point detection.SoftwareX, 14:100689, 2021

    Nazmus Sakib, Shiyu Tian, Md Munirul Haque, Rumi Ahmed Khan, and Sheikh Iqbal Ahamed. Sepinav (sepsis icu navigator): A data-driven software tool for sepsis monitoring and intervention using bayesian online change point detection.SoftwareX, 14:100689, 2021

    work page 2021

  60. [60]

    Stephanie L. Hyland, Martin Faltys, Matthias Hüser, Xinrui Lyu, Thomas Gumbsch, Cristóbal Esteban, Christian Bock, Max Horn, Michael Moor, Bastian Rieck, Marc Zimmer- mann, Dean Bodenham, Karsten Borgwardt, Gunnar Rätsch, and Tobias M. Merz. Early prediction of circulatory failure in the intensive care unit using machine learning.Nature Medicine, 26(3):36...

    work page 2020