Pith. sign in

REVIEW 2 major objections 2 minor 1 cited by

A moment-based approximation of the log-likelihood ratio adapts classical change-point detectors to non-Gaussian and heavy-tailed data.

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 · grok-4.3

2026-06-30 15:20 UTC pith:2R33HZVT

load-bearing objection Moment-based LLR approximation for heavy-tailed change detection with Lean proofs, but KU-PE bound transfer to the projection needs explicit error control. the 2 major comments →

arxiv 2605.23419 v2 pith:2R33HZVT submitted 2026-05-22 stat.ME eess.SPmath.STstat.TH

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

Serhii Zabolotnii This is my paper
classification stat.ME eess.SPmath.STstat.TH
keywords sequential change-point detectionlog-likelihood ratio approximationnon-Gaussian processesCUSUMrobust detectionmoment-based methodsKullback-Leibler projectionheavy-tailed data
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 develops a unified framework for approximating the log-likelihood ratio on a generalized stochastic basis such as polynomials, logarithms, or fractional powers. The approximation relies only on moments up to order 3s and requires no closed-form distribution. It projects the Kullback-Leibler divergence to choose the basis order and converts the classical CUSUM, GRSh, and SRP procedures into versions that remain operative when the data are non-Gaussian. The method targets small relative change-points that alter tail structure or modality rather than overall scale, and it supplies a threshold from Kunchenko's probability-error bound that controls false alarms without empirical tuning. On benchmarks with extreme kurtosis the approximated detectors produce usable performance while classical procedures generate 100 percent false alarms.

Core claim

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 robus

What carries the argument

The generalized stochastic basis approximation of the LLR, which uses moments up to order 3s and treats J(s) = K^T Y as the projection of the Kullback-Leibler divergence onto the chosen basis span.

Load-bearing premise

Moments up to order 3s are sufficient to produce a usable approximation of the log-likelihood ratio when only the shape of the distribution changes.

What would settle it

Apply the method to heavy-tailed sequences with excess kurtosis above 20; if the observed false-alarm rate exceeds the KU-PE bound or detection delay matches that of unadapted CUSUM, the approximation claim fails.

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

If this is right

  • Classical CUSUM, GRSh, and SRP procedures become applicable to non-Gaussian data once the LLR is replaced by the moment-based approximation.
  • Kunchenko's probability-error bound supplies a threshold that controls false alarms without data-driven calibration.
  • The convergence functional J(s) supplies an explicit rule for choosing the approximation order s.
  • The procedure remains operative in the small-relative-change regime where only tail structure or modality shifts.

Where Pith is reading between the lines

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

  • The same moment projection could be recomputed on sliding windows to track slowly varying distributions.
  • Basis selection via J(s) might be replaced by cross-validation on a small held-out segment when moments alone are ambiguous.
  • The framework suggests a route to moment-only versions of other likelihood-ratio tests outside change detection.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The manuscript develops a unified framework for approximating the log-likelihood ratio (LLR) for sequential change-point detection in non-Gaussian processes using generalized stochastic bases (polynomial, logarithmic, or fractional-power) and moments up to order 3s. It adapts classical procedures like CUSUM, GRSh, and SRP, interprets the convergence functional J(s) = K^T Y as the projection of the Kullback-Leibler divergence, provides a criterion for approximation order, uses Kunchenko's KU-PE bound for robust thresholding without tuning, demonstrates performance on nine benchmarks including heavy-tailed data, and verifies core theorems in Lean 4.

Significance. If the central claims hold—particularly that the moment-projection approximation preserves the large-deviation behavior needed for the KU-PE bound to control false-alarm rates in the shape-change regime—the work would offer a practical advance for robust detection on heavy-tailed data where classical methods fail. The formal Lean 4 verification of the core theorems is a clear strength, providing machine-checked guarantees on the stated results.

major comments (2)
  1. [threshold derivation / KU-PE application] The transfer of the KU-PE bound to the approximated statistic J(s) is load-bearing for the no-calibration threshold claim, yet the manuscript supplies no explicit control (e.g., total-variation or tail-probability bound) on the discrepancy between the true LLR and its finite-moment projection in the small-relative-change regime. This gap is not addressed by the moment-order selection criterion alone.
  2. [regime definition and error analysis] The target regime is defined as small relative changes where only tail structure and modality change while energy is nearly constant; however, no analysis shows that the projection error remains negligible in the large-deviation regime on which KU-PE relies. The abstract states the framework but the full derivation of error propagation is missing.
minor comments (2)
  1. [preliminaries] Notation for the basis functions and the vector Y should be introduced with explicit definitions before the statement of J(s) = K^T Y to improve readability.
  2. [experiments] The nine benchmarks are mentioned but the specific datasets, excess-kurtosis values, and quantitative detection-delay comparisons at fixed PFA are not tabulated; adding a summary table would strengthen the empirical claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading, the positive assessment of the framework's potential, and the emphasis on the Lean 4 verification. We address the two major comments point by point below.

read point-by-point responses

  1. Referee: [threshold derivation / KU-PE application] The transfer of the KU-PE bound to the approximated statistic J(s) is load-bearing for the no-calibration threshold claim, yet the manuscript supplies no explicit control (e.g., total-variation or tail-probability bound) on the discrepancy between the true LLR and its finite-moment projection in the small-relative-change regime. This gap is not addressed by the moment-order selection criterion alone.

    Authors: We agree that an explicit quantitative control on the discrepancy between the true LLR and the finite-moment projection J(s) is required to justify transferring the KU-PE bound without calibration. The moment-order selection criterion provides a practical way to choose s but does not itself supply a tail-probability or total-variation bound. We will add a new subsection deriving such a bound under the finite-moment assumption (moments up to order 3s) and the small-relative-change regime, building on the existing Lean 4 formalization of the core projection and convergence results. revision: yes

  2. Referee: [regime definition and error analysis] The target regime is defined as small relative changes where only tail structure and modality change while energy is nearly constant; however, no analysis shows that the projection error remains negligible in the large-deviation regime on which KU-PE relies. The abstract states the framework but the full derivation of error propagation is missing.

    Authors: The small-relative-change regime is defined so that the method targets shape (tail and modality) changes with nearly constant energy, precisely the setting where classical parametric procedures break down on heavy tails. We acknowledge that while the regime is stated and the projection J(s) is interpreted as a KL divergence projection, the manuscript does not contain an explicit derivation showing that the approximation error remains negligible at the large-deviation scale needed for the KU-PE bound. The Lean 4 verification covers the stated core theorems on approximation and convergence but does not extend to this error-propagation step. We will supply the missing derivation in the revision, establishing that the projection error is controlled asymptotically under the regime assumptions. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses external moments, external KU-PE bound, and Lean-verified theorems

full rationale

The paper constructs an LLR approximation J(s)=K^T Y from moments up to order 3s on an external generalized basis and transfers the KU-PE false-alarm bound directly; neither step reduces to a self-defined quantity or a fitted parameter renamed as prediction. The Lean 4 verification of core theorems supplies independent formal content. No self-citation chain is load-bearing, no ansatz is smuggled, and no uniqueness theorem is imported from the authors' prior work. The central claim therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Because only the abstract is available, the ledger is necessarily incomplete. The framework appears to rest on the existence of moments up to order 3s, the validity of the projection interpretation of J(s), and the applicability of KU-PE to the approximated statistic. No explicit free parameters, axioms, or invented entities are stated in the provided text.

pith-pipeline@v0.9.1-grok · 5786 in / 1427 out tokens · 27411 ms · 2026-06-30T15:20:01.682581+00:00 · methodology

0 comments
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.

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 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 2
Figure 2. Figure 2: ADD as a function of the approximation order s for several distributions, with the order chosen by the J(s)-based selection rule. The detection delay decreases as s grows, and the marginal gain saturates beyond the order indicated by J(s). • For scenario C (γ3 = 10, ρ ≈ 0.6), the gain is substantial: ADD falls from 76.6 to 48.7, a detection speedup of 36.4%. This agrees with the theoretical prediction (§2.… 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 3
Figure 3. Figure 3: 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 [PITH_FULL_IMAGE:figures/full_fig_p036_3.png] 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 4
Figure 4. Figure 4: 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. 3. The fractional-power basis Φfrac exhibits systematically lower FAR, owing to the better conditioning of the matrix F, which reduces the estimation error of Var[Λ|H0]. 4. The results confirm that the detection speedup repor… 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 5
Figure 5. Figure 5: Operating-characteristic curves (ADD versus ARL0 / FAR) for mean-shift scenarios, compar￾ing GSA-CUSUM against the classical CUSUM, SCUSUM [61], and the univariate QT-EWMA core [63]. Thresholds are Monte Carlo–calibrated so that all methods are compared at matched ARL0; a lower curve indicates a shorter detection delay at the same false alarm rate. 10 3 × 10 2 4 × 10 2 6 × 10 2 3 In-control ARL0 (achieved)… 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 6
Figure 6. Figure 6: Operating-characteristic curves for pure shape changes at constant mean and variance: a skewness change (γ3 : 0 → 1.5, left) and a kurtosis change (γ4 : 0 → 6, right). Thresholds are Monte Carlo–calibrated to each ARL0. The mean-based CUSUM and SCUSUM are near-blind (delay grows with ARL0); GSA-CUSUM (polynomial s = 3 for skewness, logarithmic basis for kurtosis) and, to a lesser extent, the distribution-f… 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 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. 5.3 Cybersecurity: NSL-KDD The NSL-KDD dataset [Tavallaee et al., 2009] — a cleaned version of KDD Cup 1999 with network intrusion… 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 ↗
Figure 8
Figure 8. Figure 8: 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. D.1.2 Student-t (ν = 5, δ = 0.3) Basis ADD (s=1) ADD (s=2) ∆ADD, % cond(F) poly 68.3 ± 2.6 57.4 ± 2.3 −16.0% ∼ 105 frac 67.8 ± 2.5 52.1 ± 2.1 −23.2% ∼ 102 log 69.5 ± 2.7 54.8 ± 2.2 −21.1% ∼ 103 For Student-t with ν = 5 … view at source ↗
Figure 9
Figure 9. Figure 9: 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 10
Figure 10. Figure 10: 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. |φ|max ADD FAR DetRate 5 195 0% 88% 10 (default) 232 0% 88% 20 248 0% 88% ∞ (no clipping) 241 0% 88% Tighter clipping (5 vs. def… view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Closed-form fractional radial links for elliptical Mahalanobis discriminant analysis

    math.ST 2026-07 conditional novelty 7.0

    The Bayes-optimal classifier for elliptical distributions is derived in closed form from the radial generator, yielding a tuning-free alternative to spline GAMs with proven consistency.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · cited by 1 Pith paper · 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

    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. 64

    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

    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. 65

    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

    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. 66

    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

  61. [61]

    Score-based quick- est change detection for unnormalized models

    Suya Wu, Enmao Diao, Taposh Banerjee, Jie Ding, and Vahid Tarokh. Score-based quick- est change detection for unnormalized models. InProceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 206 ofProceedings of Machine Learning Research, pages 10546–10565. PMLR, 2023

    work page 2023

  62. [62]

    Estimation of non-normalized statistical models by score matching.Jour- nal of Machine Learning Research, 6:695–709, 2005

    Aapo Hyvärinen. Estimation of non-normalized statistical models by score matching.Jour- nal of Machine Learning Research, 6:695–709, 2005. 67

    work page 2005

  63. [63]

    Change detection in multivariate data streams: Online analysis with kernel-QuantTree

    Michelangelo Olmo Nogara Notarianni, Filippo Leveni, Diego Stucchi, Luca Frittoli, and Giacomo Boracchi. Change detection in multivariate data streams: Online analysis with kernel-QuantTree. InAdvanced Analytics and Learning on Temporal Data: 9th ECML PKDD Workshop, AALTD 2024, Vilnius, Lithuania, September 9–13, 2024, Revised Se- lected Papers, volume154...

    work page arXiv 2024

  64. [64]

    Zoubir, and H

    Michael Fauß, Abdelhak M. Zoubir, and H. Vincent Poor. Minimax robust detection: Classic results and recent advances.IEEE Transactions on Signal Processing, 69:2252– 2283, 2021

    work page 2021

  65. [65]

    PhD thesis, Faculté des Sciences de Paris, Paris, 1939

    Jean Ville.Étude critique de la notion de collectif. PhD thesis, Faculté des Sciences de Paris, Paris, 1939

    work page 1939

  66. [66]

    Springer Series in Statistics

    Lucien Le Cam.Asymptotic Methods in Statistical Decision Theory. Springer Series in Statistics. Springer-Verlag, New York, 1986. 68

    work page 1986