Pith. sign in

REVIEW 2 major objections 5 minor 79 references

You can detect unknown distributional changes without partitioning the model class in advance, with optimal delay and finite-sample false-alarm control.

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 ↗

Aggregating point-null e-processes and minimizing over candidate no-change laws yields ARL- and PFA-valid non-partitioned change detectors with first-order optimal delay under local REGROW conditions.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid non-partitioned e-detector package with matching delay bounds; the local REGROW witness is the real regularity tax, and it is stated cleanly. the 2 major comments →

arxiv 2607.28322 v1 pith:MLRIUKBG submitted 2026-07-30 stat.ME math.STstat.MLstat.TH

Non-partitioned e-detectors for nonparametric sequential change detection

Aytijhya Saha , Aaditya Ramdas This is my paper
classification stat.ME math.STstat.MLstat.TH MSC 62L1062G1062M05
keywords sequential change detectione-processese-detectorsnon-partitionedARLPFAKL optimalitynonparametric
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 reading

Sequential change detection usually assumes you already know which distributions are "before" and which are "after." This paper drops that partition: both pre- and post-change laws are unknown members of one general class P. The method builds a detector by restarting a point-null e-process at every candidate change time, summing those restarts, and taking the worst-case (infimum) over every candidate no-change law. Weighting the sum differently yields either average-run-length control or probability-of-false-alarm control, both with finite-sample guarantees. When P has enough local structure, the resulting delay is first-order optimal and governed by the true KL divergence between the actual post- and pre-change laws. Concrete detectors cover sub-Gaussian and bounded means, Gaussians with unknown variance, and two-state Markov transition changes.

Core claim

Non-partitioned sequential change detection reduces to aggregating point-null e-processes over candidate change times and minimizing over candidate no-change laws. With suitable weights this gives finite-sample ARL or PFA validity over arbitrary P, and under a countable local REGROW witness basis the late-change detection delay is first-order asymptotic to log(1/α)/DKL(Q∥P).

What carries the argument

The SR-style composite e-detector: Dw_t = inf_R ∑_s ws M^R_s:t (or its adjusted nondecreasing version), stopped when it crosses a threshold. Aggregation weights choose ARL versus PFA control; the infimum over R handles the fully composite, non-partitioned null; local REGROW witnesses supply the uniform growth needed for instance-optimal delay.

Load-bearing premise

The model class must have local compact neighborhoods where a single e-process grows at the right rate both near the true pre-change law and far from it; without that local regularity the optimal-delay proof does not go through.

What would settle it

On a class that fails the local witness condition, or in a late-change Gaussian/sub-Gaussian experiment, check whether empirical CADD stays within (1+o(1)) log(A)/DKL(Q∥P) while ARL or PFA stays controlled; a systematic gap above that rate or a false-alarm violation would refute the claim.

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

If this is right

  • Practitioners can monitor for mean or distributional shifts without declaring in advance whether the mean will rise or fall, or which variance is fixed.
  • ARL and PFA are interchangeable design choices via the same e-process building blocks and different start weights.
  • Instance-specific KL delay rates remain attainable even when every post-change law is also a legal no-change law.
  • The same template extends beyond i.i.d. data, as shown for changes in two-state Markov transition matrices.
  • Early changes are information-theoretically hard: global PFA control can make reliable detection impossible if the pre-change sample is too short.

Where Pith is reading between the lines

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

  • Any setting that already has strong point-null e-processes (exchangeability, conformal martingales, other exponential families) is a candidate for the same non-partitioned SR reduction.
  • Computational cost of the all-start infimum will push practice toward pruning, geometric start grids, or lower-bound mixtures that preserve validity.
  • The early-versus-late change distinction suggests hybrid monitors that loosen global PFA until a minimum pre-change stretch is observed.
  • Instance-optimal full-parameter detectors (e.g., full Gaussian) will systematically beat mean-only studentized detectors once the prefix can learn nuisance parameters.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper studies sequential change detection when both pre- and post-change laws are unknown and lie in a common class P, without a pre-specified partition into pre- and post-change families. It constructs Shiryaev–Roberts-style detectors by aggregating s-delay point-null e-processes over candidate changepoints and taking an infimum over candidate no-change laws R in P; aggregation weights yield finite-sample ARL control (Theorem 2.2) or global PFA control (Theorem 2.3). Under a countable local REGROW witness basis on P (Definition 6.8), adjusted versions of these detectors attain first-order late-change delay of order log(1/α)/DKL(Q∥P), with matching change-of-measure lower bounds (Theorems 6.13, 7.1–7.3; Corollaries 6.14–6.15). Explicit constructions are given for sub-Gaussian means, bounded means (universal portfolios), Gaussians with unknown variance (studentized and full-KL versions), and two-state Markov transition matrices, with supporting simulations and public code.

Significance. Non-partitioned composite change detection is practically important and comparatively underdeveloped; classical CUSUM/SR and recent e-detector theory typically rely on a separated partition of the model class. The reduction to point-null e-processes plus an infimum, together with finite-sample ARL/PFA guarantees that do not need the REGROW assumption, is a clean and reusable contribution. The local-witness condition is stated explicitly, shown to be strictly weaker than global weak compactness (Propositions 6.9–6.10), and paired with matching lower bounds and an early-change impossibility result—giving a coherent information-theoretic picture. Concrete examples, including dependent Markov data and an instance-optimal full-Gaussian detector, plus released code, make the framework usable beyond pure existence theory. If the results hold as stated, this is a substantial advance for nonparametric sequential analysis and e-process methodology.

major comments (2)
  1. [Section 6.2, Definition 6.8, Theorem 6.11] The generic first-order optimality claim (Theorem 6.13 and Corollaries 6.14–6.15) rests on P admitting a countable local REGROW witness basis (Definition 6.8 / Theorem 6.11). The paper correctly treats this as an assumption, proves it is weaker than global weak compactness, and verifies it for the unit-variance Gaussian location family. For a reader who wants to apply the generic theorem to a new nonparametric class, however, there is little guidance on how to construct or certify the simultaneous inner/exterior witnesses in practice. A short remark or checklist in §6.2 (e.g., what must be exhibited for a new P, and when direct growth arguments as in §§3–5 and A are preferable) would make the main theorem more usable without changing the mathematics.
  2. [Section 3, Theorem 3.2; Section 8.1, Table 1] In the sub-Gaussian example, the explicit mixture e-process yields delay bounds in terms of the reverse-information-projection quantity I = (μ−ν)²/(2σ²) (Theorem 3.2–3.3, Eq. 8), which can be strictly smaller than the instance-specific DKL(Q∥P). The text acknowledges this and points to §6 for adaptation, but the experiments in §8.1 only report this I-benchmark. Clarifying in §3 and in Table 1 whether the implemented detector is the generic adjusted construction or the explicit mixture—and, if the latter, that the table does not claim instance-optimality—would avoid over-reading the empirical proximity to log(A)/I.
minor comments (5)
  1. [Section 2, after Eq. (1)] The measurability of DARL_t and DPF A_t (infima over possibly nonparametric P) is flagged in §2 but then left implicit. A one-sentence pointer that all concrete examples reduce to finite-dimensional convex optimizations (or countable dense grids) would reassure readers that the stopping times are well-defined in the cases of interest.
  2. [Section 6.1, Definition 6.3] Notation for adjusted processes (M with under-tilde vs. plain M) is easy to miss on first reading; a brief reminder when the adjuster is first used in the delay proofs (§6.3) would help.
  3. [Section 8.2, Table 2] Table 2 Panel B’s finite-prefix prediction d_FT is defined in the text but not in the table caption; adding the defining display to the caption would make the panel self-contained.
  4. [Section 1 / Section A] The Markov appendix (Section A) is a strong addition; a single forward reference in the introduction or §2 that dependent data are handled in the appendix would improve discoverability.
  5. Minor typos and consistency: “PF A” vs “PFA” spacing; “e-Shiryaev–Roberts” hyphenation; and the arXiv footer date format. None affect correctness.

Circularity Check

1 steps flagged

No significant circularity: delay rates track external KL/reverse-IP quantities; self-citations supply independent e-process primitives, not the claimed optima.

specific steps
  1. self citation load bearing [§1.2 / Def 2.1 and §6.2 Lemma 6.7 (citing Ram and Ramdas 2026b)]
    "Ram and Ramdas (2026b) prove a general existence theorem for sequential tests and e-processes for i.i.d. laws on Polish spaces: weak compactness of the null class is a sufficient condition for power-one tests against the complement, and their REGROW construction yields asymptotically relatively growth-rate optimal e-processes. Since a singleton {R} is weakly compact, their result supplies point-null primitives M^R at the level of existence."

    Overlapping-author citation supplies existence of point-null REGROW e-processes used as primitives. This is ordinary foundational self-citation, not circularity of the delay claim: the paper still proves its own aggregation/infimum validity, local-witness uniform growth, and change-of-measure lower bounds; I* remains the external KL, not a quantity defined by the cited construction.

full rationale

The paper’s load-bearing chain is: (i) restartable point-null e-processes → SR-style infimum detectors with finite-sample ARL/PFA (Thm 2.2–2.3, proved directly); (ii) under a countable local REGROW witness basis, adjusted aggregates grow uniformly at nearly D_KL(Q∥P) post-change and at a positive pre-change exterior rate (Thm 6.11 → 6.13); (iii) matching lower bounds by change of measure against the same external I* (Thm 7.1–7.2). Concrete examples either give explicit supermartingales/martingales with growth equal to reverse information projection (sub-Gaussian mixture, UP betting, studentized/full Gaussian, Jeffreys Markov) or sit in classes that admit the witness basis. None of these steps defines the target delay rate in terms of a fitted free parameter, nor renames a data fit as a first-principles prediction. Self-citations (Ram–Ramdas REGROW existence, Shin et al. e-detectors, Choe–Ramdas adjusters) provide standard primitives with independent definitions; the paper’s novelty is the non-partitioned aggregation/infimum reduction and the local-witness delay theory, which are proved in-place rather than imported as uniqueness that forbids alternatives. Residual caveats (measurability of the infimum; verifying the basis for a new P) are methodological, not circular. Score 1 only for ordinary overlapping-author primitive citations that are not load-bearing for the optimality constants.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

The validity half rests on standard e-process/optional-stopping mathematics. Optimality rests on domain assumptions (Polish observation space, existence of point-null e-processes, late-change scaling) plus the paper’s local REGROW witness basis, which packages the uniform-growth regularity needed for the infimum statistic. Tuning constants (mixture variance ρ, spending weights π_s, ridges in t-predictors) affect finite-sample power but not the first-order KL rate under the stated regimes.

free parameters (4)
  • sub-Gaussian mixture scale ρ = ρ²=1 in experiments
    Chooses the Gaussian mixture width in the point-null e-process (Eq. 6); affects finite-sample power and the log-penalty term, not the leading I = (μ−ν)²/(2σ²) rate.
  • PFA spending weights π_s = canonical form (4)
    Deterministic weights with sum ≤1 (canonical π_s ∝ 1/(s{log(es)}²)); trade a log T additive delay price for global PFA control.
  • t-process ridge hyperparameters (m0,v0,ν0)
    Regularized online mean/variance predictors in the universal-inference t e-process (22); needed for early times and inverse-moment conditions.
  • ARL/PFA thresholds A and α = A=1000 in Table 1
    User-chosen false-alarm levels; delay scales as log A or log(1/α).
axioms (6)
  • domain assumption Observations live on a Polish space; P is a class of probability laws on that space.
    Standing setup in §1; needed for weak topology and REGROW existence results.
  • domain assumption For each singleton null {R}, a point-null (s-delay) e-process exists; weak compactness of a null class yields REGROW e-processes (Ram & Ramdas 2026b).
    Primitives for Definition 2.1 and §6.2; existence is cited, not re-proved.
  • ad hoc to paper P admits a countable local REGROW witness basis (Definition 6.8).
    The paper’s sufficient regularity for uniform growth of the infimum statistic (Theorem 6.11); weaker than global weak compactness but still a structural assumption on P.
  • domain assumption Late-change regime: T/L→∞ and typically log T = o(L) (or T/B→∞ for PFA).
    Required for first-order delay upper bounds and to avoid the early-change impossibility regime of Theorem 7.3.
  • domain assumption Infima defining D_t^{ARL} and D_t^{PFA} are measurable so stopping times are well-defined.
    Explicitly assumed after (1); verified in concrete parametric examples via convexity/continuity.
  • standard math Optional stopping / Ville inequalities for nonnegative supermartingales and e-processes.
    Used throughout validity proofs (Appendix B.1).
invented entities (2)
  • Non-partitioned SR-style e-detector (inf over R of sum of s-delay point-null e-processes) independent evidence
    purpose: Define a single composite statistic with ARL or PFA control when pre- and post-change families are not disjoint.
    Core construction (1)–(3); built from existing e-process primitives rather than a new physical object.
  • Countable local REGROW witness basis independent evidence
    purpose: Encode local compactness/uniform-growth regularity so delay bounds hold without global weak compactness of P.
    Definition 6.8; sufficient conditions verified for weakly compact P and for the Gaussian location family.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Non-partitioned e-detectors for nonparametric sequential change detection." pith.science (2026). https://pith.science/paper/MLRIUKBG

@misc{pith2026260728322,
  author       = {Pith},
  title        = {Pith review of: Non-partitioned e-detectors for nonparametric sequential change detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLRIUKBG}},
  note         = {Machine review of arXiv:2607.28322}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We study the problem of sequential change detection over a general class of probability distributions ($\mathcal P$), where both the pre-change and post-change distributions are unknown and belong to $\mathcal P$. We do not assume a pre-specified partition of $\mathcal P$ into pre- and post-change families. We propose a general class of sequential change detectors obtained by aggregating point-null e-processes over possible changepoints and taking an infimum over candidate no-change distributions. The weights in the aggregation scheme determine whether they attain average run length (ARL) control and probability-of-false-alarm (PFA) control. Under suitable assumptions, we prove that our methods achieve first-order asymptotically optimal detection delay. Concrete examples include sub-Gaussian and bounded mean changes, Gaussian mean changes with unknown variance, as well as changes in Markov transition matrices.

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

79 extracted references · 3 linked inside Pith

  1. [1]

    2014 , publisher=

    Sequential Analysis: Hypothesis Testing and Changepoint Detection , author=. 2014 , publisher=

    2014

  2. [2]

    Annals of Stat

    The numeraire e-variable and reverse information projection , author=. Annals of Stat. , year=

  3. [3]

    Reverse Information Projections and Optimal E-Statistics , year=

    Lardy, Tyron and Grünwald, Peter and Harremoës, Peter , journal=. Reverse Information Projections and Optimal E-Statistics , year=

  4. [4]

    , title = "

    Hinkley, David V. , title = ". Biometrika , volume =

  5. [5]

    K. J. Worsley , journal =. Confidence Regions and Tests for a Change-Point in a Sequence of Exponential Family Random Variables , volume =

  6. [6]

    The Annals of Statistics , volume=

    Optimal change-point detection and localization , author=. The Annals of Statistics , volume=

  7. [7]

    Fast and Optimal Changepoint Detection and Localization using

    Jang, Jayoon and Walther, Guenther , journal=. Fast and Optimal Changepoint Detection and Localization using

  8. [8]

    The Annals of Statistics , year=

    The asymptotic behavior of some nonparametric change-point estimators , author=. The Annals of Statistics , year=

  9. [9]

    Journal of the American Statistical Association , volume=

    The application of statistics as an aid in maintaining quality of a manufactured product , author=. Journal of the American Statistical Association , volume=. 1925 , publisher=

    1925

  10. [10]

    Biometrika , volume=

    Continuous inspection schemes , author=. Biometrika , volume=. 1954 , publisher=

    1954

  11. [11]

    Theory of Probability & Its Applications , volume=

    On optimum methods in quickest detection problems , author=. Theory of Probability & Its Applications , volume=

  12. [12]

    Technometrics , volume=

    A comparison of some control chart procedures , author=. Technometrics , volume=. 1966 , publisher=

    1966

  13. [13]

    The Annals of Mathematical Statistics , year=

    Procedures for reacting to a change in distribution , author=. The Annals of Mathematical Statistics , year=

  14. [14]

    The Annals of Statistics , pages=

    Using the generalized likelihood ratio statistic for sequential detection of a change-point , author=. The Annals of Statistics , pages=. 1995 , publisher=

    1995

  15. [15]

    The New England Journal of Statistics in Data Science , volume =

    Jaehyeok Shin and Aaditya Ramdas and Alessandro Rinaldo , title =. The New England Journal of Statistics in Data Science , volume =. 2023 , pages =

    2023

  16. [16]

    Neural Information Proc

    Kernel change-point analysis , author=. Neural Information Proc. Systems , volume=

  17. [17]

    The Annals of Statistics , volume=

    Nonparametric change-point estimation , author=. The Annals of Statistics , volume=. 1988 , publisher=

    1988

  18. [18]

    Darkhovskh, B. S. , title =. Theory of Probability & Its Applications , volume =

  19. [19]

    Proceedings of the National Academy of Sciences , volume=

    Confidence sequences for mean, variance, and median , author=. Proceedings of the National Academy of Sciences , volume=. 1967 , publisher=

    1967

  20. [20]

    The Annals of Statistics , volume=

    Time-uniform, nonparametric, nonasymptotic confidence sequences , author=. The Annals of Statistics , volume=

  21. [21]

    arXiv preprint arXiv:2309.09111 , year=

    Reducing sequential change detection to sequential estimation , author=. arXiv preprint arXiv:2309.09111 , year=

    Pith/arXiv arXiv

  22. [22]

    , title =

    Jennison, Christopher and Turnbull, Bruce W. , title =. Journal of the Royal Statistical Society: Series B , volume =

  23. [23]

    Etude critique de la notion de collectif , author=. Bull. Amer. Math. Soc , volume=

  24. [24]

    Statistical Science , year=

    Game-theoretic statistics and safe anytime-valid inference , author=. Statistical Science , year=

  25. [25]

    Proceedings of Thirty Sixth Conference on Learning Theory , pages =

    Bregman Deviations of Generic Exponential Families , author =. Proceedings of Thirty Sixth Conference on Learning Theory , pages =. 2023 , volume =

    2023

  26. [26]

    Sequential Analysis , volume =

    Wu, Yanhong , title =. Sequential Analysis , volume =. 2005 , publisher =

    2005

  27. [27]

    Wu, Y. , isbn=. Inference for Change Point and Post Change Means After a. 2007 , publisher=

    2007

  28. [28]

    A lower confidence bound for the change point after a sequential. J. Statist. Planng Inf. , volume =. 2003 , author =

    2003

  29. [29]

    Sequential Analysis , volume =

    Boris Brodsky , title =. Sequential Analysis , volume =. 2010 , publisher =

    2010

  30. [30]

    Sequential Analysis , volume =

    M.S Srivastava and Yanhong Wu , title =. Sequential Analysis , volume =. 1999 , publisher =

    1999

  31. [31]

    Sequential Analysis , volume =

    Edit Gombay , title =. Sequential Analysis , volume =. 2003 , publisher =

    2003

  32. [32]

    Annals of the Institute of Statistical Mathematics , volume =

    Yanhong Wu , title =. Annals of the Institute of Statistical Mathematics , volume =. 2004 , publisher =

    2004

  33. [33]

    2007 , publisher=

    Stochastic orders , author=. 2007 , publisher=

    2007

  34. [34]

    1986 , publisher=

    Testing statistical hypotheses , author=. 1986 , publisher=

    1986

  35. [35]

    Inference for post-change mean by a. J. Statist. Planng Inf. , volume =. 2006 , author =

    2006

  36. [36]

    Grünwald, Peter and de Heide, Rianne and Koolen, Wouter , title =. J. R. Statist. Soc. B , volume =. 2024 , month =

    2024

  37. [37]

    Foundations and Trends in Statistics , volume =

    Hypothesis testing with e-values , author=. Foundations and Trends in Statistics , volume =

  38. [38]

    International Journal of Approximate Reasoning , volume=

    Testing exchangeability: Fork-convexity, supermartingales and e-processes , author=. International Journal of Approximate Reasoning , volume=. 2022 , publisher=

    2022

  39. [39]

    Universal

    Wasserman, Larry and Ramdas, Aaditya and Balakrishnan, Sivaraman , journal=. Universal. 2020 , publisher=

    2020

  40. [40]

    Journal of the Royal Statistical Society Series B (Methodology), with discussion , year=

    Estimating means of bounded random variables by betting , author=. Journal of the Royal Statistical Society Series B (Methodology), with discussion , year=

  41. [41]

    Huber-Robust Likelihood Ratio Tests for Composite Nulls and Alternatives , year=

    Saha, Aytijhya and Ramdas, Aaditya , journal=. Huber-Robust Likelihood Ratio Tests for Composite Nulls and Alternatives , year=

  42. [42]

    International Conference on Artificial Intelligence and Statistics , pages=

    Testing exchangeability by pairwise betting , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization=

    2024

  43. [43]

    Time-uniform

    Howard, Steven R and Ramdas, Aaditya and McAuliffe, Jon and Sekhon, Jasjeet , journal=. Time-uniform

  44. [44]

    The Annals of Mathematical Statistics , pages=

    A robust version of the probability ratio test , author=. The Annals of Mathematical Statistics , pages=. 1965 , publisher=

    1965

  45. [45]

    Proceedings of the 2013 conference on empirical methods in natural language processing , pages=

    Recursive deep models for semantic compositionality over a sentiment treebank , author=. Proceedings of the 2013 conference on empirical methods in natural language processing , pages=

    2013

  46. [46]

    Minimax tests and the

    Huber, Peter J and Strassen, Volker , journal=. Minimax tests and the. 1973 , publisher=

    1973

  47. [47]

    Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas , journal=. Distil

  48. [48]

    Econometrica: Journal of the Econometric Society , pages=

    Monitoring structural change , author=. Econometrica: Journal of the Econometric Society , pages=. 1996 , publisher=

    1996

  49. [49]

    Sequential change point tests based on

    Kirch, Claudia and Stoehr, Christina , journal=. Sequential change point tests based on. 2022 , publisher=

    2022

  50. [50]

    The state of cumulative sum sequential changepoint testing 70 years after

    Aue, Alexander and Kirch, Claudia , journal=. The state of cumulative sum sequential changepoint testing 70 years after. 2024 , publisher=

    2024

  51. [51]

    The Annals of Statistics , volume=

    Optimal detection of a change in distribution , author=. The Annals of Statistics , volume=. 1985 , publisher=

    1985

  52. [52]

    IEEE Transactions on Information Theory , volume=

    Information bounds and quick detection of parameter changes in stochastic systems , author=. IEEE Transactions on Information Theory , volume=. 1998 , publisher=

    1998

  53. [53]

    Statistica Sinica , pages=

    Sequential analysis: some classical problems and new challenges , author=. Statistica Sinica , pages=. 2001 , publisher=

    2001

  54. [54]

    International Conference on Machine Learning , pages=

    Sequential changepoint detection via backward confidence sequences , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  55. [55]

    Conformal and Probabilistic Prediction and Applications , pages=

    Retrain or not retrain: Conformal test martingales for change-point detection , author=. Conformal and Probabilistic Prediction and Applications , pages=. 2021 , organization=

    2021

  56. [56]

    Proceedings of the 20th

    Testing exchangeability on-line , author=. Proceedings of the 20th

  57. [57]

    Power one sequential tests for

    Ram, Ashwin and Ramdas, Aaditya , journal=. Power one sequential tests for

  58. [58]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , pages=

    Combining evidence across filtrations , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , pages=. 2026 , publisher=

    2026

  59. [59]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

    Saha, Aytijhya and Ramdas, Aaditya , title =. Journal of the Royal Statistical Society Series B: Statistical Methodology , year =

  60. [60]

    IEEE Transactions on Information Theory , volume=

    Non-parametric quickest mean-change detection , author=. IEEE Transactions on Information Theory , volume=. 2022 , publisher=

    2022

  61. [61]

    Sequential Analysis , year=

    Anytime-valid t-tests and confidence sequences for Gaussian means with unknown variance , author=. Sequential Analysis , year=

  62. [62]

    The Annals of Statistics , pages=

    On confidence sequences , author=. The Annals of Statistics , pages=

  63. [63]

    Mathematical Finance , volume=

    Universal portfolios , author=. Mathematical Finance , volume=

  64. [64]

    IEEE Transactions on Information Theory , volume=

    Tight concentrations and confidence sequences from the regret of universal portfolio , author=. IEEE Transactions on Information Theory , volume=. 2024 , doi=

    2024

  65. [65]

    IEEE Transactions on Information Theory , volume=

    On confidence sequences for bounded random processes via universal gambling strategies , author=. IEEE Transactions on Information Theory , volume=. 2024 , doi=

    2024

  66. [66]

    arXiv preprint arXiv:2310.01547 , year=

    On the near-optimality of betting confidence sets for bounded means , author=. arXiv preprint arXiv:2310.01547 , year=

    Pith/arXiv arXiv

  67. [67]

    Proceedings of the 37th International Conference on Machine Learning , series=

    Restarted Bayesian Online Change-point Detector Achieves Optimal Detection Delay , author=. Proceedings of the 37th International Conference on Machine Learning , series=. 2020 , publisher=

    2020

  68. [68]

    arXiv preprint arXiv:2112.07549 , year=

    Sequential Change Detection through Empirical Distribution and Universal Codes , author=. arXiv preprint arXiv:2112.07549 , year=

    Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2510.26204 , year=

    Sequential Change Detection Under a Markov Setup with Unknown Pre-Change and Post-Change Distributions , author=. arXiv preprint arXiv:2510.26204 , year=

    arXiv

  70. [70]

    , journal=

    Ross, Gordon J. , journal=. Parametric and Nonparametric Sequential Change Detection in. 2015 , doi=

    2015

  71. [71]

    and Zou, Shaofeng , journal=

    Zhang, Qi and Sun, Zhongchang and Herrera, Luis C. and Zou, Shaofeng , journal=. Data-Driven Quickest Change Detection in

  72. [72]

    Proceedings of the 26th International Conference on Artificial Intelligence and Statistics , series=

    A Contrastive Approach to Online Change Point Detection , author=. Proceedings of the 26th International Conference on Artificial Intelligence and Statistics , series=. 2023 , publisher=

    2023

  73. [73]

    Proceedings of the 39th Conference on Uncertainty in Artificial Intelligence , series=

    Online Heavy-Tailed Change-Point Detection , author=. Proceedings of the 39th Conference on Uncertainty in Artificial Intelligence , series=. 2023 , publisher=

    2023

  74. [74]

    and Fearnhead, Paul and Rigaill, Guillem , journal=

    Romano, Gaetano and Eckley, Idris A. and Fearnhead, Paul and Rigaill, Guillem , journal=. Fast Online Changepoint Detection via Functional Pruning

  75. [75]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Online Multivariate Changepoint Detection: Leveraging Links with Computational Geometry , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2026 , doi=

    2026

  76. [76]

    Validity and Efficiency of the Conformal

    Vovk, Vladimir and Nouretdinov, Ilia and Gammerman, Alexander , booktitle=. Validity and Efficiency of the Conformal. 2025 , publisher=

    2025

  77. [77]

    arXiv preprint arXiv:2602.05272 , year=

    Asymptotically Optimal Sequential Change Detection for Bounded Means , author=. arXiv preprint arXiv:2602.05272 , year=

    arXiv

  78. [78]

    arXiv preprint arXiv:2602.11846 , year=

    Universal Sequential Changepoint Detection of Quantum Observables via Classical Shadows , author=. arXiv preprint arXiv:2602.11846 , year=

    arXiv

  79. [79]

    Proceedings of the 30th International Conference on Algorithmic Learning Theory , series=

    Sequential Change-Point Detection: Laplace Concentration of Scan Statistics and Non-Asymptotic Delay Bounds , author=. Proceedings of the 30th International Conference on Algorithmic Learning Theory , series=. 2019 , publisher=

    2019

This paper was first reviewed by grok-4.5 on July 31, 2026.