Pith. sign in

REVIEW 4 minor 32 references

Universality of e-detectors for ARL control

T0 review · 0 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read E-detectors exactly capture ARL-safe alarm times, the paper proves.

desk verdict A clean, mostly self-contained paper that gives exact representation theorems for ARL and optional-horizon control; the weak-ARL converse is more definitional than deep, but the paper is honest about that and the strong theorems are genuinely useful. read the letter →

arxiv 2608.12660 v1 pith:RI5A3BZM submitted 2026-08-12 math.ST cs.ITeess.SPmath.ITstat.MEstat.TH

Aaditya Ramdas This is my paper · ORCID
classification math.STcs.ITeess.SPmath.ITstat.MEstat.TH MSC 62L1062L15
keywords e-detectoraveragerunlengthchangedetectionfalsealarmoptionalstoppinge-processuniversalitylevel-crossingtime
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether the e-detector formalism—a nonnegative process whose stopped expectation never exceeds elapsed time—exactly captures valid alarm rules in sequential change detection, and answers yes for two false-alarm criteria. It proves that an alarm time $T$ has optional-horizon linear false-alarm control, $P(T\le\sigma)\le E[\sigma]/b$ for every data-dependent stopping time $\sigma$, if and only if $T$ is the level-$b$ first crossing of a strong e-detector. It also proves that $T$ has ordinary average-run-length control, $\inf_P E_P[T]\ge b$, if and only if $T$ is the level-$b$ crossing of a weak e-detector, a variant checked only at its own crossings. The same two theorems hold with calendar time replaced by any budget clock $C$. The practical payoff is that e-detector-based procedures cannot achieve a large ARL by concentrating false alarms early; their full distribution of false alarms is constrained, not just the mean.

What carries the argument

The central objects are e-detectors: nonnegative adapted processes $M$ with $M_0=0$ such that $E_P[M_\tau]\le E_P[\tau]$ for every stopping time $\tau$ and every null law $P$. A weak e-detector only requires this inequality at the process's own level-crossing times $T_c(M)$, which is exactly the condition needed for thresholding to control the ARL. The clock $C$ generalizes the budget from units of calendar time to any nondecreasing adapted process, unifying e-processes ($C_n\equiv 1$) and e-detectors ($C_n=n$). The load-bearing construction is the all-or-nothing indicator $M_n=b\mathbf{1}\{T\le n\}$, which certifies representability of any valid alarm time; strong validity also receives a checkable Snell-envelope certificate via an integrable supermartingale dominating $M_n-C_n$.

What would settle it

Take $b=4$, reveal a Bernoulli($1/2$) at time one, and set $T=1$ on heads and $T=7$ on tails. This has $E[T]=4=b$ but $P(T=1)=1/2>1/b$, so it cannot be the level-$b$ crossing of any strong e-detector (Corollary 4.2), while its indicator process is a weak e-detector. Verifying both facts directly settles the claimed separation between the two universality theorems.

Watch

Extended reading notes

Core claim

The paper's central claim is universality: the class of threshold-crossing times of strong e-detectors coincides exactly with the class of stopping times satisfying $P(T\le\sigma)\le E[\sigma]/b$ for all stopping times $\sigma$, and the class of threshold-crossing times of weak e-detectors coincides exactly with the class of stopping times satisfying $\inf_{P\in\mathcal P}E_P[T]\ge b$. Both directions are shown by a single canonical witness, $M_n=b\mathbf{1}\{T\le n\}$, which is nonnegative, adapted, and crosses level $b$ exactly when $T$ alarms. More generally, for any nondecreasing clock $C$, optional-clock false-alarm control $P(T\le\sigma)\le E[C_\sigma]/b$ is equivalent to $T=T_b(M)$ for a strong $C$-clocked e-process, while $E[C_T]\ge b$ is equivalent to $T=T_b(M)$ for a weak $C$-clocked e-process. Strong e-detectors therefore enforce a distributional guarantee—no adaptive observer can concentrate false alarms—whereas weak e-detectors form the exact language for the mean-only criterion.

Load-bearing premise

The definitions allow the canonical certificate $M_n=b\mathbf{1}\{T\le n\}$, a process with zero evidence before the alarm and an exact jump to $b$ at it; if detector classes were required to supply graded or positive evidence, most ARL-controlled alarm times could not be represented, and the paper explicitly acknowledges this degeneracy.

Editorial extensions

If this is right

  • Every e-detector-based procedure obeys $P(T\le t)\le t/b$ at fixed times and $P(T\le\sigma)\le E[\sigma]/b$ for data-dependent monitoring horizons, so a large ARL cannot be purchased by front-loading false alarms.
  • Any alarm time with ARL at least $b$, however pathological its early-alarm probability, has a weak e-detector representation; the price is that weak e-detectors are not closed under mixtures and have no all-horizon certificate.
  • With calendar time replaced by a clock $C$, the same equivalence gives false-alarm control proportional to expected samples or expected cost when observations are intermittent or unequally priced.
  • Under the minimal alarm filtration, the null run-length laws of strong e-detectors are exactly those satisfying $bP(T\le m)\le E[T\wedge m]$ for all $m$; at exact mean $b$, this is the discrete NBUE condition.
  • Strong e-detectors are convex and stable under independent randomization, allowing a mixture with an external exponential clock to pin the actual ARL between $\gamma$ and $\gamma/\rho+1$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The universality theorems rest on allowing the degenerate indicator detector $b\mathbf{1}\{T\le n\}$; if one required detectors to accumulate graded evidence or to have positive increments, many ARL-valid alarm times would become unrepresentable, so the claimed completeness is a property of the chosen definitions.
  • The weak–strong gap translates into a sharp startup-versus-minimax tradeoff: a time-inhomogeneous Shewhart chart can have first-step power exceeding every strongly representable rule at the same ARL, but its worst-case Lorden delay is strictly worse.
  • The clock formulation suggests immediate cost-aware extensions: any predictable cost process $a_j$ yields a budget $C_n=\sum_{j\le n}a_j$, so optional-clock control charges false alarms per expected unit of exposure—a testable design for intermittent sensing and batch inspection.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 4 minor

Summary. The paper studies the exact expressive power of e-detectors in sequential change detection under average run length control. It introduces a clock abstraction that unifies e-processes and e-detectors, proves that optional-clock false-alarm control is equivalent to thresholding a strong clocked e-process (Theorem 3.3), and that expected-clock lower bounds are equivalent to thresholding a weak clocked e-process (Theorem 3.6). Specializing to calendar time, it obtains two representation theorems: a stopping time T has ARL at least b if and only if it is the level-b crossing of a weak e-detector (Theorem 5.8), and T satisfies the stronger optional-horizon inequality P(T≤σ)≤E[σ]/b for all stopping times σ if and only if it is the level-b crossing of a strong e-detector (Theorem 8.3). The paper also develops structural properties of both classes, characterizes strongly representable run-length laws under the minimal alarm filtration via truncated-mean inequalities and the NBUE condition, and gives explicit examples showing that the weak–strong gap is statistically active. Appendices extend the indicator-certificate method to weighted losses, tail envelopes, hazards, rolling windows, restart fields, repeated alarms, simultaneous threshold families, and conjunctions of criteria.

Significance. If the results are correct, the paper gives a complete answer to a natural question: e-detectors are not merely sufficient for ARL, and the original e-detector is the exact universal object for the stronger optional-horizon criterion. The proofs are short, self-contained, and the examples are explicit and checkable. The paper is unusually transparent about the fact that the converse directions rely on the degenerate all-or-nothing witness b1{T≤n}; this makes the weak universality theorem a formal completeness statement rather than a construction method, exactly as stated in Sections 5 and 9.1. The stress-test concern about circularity therefore does not land as an objection: the equivalence claims are precise properties of the admitted definitions, and the paper scopes its claims accordingly. The Snell-envelope and truncated-mean characterizations are genuine structural contributions beyond the representation theorems.

minor comments (4)
  1. [Title and Abstract] The title and abstract render 'Universality ofe-detectors' without a space, and several inline passages (for example 'Thefailureisnotcosmetic' in Section 5) have the same spacing problem; these should be corrected in the final version.
  2. [Remark 8.12] The identity M_τ = 2 1{τ=2} is correct under the liminf convention for M_τ, but the authors should add one clause stating that at τ=∞ the right-hand side is zero, to preempt confusion with the killed stopped value notation introduced in Eq. (8).
  3. [Appendix G] Condition (85) is called 'left-continuity forced by crossings of a real-valued process'; this is true but not immediate, and a one-sentence justification that crossing families always satisfy it would help readers verify the necessity direction of Theorem G.1.
  4. [Section 9.1] The caveat that weak e-detectors are a representation language rather than a construction algebra is stated clearly in the discussion; consider echoing it in the statement of Theorem 5.8 or in Remark 5.9 so the theorem is not misread as a constructive recipe.

Circularity Check

1 steps flagged · score 2.0 of 10

Weak-ARL universality rests on the degenerate certificate b1{T≤n}; the converse is a repackaging of the ARL bound, though the paper acknowledges this and the soundness directions are substantive.

  1. other [Theorem 5.8 proof; Section 9.1]
    "If 0<c≤b, then Tc(MT,b)=T. Fix P∈P. ... EP[MT,b Tc(MT,b)]=b≤EP[T]=EP[Tc(MT,b)]. ... The weak theorem should be interpreted carefully. Its canonical certificate is a restatement of the target alarm event, and the weak class is not even closed under convex combinations."

    The converse of Theorem 5.8 constructs M=b1{T≤n}. For every level c≤b, its own level-crossing time is exactly T, so the weak e-detector inequality evaluated at the crossing is literally b≤E[T]—the ARL lower bound that was to be represented. The existence of the representing detector is therefore not an independent certificate; it is the target inequality re-expressed in detector notation, made admissible only because Definition 5.1 permits all-or-nothing indicator processes with zero pre-alarm values. The paper explicitly concedes this: 'Its canonical certificate is a restatement of the target alarm event.' Thus the claimed universality of weak e-detectors for ARL has a partly definitional character: the converse direction reduces by construction to the input ARL bound.

full rationale

The paper contains no data fitting, no fitted parameter renamed as a prediction, and no external benchmark used circularly. The target scale b is an external input, and the soundness directions—strong e-detectors imply optional-horizon linear false-alarm control, weak e-detectors imply ARL control—are genuine arguments from the definitions. The converse directions of the two universality theorems are established by the canonical degenerate process M_n=b1{T≤n}. For weak detectors (Theorem 5.8), verifying that this process is weakly valid at its own crossing is exactly the ARL inequality b≤E[T] that defines the class of alarm times being represented; the paper acknowledges this in Section 9.1, calling the canonical certificate 'a restatement of the target alarm event.' This is a definitional premise rather than an empirical overreach: if the weak class were required to have positive pre-alarm values or graded evidence, many ARL-controlled alarm times would not be representable. For strong detectors (Theorem 8.3), the same indicator witness converts the stronger optional-horizon inequality into strong e-detector validity, so the restatement is less acute. The paper's self-citations to Ramdas and Wang (2025) and Ramdas et al. (2022) are contextual and not load-bearing for the main bi-implications, which are proved in the text. No unsupported uniqueness theorem is imported from the authors' prior work. On balance, the central mathematical claims are self-contained, with the acknowledged degeneracy of the canonical certificate giving the universality statements a partly definitional flavor rather than a substantive circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

No free parameters are fitted; b is an external validity level. The paper's premises are standard measure theory plus the e-detector definition from prior work. New mathematical objects (weak e-detectors, clocks) are definitions whose only support is the paper's theorems, which is normal for pure theory but means they carry no external empirical evidence.

assumptions (5)
  • domain assumption Discrete-time filtered probability space with an arbitrary pre-change class P; no independence, i.i.d., domination, or parametric assumptions.
    Section 2.1 sets this as the framework; all results are stated for general adapted processes.
  • standard math Standard measure-theoretic toolkit: Fatou's lemma, monotone convergence, Tonelli's theorem, optional sampling, Snell envelope theory, Doob decomposition.
    Used in Lemma 2.1, Theorems 5.7, 8.10 and throughout the appendices.
  • domain assumption Background definitions and sufficiency results for e-values, e-processes, and e-detectors from prior work (Ramdas and Wang 2025; Shin et al. 2024).
    The paper adopts these as background; they are not re-derived.
  • domain assumption Integrability assumption in Theorem 8.10: C_0=0 and E_P[C_n] < infinity for every finite n.
    Assumed to make Snell envelopes and optimal stopping times well-defined.
  • domain assumption Proper clocks satisfy C_0=0 and C_n upward to infinity almost surely for each P.
    Definition 3.5; required for expected-clock universality.
invented entities (2)
  • Weak e-detector
    purpose: Exactly represents the class of ARL-controlled alarm times as threshold crossings.
    Definition 5.1 introduces this new object; the paper itself notes the canonical certificate is a restatement of the alarm event and the class is not convex. Its properties are proven only inside the paper.
  • Strong C-clocked e-process and clock abstraction
    purpose: Unifies e-processes and e-detectors under a budget clock; used for exact optional-clock representation.
    Definitions 3.1 and 3.5; a new organizing framework whose content is the paper's own theorems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Universality of e-detectors for ARL control." pith.science (2026). https://pith.science/paper/RI5A3BZM

@misc{pith2026260812660,
  author       = {Pith},
  title        = {Pith review of: Universality of e-detectors for ARL control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RI5A3BZM}},
  note         = {Machine review of arXiv:2608.12660}
}
abstract

An e-detector for a pre-change class $\mathcal P$ is a nonnegative process $M$ such that $\mathbb E_P[M_\tau] \leq \mathbb E_P[\tau]$ for all stopping times $\tau$ and all $P \in \mathcal P$. Thresholding e-detectors controls the average run length (ARL): declaring a change at the first time $T_b$ when $M$ crosses $b$ ensures that $\inf_{P \in \mathcal P}\mathbb E_P[T] \geq b$. But e-detectors do substantially more than control the ARL; they also satisfy a \emph{optional-horizon inequality}: \[ P(T_b\leq\sigma)\leq \mathbb E_P[\sigma]/b \] for every data-dependent stopping time (monitoring horizon) \(\sigma\) and $P\in \mathcal P$. In particular, every e-detector-based procedure obeys $P(T\leq t)\leq t/b$ at each fixed $t$, thus avoiding early false alarms. Remarkably, the converse also holds: every stopping time $T$ that satisfies the optional-horizon inequality must in fact arise from thresholding an e-detector. We also derive a universal representation of stopping times that satisfy (only) ARL control. These are represented by \emph{weak} e-detectors, that only require $\mathbb E_P[M_\tau] \leq \mathbb E_P[\tau]$ to hold at all threshold stopping times $T_b$. Appendices present universal representations for other (less common) change detection metrics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 12 canonical work pages

  1. [25]

    doi: 10.1214/23- EJP1019. S. Shekhar and A. Ramdas. Sequential changepoint detection via backward confidence sequences. In Proceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 30908–30930,

    doi:10.1214/23-

  2. [28]

    doi: 10.1007/978-1-4757-1862-1. A. G. Tartakovsky, M. Pollak, and A. S. Polunchenko. Third-order asymptotic optimality of generalized Shiryaev–Roberts changepoint detection procedures.Theory of Probability and Its Applications, 56(3):457–484,

    doi:10.1007/978-1-4757-1862-1

  3. [29]

    doi: 10.1137/S0040585X97985534. A. Tartakovsky, I. Nikiforov, and M. Basseville.Sequential Analysis: Hypothesis Testing and Changepoint Detection. CRC Press, Boca Raton,

    work page doi:10.1137/s0040585x97985534

  4. [33]

    No monotonicity assumption is needed for the representation theorem, although in applicationsw will often be nonincreasing so that early alarms are penalized more heavily

    The quantity w(T ) may be interpreted as the loss incurred by a false alarm at timeT. No monotonicity assumption is needed for the representation theorem, although in applicationsw will often be nonincreasing so that early alarms are penalized more heavily. Theorem A.1(Universality for weighted stopping losses).Fixα> 0and a stopping timeT. The following a...

    work page 1963

  5. [1933]

    doi: 10.1098/rsta.1933.0009. K. Nishina and S. Nishiyuki. False alarm probability function of CUSUM charts.Stochastics and Quality Control, 18(1):101–112,

    arXiv 1933

  6. [1954]

    doi: 10.1093/biomet/41.1-2.100. S. M. Pergamenchtchikov and A. G. Tartakovsky. Asymptotically optimal pointwise and minimax quickest change-point detection for dependent data.Statistical Inference for Stochastic Processes, 21(1):217–259,

    doi:10.1093/biomet/41.1-2.100

  7. [1963]

    doi: 10.1137/1108002. D. Siegmund.Sequential Analysis: Tests and Confidence Intervals. Springer, New York,

    doi:10.1137/1108002

  8. [1966]

    doi: 10.2307/1266688. J. Ruf, M. Larsson, W. M. Koolen, and A. Ramdas. A composite generalization of Ville’s martingale theorem using e-processes.Electronic Journal of Probability, 28:1–21,

    work page doi:10.2307/1266688

Show all 32 references
  1. [1971]

    doi: 10.1214/aoms/1177693055. J. M. Lucas and R. B. Crosier. Fast initial response for CUSUM quality-control schemes: Give your CUSUM a head start.Technometrics, 24(3):199–205,

  2. [1972]

    doi: 10.1093/biomet/59.3.539. T. L. Lai. Sequential changepoint detection in quality control and dynamical systems.Journal of the Royal Statistical Society: Series B, 57(4):613–644,

    doi

  3. [1982]

    doi: 10.1080/00401706.1982.10487759. A. W. Marshall and I. Olkin.Life Distributions: Structure of Nonparametric, Semiparametric, and Parametric Families. Springer, New York,

    1982

  4. [1985]

    doi: 10.1214/aos/1176346587. M. Pollak and A. G. Tartakovsky. Asymptotic exponentiality of the distribution of first exit times for a class of Markov processes with applications to quickest change detection.Theory of Probability and Its Applications, 53(3):430–442,

  5. [1986]

    doi: 10.1214/aos/1176350164. N. U. Nair, P. G. Sankaran, and M. Preeth. Reliability aspects of discrete equilibrium dis- tributions.Communications in Statistics—Theory and Methods, 41(3):500–515,

  6. [1990]

    doi: 10.1214/aos/1176347761. S. W. Roberts. A comparison of some control chart procedures.Technometrics, 8(3):411–430,

  7. [1995]

    doi: 10.1111/j.2517-6161.1995.tb02052.x. G. Lorden. Procedures for reacting to a change in distribution.The Annals of Mathematical Statistics, 42(6):1897–1908,

    1995

  8. [2003]

    doi: 10.1515/EQC.2003.101. E. S. Page. Continuous inspection schemes.Biometrika, 41(1/2):100–115,

    2003 doi

  9. [2006]

    doi: 10.1214/009053605000000859. Y. Mei. Is average run length to false alarm always an informative criterion?Sequential Analysis, 27(4):354–376,

    doi

  10. [2007]

    doi: 10.1007/978-0-387-68477-2. 29 Y. Mei. Sequential change-point detection when unknown parameters are present in the pre-change distribution.The Annals of Statistics, 34(1):92–122,

    doi

  11. [2008]

    doi: 10.1080/07474940802445790. T. L. Molloy and J. J. Ford. Misspecified and asymptotically minimax robust quickest change detection.IEEE Transactions on Signal Processing, 65(21):5730–5742,

    doi

  12. [2009]

    doi: 10.1137/S0040585X97983742. M. Pollak and A. G. Tartakovsky. Optimality properties of the Shiryaev–Roberts procedure. Statistica Sinica, 19(4):1729–1739,

    doi

  13. [2010]

    doi: 10.1214/09-AOS775. H. V. Poor and O. Hadjiliadis.Quickest Detection. Cambridge University Press, Cambridge,

    doi

  14. [2011]

    doi: 10.1109/TIT.2011.2104993. J. Ville.Étude critique de la notion de collectif. Gauthier-Villars, Paris,

    2011

  15. [2012]

    doi: 10.1080/03610926.2010.529525. J. Neyman and E. S. Pearson. On the problem of the most efficient tests of statistical hypotheses. Philosophical Transactions of the Royal Society of London. Series A, 231:289–337,

    2010

  16. [2013]

    doi: 10.1214/13-AOS1094. 31 Guide to the appendices The main text develops the common clock abstraction and then focuses on its two calendar- time specializations: weak e-detectors for ARL and strong e-detectors for optional-horizon linear false-alarm control. The appendices s...

    doi

  17. [2017]

    doi: 10.1109/TSP.2017.2740202. G. V. Moustakides. Optimal stopping times for detecting changes in distributions.The Annals of Statistics, 14(4):1379–1387,

    2017

  18. [2018]

    doi: 10.1007/s11203-016-9149-x. M. Pollak. Optimal detection of a change in distribution.The Annals of Statistics, 13(1):206–227,

    doi

  19. [2021]

    doi: 10.1109/JSAIT.2021.3072962. Y. Xie and D. Siegmund. Sequential multi-sensor change-point detection.The Annals of Statistics, 41(2):670–692,

    2021

  20. [2022]

    arXiv:2009.03167. 30 A. Ramdas, P. Grünwald, V. Vovk, and G. Shafer. Game-theoretic statistics and safe anytime-valid inference.Statistical Science, 38(4):576–601,

    2009 arXiv

  21. [2023]

    doi: 10.1214/23-STS894. A. Ramdas and R. Wang.Hypothesis Testing with E-values. Foundations and Trends in Statistics, 1(1–2):1–390,

    doi

  22. [2024]

    doi: 10.51387/23-NEJSDS51. A. N. Shiryaev. On optimum methods in quickest detection problems.Theory of Probability and Its Applications, 8(1):22–46,

    doi

  23. [2025]

    doi: 10.1561/3600000002. Y. Ritov. Decision theoretic optimality of the CUSUM procedure.The Annals of Statistics, 18(3):1464–1469,

    doi

  24. [2026]

    arXiv:2602.05272. A. Ramdas, J. Ruf, M. Larsson, and W. M. Koolen. Admissible anytime-valid sequential infer- ence must rely on nonnegative martingales.arXiv preprint arXiv:2009.03167, version 3,

    2009

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.