{"id":"ba919c53-93b2-4769-8412-cd79e72d2f3d","arxiv_id":"2608.12660","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Any stopping time with expected alarm time at least b is the level-b crossing of a weak e-detector, and any stopping time satisfying optional-horizon false-alarm control is the level-b crossing of a strong e-detector.","lead":"This paper proves exact representation theorems: every change-detection rule that only controls the average run length is a threshold crossing of a weak e-detector, and every rule that also controls false alarms at data-dependent horizons is a threshold crossing of a strong e-detector. Read it to understand what ARL guarantees really mean and why e-detector constructions are stronger than a bare ARL bound.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Weak-ARL universality is exact but rests on the degenerate certificate b1{T≤n}; the theorem is close to a restatement of the ARL bound unless the all-or-nothing indicator is admitted.","rationale":"I read the paper as establishing two exact representation theorems: strong e-detectors correspond to optional-horizon linear false-alarm control (Theorem 8.3), and weak e-detectors correspond to ARL control (Theorem 5.8), with clocked generalizations in Section 3. I checked the key proof steps: Lemma 2.1's localization by Fatou/monotone convergence is sound; the canonical witnesses satisfy the required inequalities essentially by substitution; the Snell-envelope and martingale-domination argument in Theorem 8.10 is valid under the stated integrability assumptions; and the minimal-filtration characterization in Theorem 8.6 correctly reduces all stopping horizons to deterministic self-censoring horizons via Lemma 8.5. The Gaussian fast-start example and the adaptive-horizon counterexample Example 8.13 check out. The one caveat that is genuinely load-bearing is the degenerate all-or-nothing certificate used in the converse theorems. For weak e-detectors this makes the universality statement nearly tautological: the certificate encodes T itself and validates itself by the same ARL inequality it is supposed to represent. The paper acknowledges this, and the weak class's nonconvexity (Proposition 5.4) confirms that it is a representation language rather than a constructive algebra. Because the definitions are stated explicitly and the degeneracy is flagged, this does not undermine the correctness of the theorems; it only delimits their interpretive force. The reader's weakest_assumption identifies exactly this point, and I agree with the ACCEPT verdict.","tokens_in":30451,"tokens_out":14696,"duration_ms":152640,"concrete_test":"Analytic check: add a minimal non-degeneracy axiom to Definition 5.1 (e.g., require M_n>0 on {T>n}, or require M to have finite range larger than {0,b}) and re-run Theorem 5.8 on Example 4.5 (T=1 w.p. 1/2, T=7 w.p. 1/2, b=4). If no such weak e-detector exists, the universality theorem is an artifact of admitting zero-before-alarm indicator certificates, confirming the caveat; if one exists, the caveat is weaker than thought.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The converse directions of Theorems 5.8 and 8.3 are proved by the canonical process M_n = b 1{T≤n}. For weak detectors this witness verifies Definition 5.1 by essentially repackaging the target inequality: for 0<c≤b, T_c(M)=T and E[M†_{T_c}] = b P(T<∞) ≤ b ≤ E[T] = E[T_c] when E[T] is finite, while for c>b there is no crossing. Thus the only existence claim is that the path that records T is an admissible detector. If the weak definition were strengthened to require positive pre-alarm values, graded evidence, or bounded jumps, the early-alarm lottery E[T]=4, P(T=1)=1/2 would not be representable at b=4, and most ARL-controlled rules would fall outside the class. The paper is transparent about this in Sections 5 and 9.1, so it is a scope limitation rather than an internal inconsistency. The strong theorem has the same degenerate witness, but its hypothesis is already an optional-horizon property that is strictly stronger than ARL, so the circularity is less acute. No contradiction or unproved step in the core arguments was found.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":30718,"tokens_out":26246,"duration_ms":234550,"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.","major_comments":[],"minor_comments":[{"comment":"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.","section":"Title and Abstract"},{"comment":"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).","section":"Remark 8.12"},{"comment":"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.","section":"Appendix G"},{"comment":"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.","section":"Section 9.1"}],"recommendation":"accept","confidential_remarks":"I see no grounds for reject. The stress-test concern about the degenerate canonical witness is real but is explicitly disclosed by the authors and is a definitional scope caveat, not a technical defect. The paper is well within the scope of math.ST. The appendices are extensive and could plausibly be moved to supplemental material, but they are clearly marked as extensions and do not hinder the main narrative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper gives exact representation theorems that have been missing. ARL-controlled alarm times are shown to be exactly the level-crossings of weak e-detectors; optional-horizon-controlled alarm times are shown to be exactly the level-crossings of the original strong e-detectors. The clocked framing (C_n ≡ 1 gives e-processes, C_n = n gives e-detectors) is a genuinely useful unification, and the truncated-mean / NBUE characterization of strong run-length laws under the minimal alarm filtration is a clean standalone result. The Gaussian fast-start example is a nice touch: it turns a definitional distinction into a concrete power/delay tradeoff.\n\nWhat is solid: the core proofs are short, self-contained, and handle infinite horizons carefully. The examples are checkable. The citation pattern is appropriate; the author leans on his own monograph for e-process representation, but that is exactly where the relevant result lives. I checked the main theorems for hidden circularity and found none beyond what the paper itself concedes.\n\nThe soft spot is exactly where the stress-test note points. The converse of weak universality (Theorem 5.8) is proved by the canonical all-or-nothing process b 1{T≤n}. That makes the weak theorem close to a restatement of ARL control: the object certifies T by essentially recording T. The paper says this plainly in Sections 5 and 9.1, so it is a scope limitation, not a hidden flaw. The same degenerate witness also proves the strong converse in Theorem 8.3, but there the hypothesis is strong enough that the result is substantive. I would have liked the author to state even more forcefully that weak universality is a completeness result for a language, not a construction toolkit; he effectively does, given the nonconvexity result in Proposition 5.4. The appendices are less important; some of them (rolling e-fields) are explicitly self-referential and read as taxonomical filler rather than results anyone will use. The Remark 8.12 concern in the reader's report looks like a false alarm to me—the identity M_tau = 2 1{tau=2} is correct under the paper's killed/liminf convention.\n\nVerdict: this deserves a serious referee and, in my view, acceptance after minor revision. Anyone working on e-detectors, anytime-valid inference, or quickest detection will want the strong universality theorem and the run-length law characterization. I would cite it and bring it to the reading group.","headline":"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.","tokens_in":31237,"tokens_out":2420,"would_cite":true,"duration_ms":24988,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L10","62L15"],"pacs":[],"model":"deepseek-v4-flash","headline":"E-detectors exactly capture ARL-safe alarm times, the paper proves.","keywords":["e-detector","average run length","change detection","false alarm","optional stopping","e-process","universality","level-crossing time"],"falsifier":"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.","tokens_in":30254,"feed_emoji":"🚨","tokens_out":5285,"duration_ms":45396,"temperature":0.7,"pith_summary":"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.","feed_headline":"E-detectors exactly capture ARL-safe alarm times","feed_subtitle":"Any alarm rule with average run length at least b is a threshold crossing of an e-detector; strong rules also bound every adaptive horizon.","key_machinery":"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 $M_n-C_n$.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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 $\\gamma/\\rho+1$."],"supporting_citations":[{"why":"Introduces e-detectors and proves that thresholding them controls ARL; this is the object whose universality is at issue.","marker":"Shin et al. (2024)"},{"why":"Establishes e-value and e-process representation theorems that the paper extends to ARL-style metrics.","marker":"Ramdas and Wang (2025)"},{"why":"Develops the Snell-envelope and martingale-domination tools used to certify strong e-detectors.","marker":"Ramdas et al. (2022)"},{"why":"Documents that a large ARL can conceal high early false-alarm risk, motivating the optional-horizon inequality.","marker":"Mei (2008)"},{"why":"Provides fast-initial-response CUSUM ideas that the Gaussian fast-start example is compared against.","marker":"Lucas and Crosier (1982)"},{"why":"Shows that composite-null e-processes need not have a single common supermartingale, paralleling the pointwise domination discussion here.","marker":"Ruf et al. (2023)"},{"why":"Originates the martingale-based viewpoint from which e-processes descend.","marker":"Ville (1939)"}],"fun_headline_variants":["E-detectors: exact universal rules for ARL control","All ARL-safe alarm times are e-detector thresholds","E-detectors fully characterize adaptive false-alarm bounds","Strong e-detectors exactly match optional-horizon safety","Universality proven: e-detectors capture all safe alarms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["E-detectors: exact universal rules for ARL control","All ARL-safe alarm times are e-detector thresholds","E-detectors fully characterize adaptive false-alarm bounds","Strong e-detectors exactly match optional-horizon safety","Universality proven: e-detectors capture all safe alarms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2967,"prompt_tokens":1085,"completion_tokens":1882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":1802}},"tokens_in":701,"tokens_out":1882,"duration_ms":10613,"temperature":1.0,"reasoning_tokens":1802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:03:55.504188+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}