Pith. sign in

REVIEW 5 minor 41 references

Dynamic $e$-closure for online hypotheses with any-time-valid evidence: closure principles and projective mergers

T0 review · 0 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A coherence condition called future-extension coherence lets dynamic e-closure certify rejection sets that remain valid at arbitrary global stopping times, even as hypotheses arrive and evidence evolves.

desk verdict A genuine unification of fixed-time and online e-closure with a strong merger rigidity theorem, hampered only by the explicitly stated deterministic-arrival scope condition. read the letter →

arxiv 2608.09927 v1 pith:W4SNZVPR submitted 2026-08-10 math.ST stat.MEstat.TH

Rianne de Heide This is my paper · ORCID
classification math.STstat.MEstat.TH MSC 62L1062J1562F03
keywords dynamice-closureonlinehypothesesany-time-validevidencestoppedFDRSupe-processmergersfuture-extensioncoherenceneutral-paddingprojectivity
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 develops closure for multiple testing when both axes are sequential: new hypotheses arrive over time, and evidence for active hypotheses keeps updating and may be inspected at arbitrary stopping times. It proves that a single condition—future-extension coherence, meaning adding a dormant hypothesis never lowers the current intersection certificate—makes the closed family of certified rejection sets control the false discovery rate at every global stopping time, and additionally control the supremum over time when certificates are time-monotone. The result is sharp: every procedure that satisfies the error criterion is contained in a closure generated by its own normalized loss certificates, so closure is necessary and sufficient up to containment. The same logic extends from FDR to bounded losses that are monotone in the true-null configuration and local in the reported action.

What carries the argument

The engine is the random-intersection localization lemma. It compares the certificate of the random active true-null intersection at a stopping time with one fixed terminal-intersection e-process, using the fact that the active true-null set at a bounded stopping time is the terminal true-null set intersected with the active set; future-extension coherence then bounds the random-index certificate by the fixed-index process, to which optional stopping applies. The converse direction uses the canonical current-loss and running-loss processes, normalized by alpha, which are the pointwise smallest coherent certificates covering a given procedure.

What would settle it

A concrete counterexample would be a future-extension coherent dynamic intersection e-process collection, with deterministic finite active sets and all processes valid in one global filtration, whose dynamic closure at an almost surely finite global stopping time has expected FDP greater than alpha; such an example would refute Theorem 3.6, while Proposition 3.2 already shows that without coherence the failure is real.

Watch

Extended reading notes

Core claim

The central discovery is a dynamic closure principle for simultaneous stopped FDR and SupFDR. For a future-extension coherent dynamic intersection e-process collection, the candidate family of rejection sets defined by the closure constraints controls simultaneous stopped FDR at level alpha (Theorem 3.6); with time-monotone certificates it controls simultaneous SupFDR and the family is setwise persistent (Theorem 3.7). Conversely, any candidate-family process satisfying either guarantee is contained in a closure generated by canonical current-loss or running-loss e-processes (Theorems 3.10, 3.11). On the construction side, coherent pointwise mergers of arbitrary e-processes are exactly the affine rules with a single globally summable weight sequence, so arithmetic-mean mergers that renormalize per horizon break coherence (Theorem 4.8).

Load-bearing premise

The whole theory assumes the active hypothesis set at every time is deterministic and finite, and that every coordinate or intersection e-process is valid in one common global filtration; if hypothesis arrivals are random or processes are valid only in local filtrations, the localization step that converts the random true-null intersection into a fixed terminal intersection fails.

Editorial extensions

If this is right

  • Any procedure that controls simultaneous stopped FDR can be represented, up to containment, as a dynamic e-closure from coherent intersection evidence, so the closure framework is complete for this guarantee.
  • If the intersection certificates are time-monotone, the same closure is setwise persistent and controls simultaneous SupFDR, making previously certified rejection sets remain certified.
  • On an open-ended hypothesis universe, admissible pointwise mergers under arbitrary dependence must be affine with one globally summable weight budget; renormalizing weights when new hypotheses arrive is impossible within this class.
  • On a countably infinite universe, no nontrivial symmetric admissible pointwise merger family is coherent under neutral padding; only the trivial merger remains.
  • The results are not specific to the false discovery proportion: bounded configuration-monotone local losses inherit the same closure, representation, and minimality results.

Reading between the lines

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

  • The localization premise (deterministic finite active sets, single global filtration) is where the theory is most exposed; random predictable arrivals likely require a random-terminal-set version of the localization lemma, which the paper explicitly leaves open.
  • A practical consequence the paper states but does not emphasize: designs should commit to a global weight budget before the hypothesis family is known, since per-horizon renormalization is incoherent; platform-trial protocols can test this by comparing pre-specified versus renormalized weights in simulation.
  • The rigidity result suggests that dependence structure between hypotheses is the natural route to richer coherent mergers; under arbitrary dependence the admissible class is exhausted, but under known correlation structure (like shared-control arms) non-affine coherent mergers may exist.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 5 minor

Summary. Many modern testing problems are sequential along two axes: new hypotheses may arrive over time, and evidence for existing hypotheses continues to evolve and may be inspected at arbitrary global stopping times. The paper develops dynamic e-closure for this setting under explicit assumptions of deterministic finite active sets I_t (Assumption (2.3)) and validity of all coordinate and intersection processes in a single global filtration.

Significance. Within its stated scope, this is a substantial contribution to e-process multiple testing. I spot-checked the localization lemma, the necessity counterexample, the rigidity coefficient argument, and the stopped-BH example and found no internal errors; the appendix contains complete proofs of all main results. The necessary-and-sufficient representation up to containment cleanly extends fixed-time e-closure to the joint online/any-time setting, and Theorem 4.8 sharply delimits what admissible pointwise arbitrary-dependence mergers can achieve under horizon coherence. The paper is well documented for a theory paper: complete proofs, constructive counterexamples, and explicit limitation statements in Section 8.1. The main caveat on significance is that the headline stopped-FDR and SupFDR guarantees require deterministic finite arrival schedules and a single global filtration; the platform-trial motivation in Section 1.1 involves random arm additions, deferred to future work in Section 8.2. This is a disclosed scope restriction rather than an internal inconsistency, but the abstract should carry the restriction explicitly.

minor comments (5)
  1. [Abstract; §1.1; §2.1; §8.1] The forward theorems are explicitly conditional on deterministic finite active sets (Assumption (2.3)) and on global-filtration validity, and Section 8.1 discloses this; nevertheless, the abstract's opening sentence and the platform-trial motivation in Section 1.1 invite the reader to expect coverage of random hypothesis arrivals, a setting in which Lemma 3.4's fixed-terminal-intersection comparison has no immediate substitute. The stress-test concern about random predictable arrivals therefore lands as a real but disclosed scope limitation; I recommend stating the deterministic-arrival restriction explicitly in the abstract and in Section 1.1.
  2. [Remark 3.8] The reduction claim that the online SupFDR e-closure theorem of Xu et al. [2026] is recovered as a one-axis special case of Theorem 3.7 is asserted without verifying that the reduced process (E-hat)^S_t := E^{S ∩ [t]} is an e-process in the global filtration when the index S ∩ [τ] is random at a stopping time τ. The claim is correct: for bounded τ ≤ T, increasingness of the online e-collection gives E^{S ∩ [τ]} ≤ E^S pointwise, so E[(E-hat)^S_{τ∧T}] ≤ 1, and Fatou extends this to unbounded almost surely finite τ; adding this one-line justification, or a pointer to the corresponding property in Xu et al. [2026], would make the remark self-contained.
  3. [§6.4] In the stopped-BH counterexample, the step-up rule k* = max{k : p_(k) ≤ kα/m} is what makes BH reject both hypotheses at time three when (P_{1,3}, P_{2,3}) = (α, α), even though p_(1) = α > α/2; stating this rule explicitly would prevent readers from concluding that the individual threshold p_(1) ≤ α/2 must also hold.
  4. [Abstract] The abstract as rendered contains spacing and hyphenation artifacts ('Dynamice-closure', 'forpointwisemergers', 'ordinaryarbitrarydependencee-merging') that should be cleaned before submission; if these artifacts come from the extraction rather than the source, the authors should confirm that the source compiles cleanly.
  5. [AI-assisted editing statement] In the AI-assisted editing statement, 'the LLMh' appears to be a typo for 'the LLM'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the forward theorems are direct consequences of explicit premises, and the converse results are formal representation theorems rather than fitted predictions.

full rationale

The derivation chain is self-contained. Theorem 3.6 follows from Lemma 3.4, which bounds the random active true-null certificate by the fixed terminal-intersection process E^{N_T(P)} and then applies optional stopping to that fixed e-process; the argument is a proof from stated assumptions, not a restatement of the conclusion. Proposition 3.2 shows that fixed-intersection validity alone is insufficient without future-extension coherence, so the coherence condition is an explicit premise rather than a disguised version of the target result. The converse theorems (3.10, 3.11, 7.1, 7.2) construct canonical certificates from the procedure's own loss profile and prove in the appendix that these certificates are future-extension-coherent e-processes under the assumed stopped-FDR or SupFDR bound; this is a standard necessity-and-representation argument, not a fitted input being renamed as a prediction. The pointwise merger results reduce to Wang's fixed-dimensional characterization through Theorem 4.2, whose proof embeds arbitrary e-values into constant processes and is given in full; the cross-horizon rigidity theorem then derives the global weight sequence by an explicit coefficient-comparison argument. Citations to Xu et al. (2025) and Tavyrikov et al. (2026), which include the author, are used for context, terminology, and as starting points, with the essential proofs supplied in this paper rather than imported as unverified self-citations. Section 8.1 explicitly limits the theory to deterministic finite active sets and global-filtration validity; this narrows the scope of the stopped-FDR guarantee but does not make it circular, since it identifies a premise rather than an equivalent conclusion. No fitted parameter is relabeled as a prediction, and no uniqueness claim is imported solely from the author's prior work. Score 0.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The central claims depend on the e-process framework and on Wang's external characterization of admissible e-merging functions. No free parameters are fitted and no new entities are introduced; the remaining axioms are domain conditions that define the setting (deterministic active sets, global-filtration validity, future-extension coherence, neutral padding by one).

assumptions (6)
  • domain assumption Active sets I_t are deterministic, finite, nondecreasing, and exhaust I.
    Intro Section 2.1: I_0 subset I_1 subset ..., |I_t| finite, union I_t = I. Lemma 3.4 uses deterministic I_t; with random arrivals, N_T(P) is no longer a fixed set.
  • domain assumption All coordinate and intersection e-processes are valid in the common global filtration (F_t) used for stopping.
    Section 2.1 and Wang et al. 2025; Section 8.1 Limitations explicitly notes this is essential for the stopping-time guarantees.
  • domain assumption Future-extension coherence: E^(S cap I_t)_t <= E^S_t for all S with nonempty intersection with I_t.
    Definition 2.5; this is the key condition making the forward theorem possible, and Proposition 3.2 shows necessity.
  • domain assumption Dormant coordinates have neutral e-value one: M_{i,t} = 1 before activation.
    Section 4.2, Equation (4.3); the rigidity Theorem 4.8 depends on padding by the value one.
  • standard math Wang's fixed-dimensional characterization of admissible arbitrary-dependence e-merging functions (affine form).
    Used in Theorems 4.2 and 4.8 as an external proved theorem (Wang 2025, with geometric proof by Clerico 2026).
  • standard math Measurability and regularity conditions, with merger domain [0, infinity) so the x to infinity and x = 0 arguments apply.
    Section 2.2 and Appendix A.10; needed for the coefficient-comparison argument in Theorem 4.8.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic $e$-closure for online hypotheses with any-time-valid evidence: closure principles and projective mergers." pith.science (2026). https://pith.science/paper/W4SNZVPR

@misc{pith2026260809927,
  author       = {Pith},
  title        = {Pith review of: Dynamic $e$-closure for online hypotheses with any-time-valid evidence: closure principles and projective mergers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W4SNZVPR}},
  note         = {Machine review of arXiv:2608.09927}
}
abstract

Many modern testing problems are sequential along two axes: new hypotheses may arrive over time, while evidence for hypotheses already under consideration continues to evolve and may be inspected at arbitrary stopping times. We develop dynamic $e$-closure for this setting. At a global stopping time the active true-null intersection is random. Future-extension coherence allows its certificate to be compared with that of a fixed terminal intersection, yielding simultaneous stopped-FDR control. If the certificates are also time-monotone, the resulting closure controls simultaneous SupFDR and is setwise persistent. Conversely, every procedure satisfying either criterion is contained in a dynamic closure generated by canonical normalized-loss processes. For pointwise mergers, fixed-dimensional admissibility is equivalent to ordinary arbitrary-dependence $e$-merging. Coherence across horizons then forces a single globally summable weight sequence and exact neutrality under padding by the $e$-value one; on a countably infinite hypothesis universe, this rules out nontrivial symmetric mergers in the admissible pointwise class. The theory extends from FDP to bounded losses that are monotone in the possible true-null configuration and local in the reported action. We also give a coherence counterexample, persistent constructions, and a globally valid shared-control model.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 25 canonical work pages

  1. [1]

    2006 , publisher=

    Prediction, learning, and games , author=. 2006 , publisher=

    2006

  2. [2]

    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=

    work page 2026

  3. [3]

    The Annals of Statistics , volume=

    The online closure principle , author=. The Annals of Statistics , volume=. 2024 , publisher=

    2024

  4. [4]

    arXiv preprint arXiv:2407.15733 , year=

    Admissible online closed testing must employ e-values , author=. arXiv preprint arXiv:2407.15733 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2407.20683 , year=

    An online generalization of the (e-) Benjamini-Hochberg procedure , author=. arXiv preprint arXiv:2407.20683 , year=

    arXiv

  6. [6]

    A Uniform Improvement of the Benjamini-Hochberg Procedure via e-Closure

    A Uniform Improvement of the Benjamini-Hochberg Procedure via e-Closure , author=. arXiv preprint arXiv:2606.01854 , year=

    work page Pith review arXiv

  7. [7]

    The Annals of Statistics , volume=

    Only closed testing procedures are admissible for controlling false discovery proportions , author=. The Annals of Statistics , volume=. 2021 , publisher=

    work page 2021

  8. [8]

    Statistical Science , pages=

    Multiple testing for exploratory research , author=. Statistical Science , pages=. 2011 , publisher=

    work page 2011

Show all 41 references
  1. [9]

    The Annals of Statistics , volume=

    Time-uniform, nonparametric, nonasymptotic confidence sequences , author=. The Annals of Statistics , volume=. 2021 , publisher=

    2021

  2. [10]

    Cambridge University , volume=

    Bandit algorithms , author=. Cambridge University , volume=

  3. [11]

    Trials , volume=

    Statistical consideration when adding new arms to ongoing clinical trials: the potentials and the caveats , author=. Trials , volume=. 2021 , publisher=

    2021

  4. [12]

    Biometrika , volume=

    On closed testing procedures with special reference to ordered analysis of variance , author=. Biometrika , volume=. 1976 , publisher=

    1976

  5. [13]

    Statistical Science , volume=

    Game-theoretic statistics and safe anytime-valid inference , author=. Statistical Science , volume=. 2023 , publisher=

    2023

  6. [14]

    arXiv preprint arXiv:2009.03167 , year=

    Admissible anytime-valid sequential inference must rely on nonnegative martingales , author=. arXiv preprint arXiv:2009.03167 , year=

    2009 arXiv

  7. [15]

    Statistical science: a review journal of the Institute of Mathematical Statistics , volume=

    Online multiple hypothesis testing , author=. Statistical science: a review journal of the Institute of Mathematical Statistics , volume=

  8. [16]

    The Review of Symbolic Logic , volume=

    On the truth-convergence of open-minded Bayesianism , author=. The Review of Symbolic Logic , volume=. 2022 , publisher=

    2022

  9. [17]

    arXiv preprint arXiv:2607.14380 , year=

    Admissibility and Complete Classes for False Discovery Rate Control with E-values , author=. arXiv preprint arXiv:2607.14380 , year=

    arXiv

  10. [18]

    Electronic Journal of Statistics , volume=

    Carefree multiple testing with e-processes , author=. Electronic Journal of Statistics , volume=. 2026 , publisher=

    2026

  11. [19]

    The Annals of Statistics , volume=

    E-values: Calibration, combination and applications , author=. The Annals of Statistics , volume=. 2021 , publisher=

    2021

  12. [20]

    Electronic Journal of Statistics , volume=

    Merging sequential e-values via martingales , author=. Electronic Journal of Statistics , volume=. 2024 , publisher=

    2024

  13. [21]

    Biometrika , volume=

    The only admissible way of merging arbitrary e-values , author=. Biometrika , volume=. 2025 , publisher=

    2025

  14. [22]

    Statistics & Probability Letters , pages=

    Anytime-valid FDR control with the stopped e-BH procedure , author=. Statistics & Probability Letters , pages=. 2025 , publisher=

    2025

  15. [23]

    arXiv preprint arXiv:2509.02517 , year=

    Bringing closure to false discovery rate control: A general principle for multiple testing , author=. arXiv preprint arXiv:2509.02517 , year=

    arXiv

  16. [24]

    arXiv preprint arXiv:2603.24792 , year=

    Improving online FDR procedures via online analogs of e-closure and compound e-values , author=. arXiv preprint arXiv:2603.24792 , year=

    arXiv

  17. [25]

    Journal of the Royal statistical society: series B (Methodological) , volume=

    Controlling the false discovery rate: a practical and powerful approach to multiple testing , author=. Journal of the Royal statistical society: series B (Methodological) , volume=. 1995 , publisher=

    1995

  18. [26]

    Annals of statistics , pages=

    The control of the false discovery rate in multiple testing under dependency , author=. Annals of statistics , pages=. 2001 , publisher=

    2001

  19. [27]

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

    Safe testing , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2024 , publisher=

    2024

  20. [28]

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

    False discovery rate control with e-values , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2022 , publisher=

    2022

  21. [29]

    Statistics & Probability Letters , volume=

    Insuring against loss of evidence in game-theoretic probability , author=. Statistics & Probability Letters , volume=. 2011 , publisher=

    2011

  22. [30]

    Statistical Science , pages=

    Test martingales, Bayes factors and p-values , author=. Statistical Science , pages=. 2011 , publisher=

    2011

  23. [31]

    Foundations and Trends

    Hypothesis testing with e-values , author=. Foundations and Trends. 2025 , publisher=

    2025

  24. [32]

    Statistics & Probability Letters , pages=

    A simple geometric proof for the characterisation of e-merging functions , author=. Statistics & Probability Letters , pages=. 2026 , publisher=

    2026

  25. [33]

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

    -investing: a procedure for sequential control of expected false discoveries , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2008 , publisher=

    2008

  26. [34]

    The Annals of statistics , volume=

    Online rules for control of false discovery rate and false discovery exceedance , author=. The Annals of statistics , volume=. 2018 , publisher=

    2018

  27. [35]

    Ruben , title =

    Gabriel, K. Ruben , title =. Annals of Mathematical Statistics , year =

  28. [36]

    and Singer, Yoram and Warmuth, Manfred K

    Freund, Yoav and Schapire, Robert E. and Singer, Yoram and Warmuth, Manfred K. , title =. Proceedings of the Twenty-Ninth Annual ACM Symposium on Theory of Computing , year =

  29. [37]

    and Niculescu-Mizil, Alexandru and Sharma, Yogeshwer , title =

    Kleinberg, Robert D. and Niculescu-Mizil, Alexandru and Sharma, Yogeshwer , title =. Mach. Learn. , year =

  30. [38]

    and van Erven, Tim , title =

    Koolen, Wouter M. and van Erven, Tim , title =. 2010 , howpublished =

    2010

  31. [39]

    Mathematical and Scientific Machine Learning , series =

    Xu, Ziyu and Ramdas, Aaditya , title =. Mathematical and Scientific Machine Learning , series =. 2022 , publisher =

    2022

  32. [40]

    1939 , publisher =

    Ville, Jean , title =. 1939 , publisher =

    1939

  33. [41]

    Simes, R. J. , title =. Biometrika , year =

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.