Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Global Sequential Testing for Multi-Stream Auditing

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper constructs a balanced sequential test for a global null across k data streams that provably matches the best expected stopping time in both sparse and dense alternative regimes.

desk verdict Useful and mostly careful paper, but the product and balanced tests need an unstated cross-stream independence assumption, and the dense-rate claim is the first thing to check in revision. read the letter →

arxiv 2602.21479 v3 pith:MDR65RNE submitted 2026-02-25 stat.ML cs.LG

Beepul Bharti , Ambar Pal , Jeremias Sulam This is my paper
classification stat.MLcs.LG
keywords sequentialhypothesistestinganytime-validinferencetestmartingalese-valuesmultiplemachinelearningauditingstoppingtimemergingstrategies
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

The paper studies how to continuously audit a machine learning system across k parallel data streams, where the global null says every stream has mean zero. Standard Bonferroni testing has an expected stopping time of O(ln(k/alpha)), which is unnecessarily slow when many streams are failing. The authors show that merging per-stream betting wealth by product or by average gives complementary strengths: product is fast for dense alternatives but slow for sparse ones, and average behaves the reverse. They then construct a balanced test that takes a convex combination of the two merged processes, achieving expected stopping time O(ln(k/alpha)) under sparse alternatives and O((1/k)ln(1/alpha)) under dense alternatives. This matters because it gives auditors a single, anytime-valid test that adapts to the unknown number of failing streams.

What carries the argument

The central object is the merged wealth process: per-stream betting wealth W_i,t computed via Online Newton Step, merged by averaging, by product, and by the equal-weight mixture W_balance = (1/2)W_ave + (1/2)W_prod. Both the average and product processes are P-martingales under the global null; the balanced mixture is their convex combination. Stopping-time bounds follow from log-wealth lower bounds derived from online-learning regret bounds, plus concentration inequalities (Hoeffding and McDiarmid) on sums of squared stream sums.

What would settle it

Construct k streams with a common latent factor that induces positive correlation among mean-zero observations; run the product and balanced tests at level alpha and measure the empirical probability of false rejection before any time T. If it exceeds alpha for any T, the martingale property fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the balanced wealth process W_balance = (1/2)W_ave + (1/2)W_prod, built from per-stream betting wealth processes, is a valid test martingale under the global null and its expected stopping time under the alternative satisfies E[tau] <= min{T_prod, T_bonf}. Consequently, when few streams are non-null it inherits the Bonferroni rate O(ln(k/alpha)), and when many streams are non-null it inherits the product rate O((1/k)ln(1/alpha)). The proof combines Ville's inequality with regret guarantees for online Newton step and follow-the-regularized-leader.

Load-bearing premise

The k data streams are mutually independent, an assumption needed for the product of per-stream test martingales to remain a martingale under the global null.

Editorial extensions

If this is right

  • Auditors can deploy one test without knowing whether a failure affects one stream or many; the test automatically adapts its expected detection time to the sparsity of the alternative.
  • In dense failure settings, detection time improves by a factor of k compared to Bonferroni, while in sparse settings it stays within a log factor of the best single-stream test.
  • The level-alpha guarantee holds at every time, so the test can be monitored continuously without adjusting for multiple looks.
  • The framework applies to bounded-mean auditing tasks, such as checking multiaccuracy or fairness across demographic groups or imaging modalities.

Reading between the lines

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

  • If streams are correlated, the product martingale property can break and the balanced test's validity may be lost; a natural extension is to design merging weights that are robust to dependence or to state mutual independence as an explicit assumption.
  • The same averaging-balancing idea could be applied to other base test martingales, not just ONS-based wealth, potentially yielding better constants or adapting to heavy-tailed payoffs.
  • The bound is on expected stopping time; risk-sensitive auditing may also want high-probability tail bounds, which the paper does not provide.
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

2 major / 5 minor

Summary. The paper studies sequential testing of a global null hypothesis over k parallel data streams with bounded observations. It analyzes several tests—Bonferroni, multivariate ONS, FTRL, and two merging strategies (product and average of stream-specific wealth processes)—and proposes a balanced test that combines the average and product processes. The central theoretical claim is that the product-based test achieves an expected stopping time of O((1/k) ln(1/α)) under dense alternatives, the average/Bonferroni approach achieves O(ln(k/α)) under sparse alternatives, and the balanced test attains the minimum of the two rates without knowing the sparsity regime. The claims are supported by proofs in Appendix A and by synthetic and medical-imaging experiments.

Significance. If the results hold as stated, the paper makes a useful contribution to multi-stream sequential testing and to martingale-based auditing. The proofs are detailed and include explicit constants, and the manuscript even identifies and corrects a constant in a cited ONS lemma, which is a useful service to the community. The balanced test's sparse/dense trade-off is a principled and practical idea, and the experiments qualitatively match the theory. The main caveat is that the central product-based claims currently rely on an unstated mutual-independence assumption across streams; this is fixable but must be addressed before the results can be accepted as stated.

major comments (2)
  1. [Section 2, Theorems 5.1 and 5.3, Appendix A.5] The problem formulation states that each stream is iid within the stream, but it never states that the k streams are mutually independent. The level-α proof for W_prod uses E[∏_i W_i,t | F_{t-1}] = ∏_i E[W_i,t | F_{i,t-1}], which requires conditional independence across i. Without it, for k=2 with Z_1,t = Z_2,t = X_t and X_t iid mean zero, E[(1+λ_1,t X_t)(1+λ_2,t X_t)|F_{t-1}] = 1+λ_1,t λ_2,t Var(X_t) > 1, so W_prod is a strict submartingale and Ville's inequality does not apply. W_balance inherits this defect because it places positive weight on W_prod. The same missing assumption also affects the vector tests in Theorems 4.1 and 4.2 when cross-stream dependence has memory, since their proofs need E[Z_t|F_{t-1}]=0. Please add mutual independence of the streams to Section 2, or state and prove a weaker condition that still guarantees the required supermartingale property, and adjust Theo
  2. [Appendix A.5, Eq. (99)] The McDiarmid concentration step treats the kt variables Z_1,1,...,Z_k,t as independent. This is used to lower-bound ψ = Σ_i A_i,t^2 and hence to obtain the T_prod stopping-time bound in Theorem 5.1; Theorem 5.2's T_1 and Theorem 5.3's balanced guarantee inherit this step. If streams are cross-correlated, even only contemporaneously, the kt coordinates are not independent and the stated high-probability lower bound on Σ_i A_i,t^2 is not established. This directly affects the dense-alternative claim O((1/k) ln(1/α)) and the balanced test's guarantee. Please either make independence across streams an explicit assumption or replace the McDiarmid argument with a concentration inequality that holds under the dependence structure actually assumed.
minor comments (5)
  1. [Abstract] Typo: 'steams' should be 'streams'.
  2. [A.2, Eq. (37)] The displayed line contains a stray '= 1' after k/α, making the equation read incorrectly.
  3. [Appendix A, definitions] The definition of an adapted process is nonstandard: adaptedness is measurability of M_t with respect to F_t, not E[M_t|Z_1,...,Z_t]=M_t. This should be corrected to avoid confusion.
  4. [A.7, Eq. (135)] The conditional expectation conditions on Z_1,1,...,Z_1,t,...,Z_k,1,...,Z_k,t, i.e. including time t; it should condition on F_{t-1}.
  5. [Theorem 5.3] The theorem statement says E[τ] ≤ min{T_prod, T_bonf} with T_prod and T_bonf as in Eqs. (12) and (15), but the proof in A.7 yields the same forms with 2/α and 2k/α in place of 1/α and k/α. This is asymptotically immaterial but the statement and proof should be aligned, e.g. by stating the bound up to universal constants.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claimed stopping-time guarantees are derived analytically from standard martingale and concentration tools, with no fitted parameters and no author self-citation chain.

full rationale

Walked the full derivation chain. The single-stream ONS test's level-α property and expected stopping time are proved in Section A.1 using Ville's inequality and Hoeffding concentration; the citation to Chugg et al. is not load-bearing because the proof is included. The Bonferroni bound in Proposition 4.1 is proved by union bound in A.2. The 'not better than Bonferroni' claims for MV-ONS and FTRL (Theorems 4.1 and 4.2) are proved from the stated ONS/FTRL regret bounds and concentration arguments. The product and average merging tests (Theorems 5.1 and 5.2) are derived independently: their level-α control follows from martingale properties and their stopping-time bounds follow from log-wealth lower bounds, McDiarmid/Hoeffding concentration, and Lemma B.6; no parameter is fitted to the predicted stopping times. The balanced test (Theorem 5.3) is obtained from the elementary observation that if (W_ave + W_prod)/2 < 1/α then W_prod < 2/α and W_ave < 2/α, so the stopping time is bounded by the product and Bonferroni bounds up to a factor of 2 in α; this is a derivation from the test's definition, not a restatement of the conclusion. The paper does not rely on any self-citation by its authors for load-bearing results, and the external cited results (Ville, Hoeffding, McDiarmid, Cutkosky-Orabona, Orabona) are standard and either stated with proof or referenced with proofs. The main caveat found in review — that the problem formulation only states each stream is i.i.d., not that streams are mutually independent, while the product/multivariate proofs require cross-stream independence — is a genuine assumption gap and a correctness risk, but it is not a circularity: the theorems do not assume the stopping-time bounds they purport to establish. Accordingly, the circularity score is 0.

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

The central claim rests on standard martingale/concentration tools plus a genuine but unstated independence assumption. There are no fitted parameters, no invented entities, and no self-citations in the load-bearing lemmas. The balance weight 1/2 is a fixed constant that does not affect the rates.

assumptions (5)
  • domain assumption The k streams are mutually independent (across i).
    Used in Theorem 5.1 proof (Section A.5, E[prod_i W_i,t | past] = prod_i E[W_i,t | ...]) and in McDiarmid's inequality application. Not stated in Section 2. Without it, the product test and balanced test may lose type-I error control.
  • domain assumption Each Z_i,t is in [-1,1] almost surely.
    Bounded mean testing assumption from Section 2, used for Hoeffding/McDiarmid and the ONS regret bounds.
  • domain assumption Under the global null, each stream has mean zero, E[Z_i,1]=0.
    The null hypothesis being tested (Eq. 2).
  • standard math ONS regret bound (Theorem B.2) and FTRL regret bound (Theorem B.3) hold as stated, with the paper's correction of the 2/beta typo in [9].
    The log-wealth lower bounds (Lemma B.4) rest on these; the paper itself flags and corrects a typo in the cited proof, so the constants are not independently verified here.
  • standard math Ville's inequality, Hoeffding's inequality, and McDiarmid's inequality.
    Standard concentration tools invoked in the stopping-time proofs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Global Sequential Testing for Multi-Stream Auditing." pith.science (2026). https://pith.science/paper/MDR65RNE

@misc{pith2026260221479,
  author       = {Pith},
  title        = {Pith review of: Global Sequential Testing for Multi-Stream Auditing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDR65RNE}},
  note         = {Machine review of arXiv:2602.21479}
}
abstract

Across many risk-sensitive areas, it is critical to continuously audit machine learning systems as we receive more data to quickly determine if they are performing as designed. This auditing task can be modeled as a sequential hypothesis testing problem with $k$ data streams and a global null hypothesis that asserts the system operates as intended across all $k$ streams. Under the alternative, the standard global sequential test, which uses a Bonferroni correction, has an expected stopping time of $O\left(\ln \frac{k}{\alpha}\right)$ for large $k$ and significance level $\alpha$. In this work, we demonstrate that efficient sequential tests, relying on merging martingales via averaging and products rules, provide improved stopping times, and thus more powerful tests against the null. Using these results, we show that a balanced test can match the Bonferroni rate of $O\left(\ln \frac{k}{\alpha}\right)$ in the sparse regime (just a few non-null streams) while achieving $O\left(\frac{1}{k}\ln \frac{1}{\alpha}\right)$ under dense alternatives (many non-null steams). We validate our theory through experiments on both synthetic and real-world data.

Figures

Figures reproduced from arXiv: 2602.21479 by the authors.

Figure 2
Figure 2. Top: Distribution of stopping times, over 1,000 simulations, for various sequential tests across settings with varying proportions of streams with nonzero means. A test rejects when its corresponding wealth process exceeds 1/α for α = 0.01. The dashed vertical line is the empirical mean of the stopping times. Bottom: Trajectories of various wealth processes across settings with different amounts of nonzero means. Ea… view at source ↗
Figure 3
Figure 3. Left plot of each figure: Distribution of stopping times, over 1,000 runs, for various sequential tests. A test rejects when its corresponding wealth process exceeds 1/α for α = 0.01. The dashed vertical line is the empirical mean of the stopping times. Right plot of each figure: Various wealth process trajectories. Each line represents the median trajectory of a wealth process over 1,000 runs, with shaded areas ind… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 2 linked inside Pith

  1. [1]

    Problematic machine behavior: A systematic literature review of algorithm audits

    Jack Bandy. “Problematic machine behavior: A systematic literature review of algorithm audits”. In:Proceedings of the acm on human-computer interaction5.CSCW1 (2021), pp. 1– 34

    2021

  2. [2]

    From transparency to accountability and back: A discussion of access and evidence in ai auditing

    Sarah H Cen and Rohan Alur. “From transparency to accountability and back: A discussion of access and evidence in ai auditing”. In:Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. 2024, pp. 1–14

    2024

  3. [3]

    Online Detection of LLM-Generated Texts via Sequential Hy- pothesis Testing by Betting

    Can Chen and Jun-Kun Wang. “Online Detection of LLM-Generated Texts via Sequential Hy- pothesis Testing by Betting”. In:Forty-second International Conference on Machine Learning. 2025.url:https://openreview.net/forum?id=khFk7sdv9o

    2025

  4. [4]

    Peeking with PEAK: sequential, nonparametric composite hypothesis tests for means of multiple data streams

    Brian Cho, Kyra Gan, and Nathan Kallus. “Peeking with PEAK: sequential, nonparametric composite hypothesis tests for means of multiple data streams”. In:Proceedings of the 41st International Conference on Machine Learning. 2024, pp. 8487–8509

    2024

  5. [5]

    Auditingfairnessbybetting

    BenChuggetal.“Auditingfairnessbybetting”.In:Advances in Neural Information Processing Systems36 (2023), pp. 6070–6091

    2023

  6. [6]

    European Union, 2021

    European Commission.Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act). European Union, 2021

    2021

  7. [7]

    Universal portfolios

    Thomas M Cover. “Universal portfolios”. In:Mathematical finance1.1 (1991), pp. 1–29

    1991

  8. [8]

    20 nonparametric methods for changepoint problems

    Miklós Csörgő and Lajos Horváth. “20 nonparametric methods for changepoint problems”. In: Handbook of statistics7 (1988), pp. 403–425

    1988

Show all 45 references
  1. [9]

    Black-box reductions for parameter-free online learning in banach spaces

    Ashok Cutkosky and Francesco Orabona. “Black-box reductions for parameter-free online learning in banach spaces”. In:Conference On Learning Theory. PMLR. 2018, pp. 1493–1529

    2018

  2. [10]

    From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms

    Jessica Dai et al. “From Individual Experience to Collective Evidence: A Reporting-Based Framework for Identifying Systemic Harms”. In:Forty-second International Conference on Machine Learning. 2025

    2025

  3. [11]

    Amazon scraps secret AI recruiting tool that showed bias against women

    Jeffrey Dastin. “Amazon scraps secret AI recruiting tool that showed bias against women”. In:Ethics of data and analytics. Auerbach Publications, 2022, pp. 296–299

    2022

  4. [12]

    Logarithmic regret algorithms for online convex optimization

    Elad Hazan, Amit Agarwal, and Satyen Kale. “Logarithmic regret algorithms for online convex optimization”. In:Machine Learning69.2 (2007), pp. 169–192

    2007

  5. [13]

    Multicalibration:Calibrationforthe(computationally-identifia- ble) masses

    UrsulaHébert-Johnsonetal.“Multicalibration:Calibrationforthe(computationally-identifia- ble) masses”. In:International Conference on Machine Learning. PMLR. 2018, pp. 1939–1948

    2018

  6. [14]

    Probability inequalities for sums of bounded random variables

    Wassily Hoeffding. “Probability inequalities for sums of bounded random variables”. In:Jour- nal of the American statistical association58.301 (1963), pp. 13–30. 14

    1963

  7. [15]

    Online rules for control of false discovery rate and false discovery exceedance

    Adel Javanmard and Andrea Montanari. “Online rules for control of false discovery rate and false discovery exceedance”. In:The Annals of statistics46.2 (2018), pp. 526–554

    2018

  8. [16]

    Mixture martingales revisited with applications to sequential tests and confidence intervals

    Emilie Kaufmann and Wouter M Koolen. “Mixture martingales revisited with applications to sequential tests and confidence intervals”. In:Journal of Machine Learning Research22.246 (2021), pp. 1–44

    2021

  9. [17]

    Multiaccuracy:Black-boxpost-processing for fairness in classification

    MichaelPKim,AmirataGhorbani,andJamesZou.“Multiaccuracy:Black-boxpost-processing for fairness in classification”. In:Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. 2019, pp. 247–254

    2019

  10. [18]

    On the method of bounded differences

    Colin McDiarmid et al. “On the method of bounded differences”. In:Surveys in combinatorics 141.1 (1989), pp. 148–188

    1989

  11. [19]

    Auditing algorithms: Understanding algorithmic systems from the out- side in

    Danaë Metaxa et al. “Auditing algorithms: Understanding algorithmic systems from the out- side in”. In:Foundations and Trends®in Human–Computer Interaction14.4 (2021), pp. 272– 344

    2021

  12. [20]

    IX. On the problem of the most efficient tests of statistical hypotheses

    Jerzy Neyman and Egon Sharpe Pearson. “IX. On the problem of the most efficient tests of statistical hypotheses”. In:Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of a Mathematical or Physical Character231.694-706 (1933), pp. 289– 337

    1933

  13. [21]

    Conceptclip: Towards trustworthy medical ai via concept-enhanced con- trastive langauge-image pre-training

    Yuxiang Nie et al. “Conceptclip: Towards trustworthy medical ai via concept-enhanced con- trastive langauge-image pre-training”. In:arXiv e-prints(2025), arXiv–2501

    2025

  14. [22]

    Dissecting racial bias in an algorithm that guides health decisions for 70 million people

    Ziad Obermeyer and Sendhil Mullainathan. “Dissecting racial bias in an algorithm that guides health decisions for 70 million people”. In:Proceedings of the conference on fairness, account- ability, and transparency. 2019, pp. 89–89

    2019

  15. [23]

    Francesco Orabona.A Modern Introduction to Online Learning. 2025. arXiv:1912 . 13213 [cs.LG].url:https://arxiv.org/abs/1912.13213

    2025 arXiv

  16. [24]

    Tight concentrations and confidence sequences from the regret of universal portfolio

    Francesco Orabona and Kwang-Sung Jun. “Tight concentrations and confidence sequences from the regret of universal portfolio”. In:IEEE Transactions on Information Theory70.1 (2023), pp. 436–455

    2023

  17. [25]

    Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial ai products

    Inioluwa Deborah Raji and Joy Buolamwini. “Actionable auditing: Investigating the impact of publicly naming biased performance results of commercial ai products”. In:Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. 2019, pp. 429–435

    2019

  18. [26]

    Hypothesis Testing with E-values

    Aaditya Ramdas and Ruodu Wang. “Hypothesis Testing with E-values”. In:Foundations and Trends®in Statistics1.1–2 (July 2025), pp. 1–390.issn: 2978-4220.doi:10 . 1561 / 3600000002.url:http://dx.doi.org/10.1561/3600000002

    2025 doi

  19. [27]

    Game-theoretic statistics and safe anytime-valid inference

    Aaditya Ramdas et al. “Game-theoretic statistics and safe anytime-valid inference”. In:Sta- tistical Science38.4 (2023), pp. 576–601

    2023

  20. [28]

    An Auditing Test to Detect Behavioral Shift in Language Models

    Leo Richter et al. “An Auditing Test to Detect Behavioral Shift in Language Models”. In: The Thirteenth International Conference on Learning Representations. 2025.url:https : //openreview.net/forum?id=h0jdAboh0o

    2025

  21. [29]

    The age of secrecy and unfairness in re- cidivism prediction

    Cynthia Rudin, Caroline Wang, and Beau Coker. “The age of secrecy and unfairness in re- cidivism prediction”. In:Harvard Data Science Review2.1 (2020), p. 1

    2020

  22. [30]

    On confidence sequences for bounded random processes via universal gambling strategies

    J Jon Ryu and Alankrita Bhatt. “On confidence sequences for bounded random processes via universal gambling strategies”. In:IEEE Transactions on Information Theory(2024)

    2024

  23. [31]

    Testing by betting: A strategy for statistical and scientific communication

    Glenn Shafer. “Testing by betting: A strategy for statistical and scientific communication”. In: Journal of the Royal Statistical Society Series A: Statistics in Society184.2 (2021), pp. 407– 431. 15

    2021

  24. [32]

    Machine learning in healthcare: A re- view

    K Shailaja, Banoth Seetharamulu, and MA Jabbar. “Machine learning in healthcare: A re- view”. In:2018 Second international conference on electronics, communication and aerospace technology (ICECA). IEEE. 2018, pp. 910–914

    2018

  25. [33]

    Nonparametric two-sample testing by betting

    Shubhanshu Shekhar and Aaditya Ramdas. “Nonparametric two-sample testing by betting”. In:IEEE Transactions on Information Theory70.2 (2023), pp. 1178–1203

    2023

  26. [34]

    Zillow’s artificial intelligence failure and its impact on perceived trust in information systems

    Padma Susarla, Dexter Purnell, and Ken Scott. “Zillow’s artificial intelligence failure and its impact on perceived trust in information systems”. In:Journal of Information Technology Teaching Cases(2024), p. 20438869241279865

    2024

  27. [35]

    UNESCO.Recommendation on the Ethics of Artificial Intelligence. 2021

    2021

  28. [36]

    Algorithmic auditing and social justice: Lessons from the history of audit studies

    Briana Vecchione, Karen Levy, and Solon Barocas. “Algorithmic auditing and social justice: Lessons from the history of audit studies”. In:Proceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. 2021, pp. 1–9

    2021

  29. [37]

    Jean Ville.Etude critique de la notion de collectif. Vol. 3. Gauthier-Villars Paris, 1939

    1939

  30. [38]

    E-values: Calibration, combination and applications

    Vladimir Vovk and Ruodu Wang. “E-values: Calibration, combination and applications”. In: The Annals of Statistics49.3 (2021), pp. 1736–1754

    2021

  31. [39]

    The only admissible way of merging arbitrary e-values

    Ruodu Wang. “The only admissible way of merging arbitrary e-values”. In:Biometrika112.2 (2025), asaf020

    2025

  32. [40]

    Estimating means of bounded random variables by betting

    Ian Waudby-Smith and Aaditya Ramdas. “Estimating means of bounded random variables by betting”. In:Journal of the Royal Statistical Society Series B: Statistical Methodology86.1 (2024), pp. 1–27

    2024

  33. [41]

    Jordan.Universal Log-Optimality for General Classes of e-processes and Sequential Hypothesis Tests

    Ian Waudby-Smith, Ricardo Sandoval, and Michael I. Jordan.Universal Log-Optimality for General Classes of e-processes and Sequential Hypothesis Tests. 2025. arXiv:2504 . 02818 [math.ST].url:https://arxiv.org/abs/2504.02818

    2025 arXiv

  34. [42]

    2022.url:https://www.whitehouse.gov/ ostp/ai-bill-of-rights

    WHO.The Blueprint for an AI Bill of Rights. 2022.url:https://www.whitehouse.gov/ ostp/ai-bill-of-rights

    2022

  35. [43]

    Sequential (quickest) change detection: Classical results and new directions

    Liyan Xie et al. “Sequential (quickest) change detection: Classical results and new directions”. In:IEEE Journal on Selected Areas in Information Theory2.2 (2021), pp. 494–514

    2021

  36. [44]

    Online multiple testing with e-values

    Ziyu Xu and Aaditya Ramdas. “Online multiple testing with e-values”. In:International Con- ference on Artificial Intelligence and Statistics. PMLR. 2024, pp. 3997–4005

    2024

  37. [45]

    Asynchronousonlinetestingofmultiple hypotheses

    TijanaZrnic,AadityaRamdas,andMichaelIJordan.“Asynchronousonlinetestingofmultiple hypotheses”. In:Journal of Machine Learning Research22.33 (2021), pp. 1–39. 16 A Proofs For the following proofs, we recall our setting established in Section 2, introduce some additional notation...

    2021

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.