Pith. sign in

REVIEW 5 minor 262 references

The Gavrilov–Benjamini–Sarkar step-down procedure matches the Bayes Oracle risk under sparsity.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 00:24 UTC pith:65HKNY4K

load-bearing objection First ABOS proof for GBS via a direct risk decomposition that avoids BH-style threshold localization; solid, self-contained theory that belongs in the literature.

arxiv 2607.03719 v1 pith:65HKNY4K submitted 2026-07-04 math.ST stat.TH

Asymptotic Bayes Optimality Under Sparsity of the Gavrilov-Benjamini-Sarkar Step-Down Testing Procedure

Prasenjit Ghosh , Arijit Chakrabarti This is my paper
classification math.ST stat.TH MSC 62C1562C2062F0362G1062J15
keywords asymptotic Bayes optimality under sparsityfalse discovery rateGavrilov–Benjamini–Sarkar proceduremultiple testingsparse Gaussian sequence modelstep-down procedures
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper shows that a practical step-down multiple-testing rule, the Gavrilov–Benjamini–Sarkar (GBS) procedure, is asymptotically Bayes-optimal under sparsity in the Gaussian sequence model. Under a spike-and-slab prior and the usual sparse asymptotics, its Bayes risk becomes equivalent to that of the ideal Bayes Oracle that knows the unknown sparsity and signal strength. The result matters because GBS is already known for strong finite-sample power while controlling the false discovery rate, yet until now it lacked a decision-theoretic optimality guarantee of the kind long established for the Benjamini–Hochberg procedure. The authors obtain the guarantee by a direct risk analysis that separately bounds false discoveries and missed signals, avoiding the usual approximation of a random threshold by a deterministic surrogate.

Core claim

Under the sparse Gaussian two-groups model and a natural calibration of the nominal FDR level to the unknown sparsity and signal strength, the Bayes risk of the GBS step-down procedure is asymptotically equivalent to the Bayes Oracle risk; that is, the ratio of the two risks tends to one.

What carries the argument

A finite-sample inequality for shifted order statistics of the GBS critical constants (controlling false discoveries) together with a signal-crossing argument for ordered alternative p-values that exploits the sequential step-down structure (controlling missed signals).

Load-bearing premise

The nominal FDR level must be tuned so that the GBS critical sequence lines up with the unknown sparsity and signal strength in a precise asymptotic way; without that alignment the Type-II crossing argument fails.

What would settle it

Compute the ratio of GBS Bayes risk to Oracle risk under the paper's sparse regime but with a deliberately mis-calibrated α_m that violates the two calibration conditions (15); if the ratio stays near 1, the claim fails.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 5 minor

Summary. The paper studies the Gavrilov–Benjamini–Sarkar (GBS) step-down FDR procedure in the sparse Gaussian sequence model under the spike-and-slab formulation of Bogdan et al. (2011). Under Assumption 1 and the GBS calibration/compatibility conditions (15), Theorem 1 asserts that the Bayes risk of GBS is asymptotically equivalent to the Bayes Oracle risk, i.e., R_GBS_m / R_BO_m → 1 (ABOS). The argument proceeds by direct risk decomposition rather than threshold localization: Corollary 1 (from the finite-sample shifted-order-statistic inequality of Lemma 1) controls the false-discovery term, while Lemmas 2–4 establish a deterministic and then empirical signal-crossing property that bounds the false-nondiscovery term. The appendix supplies complete inductive proofs of these lemmas.

Significance. If correct, the result places GBS among the small set of practically implementable procedures known to be ABOS, matching the classical status of BH under the same framework. The technical contribution is genuine: Lemma 1 is a new finite-sample inequality tailored to the GBS critical constants, and the three-stage Type-II argument (deterministic crossing, initial crossing, empirical crossing) exploits the step-down structure without approximating a random threshold. The proofs are fully written out and self-contained; the calibration conditions (15) are stated explicitly as the natural analogue of those used for BH. This is a clean, decision-theoretic justification for a widely used procedure that previously lacked an ABOS analysis.

minor comments (5)
  1. Section 2.3, Assumption 1: the phrasing “satisfy satisfies” is a typographical slip; correct to “satisfies”.
  2. Section 3, display (15) and the surrounding paragraph: it would help the reader to state more explicitly that the same constant C appears in both Assumption 1(v) and the first calibration condition, so that the Oracle Type-II rate 2Φ(√C)−1 is recovered.
  3. Lemma 2 proof: the passage from (47) to the uniform bound (51) uses a fixed multiplicative margin η_ε; a one-sentence remark that the margin can be taken arbitrarily close to ε/(q*−ε) would clarify sharpness.
  4. References: several of the authors’ own 2026 preprints are cited for motivation (MRD–GBS, admissibility). A brief parenthetical note that those works are not used in the proof of Theorem 1 would avoid any appearance of circular dependence.
  5. Notation: the same symbol R is used both for the Bayes risk and for the number of rejections (R_GBS_m). Distinguishing them (e.g., R vs. R̂) would improve readability in Section 3.3.

Circularity Check

0 steps flagged

No significant circularity: ABOS of GBS is derived from model assumptions, GBS critical constants, and explicit calibration conditions via new finite-sample and signal-crossing arguments.

full rationale

The paper's central claim (Theorem 1) is that under Assumption 1 plus the GBS calibration/compatibility conditions (15), the Bayes risk of the GBS procedure is asymptotically equivalent to the Bayes Oracle risk. This is not assumed or fitted; it is proved by separately bounding the Type-I and Type-II components of the risk. Type-I control follows from a new finite-sample inequality for shifted order statistics of the GBS critical constants (Lemma 1 and Corollary 1), which yields E(V_GBS_m) = O(α_m m p_m) and hence a negligible false-discovery contribution once α_m δ_m o 0. Type-II control proceeds via a deterministic signal-crossing bound (Lemma 2), an initial-crossing condition (Lemma 3), and an empirical crossing result for ordered alternative p-values (Lemma 4) that exploits the step-down structure; these together give E(T_GBS_m) ≤ m p_m (2Φ(√C)-1) + o(m p_m). Combining the bounds with the known Oracle expansion produces R_GBS_m / R_BO_m o 1. The calibration conditions (15) are stated explicitly as hypotheses of Theorem 1 (they are the natural analogues of the BH calibration of Bogdan et al. 2011) and are not hidden. Self-citations to the authors' related dependence papers appear only for motivation and empirical context and do not enter the proof of Theorem 1. The derivation is therefore self-contained against the stated model and assumptions; no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper works entirely inside the classical sparse Gaussian sequence model and the ABOS framework of Bogdan et al. (2011). No free parameters are fitted to data; the only free sequences are the model parameters (p_m, ψ_m^{2}, δ_m) and the nominal FDR level α_m, all of which are required to satisfy explicit asymptotic conditions. No new physical or statistical entities are postulated.

axioms (5)
  • domain assumption Sparse Gaussian sequence model with two-groups (spike-and-slab) prior: X_i | μ_i ~ N(μ_i,1), μ_i ~ (1-p_m)δ_0 + p_m N(0,ψ_m^{2})
    Section 2.1; standard setting for ABOS analyses since Bogdan et al. (2011).
  • domain assumption Assumption 1: p_m o0, m p_m o∞, u_m=ψ_m^{2} o∞, v_m o∞, log(v_m)/u_m o C ∈ (0,∞)
    Section 2.3; the sparse asymptotic regime under which the oracle risk expansion is known.
  • ad hoc to paper GBS calibration: α_m o0, α_m f_m o∞, 2 log(f_m/r_α_m)/u_m o C, log{1/(r_α_m δ_m)}/u_m o0
    Display (15) and surrounding text; natural analogue of the BH calibration of Bogdan et al., but still an extra tuning requirement for GBS.
  • domain assumption Independence of the X_i (hence of the p-values)
    Model (1); required for both the FDR control of GBS and the order-statistic arguments.
  • standard math Additive Bayes risk R_m = δ_{0,m} E(V_m) + δ_{A,m} E(T_m)
    Section 2.2; standard decision-theoretic loss used throughout the ABOS literature.

pith-pipeline@v1.1.0-grok45 · 24940 in / 2802 out tokens · 18097 ms · 2026-07-12T00:24:15.869481+00:00 · methodology

0 comments
read the original abstract

In this article, we investigate the asymptotic Bayes optimality under sparsity (ABOS) of the Gavrilov-Benjamini-Sarkar (GBS) step-down multiple testing procedure of Gavrilov et al. (2009) in the sparse Gaussian sequence model. While the asymptotic optimality properties of the Benjamini-Hochberg procedure have been extensively studied, corresponding results for the GBS procedure remain unavailable despite its favorable finite-sample performance and widespread applicability. Within the spike-and-slab Bayesian formulation and the asymptotic decision-theoretic framework of Bogdan et al. (2011), we establish that the GBS procedure is ABOS over a broad class of sparse asymptotic regimes. Existing ABOS analyses of the Benjamini-Hochberg procedure rely on approximating the random rejection threshold by a suitable deterministic surrogate. In contrast, our approach analyzes the Bayes risk directly, separately controlling the false discovery and false nondiscovery components. The analysis combines two key ingredients: a new finite-sample inequality for shifted order statistics associated with the GBS critical constants and a signal-crossing argument for ordered alternative p-values that exploits the procedure's sequential step-down structure. Together, these tools yield the desired ABOS property without resorting to threshold-localization arguments. To the best of our knowledge, this is the first asymptotic decision-theoretic analysis of the GBS procedure and the first proof of its asymptotic Bayes optimality under sparsity.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

262 extracted references · 2 linked inside Pith

  1. [1]

    2026 , eprint =

    Ghosh, Prasenjit and Chakrabarti, Arijit , title =. 2026 , eprint =

    2026

  2. [2]

    Electronic Journal of Statistics , year =

    Arias-Castro, Ery and Chen, Shuo , title =. Electronic Journal of Statistics , year =

  3. [3]

    and Wainwright, Martin J

    Rabinovich, Maxim and Ramdas, Aaditya and Jordan, Michael I. and Wainwright, Martin J. , title =. Statistica Sinica , year =

  4. [4]

    The Annals of Statistics , year =

    Belitser, Eduard and Nurushev, Nurlan , title =. The Annals of Statistics , year =

  5. [5]

    The Annals of Statistics , year =

    Fromont, Magalie and Lerasle, Matthieu and Reynaud-Bouret, Patricia , title =. The Annals of Statistics , year =

  6. [6]

    arXiv preprint arXiv:1711.08705 , year =

    Salomond, Jean-Baptiste , title =. arXiv preprint arXiv:1711.08705 , year =. 1711.08705 , archivePrefix =

    Pith/arXiv arXiv

  7. [7]

    Journal of Multivariate Analysis , volume =

    Ledoit, Olivier and Wolf, Michael , title =. Journal of Multivariate Analysis , volume =

  8. [8]

    and Levina, Elizaveta , title =

    Bickel, Peter J. and Levina, Elizaveta , title =. The Annals of Statistics , volume =

  9. [9]

    Tony and Liu, Weidong , title =

    Cai, T. Tony and Liu, Weidong , title =. Journal of the American Statistical Association , volume =

  10. [10]

    Biostatistics , volume =

    Friedman, Jerome and Hastie, Trevor and Tibshirani, Robert , title =. Biostatistics , volume =

  11. [11]

    2025 , eprint =

    Paul, Sayantan and Chakrabarti, Arijit , title =. 2025 , eprint =

    2025

  12. [12]

    Annals of the Institute of Statistical Mathematics , year =

    Paul, Sayantan and Chakrabarti, Arijit , title =. Annals of the Institute of Statistical Mathematics , year =

  13. [13]

    Asymptotic Bayes Optimality Under Sparsity of Some Multiple Testing Procedures , journal =

    Frommlet, Florian and Bogdan, Ma. Asymptotic Bayes Optimality Under Sparsity of Some Multiple Testing Procedures , journal =. 2013 , volume =

    2013

  14. [14]

    2015 , eprint =

    Prasenjit Ghosh and Arijit Chakrabarti , title =. 2015 , eprint =

    2015

  15. [15]

    arXiv preprint arXiv:2606.07466 , year =

    Ghosh, Prasenjit and Chakrabarti, Arijit , title =. arXiv preprint arXiv:2606.07466 , year =. 2606.07466 , archivePrefix =

    Pith/arXiv arXiv

  16. [16]

    , year =

    Ghosh, Prasenjit , title =. , year =. , archivePrefix =

  17. [17]

    2025 , eprint =

    Paul, Sayantan and Ghosh, Prasenjit and Chakrabarti, Arijit , title =. 2025 , eprint =

    2025

  18. [18]

    and Castillo, I

    Abraham, K. and Castillo, I. and Roquain, E. , title =. The Annals of Statistics , volume =

  19. [19]

    Lehmann, E. L. and Casella, G. , title =

  20. [20]

    Lehmann, E. L. and Romano, J. P. , title =

  21. [21]

    Armagan and D

    A. Armagan and D. B. Dunson and M. Clyde , editor =. Generalized Beta Mixtures of Gaussians , booktitle =

  22. [22]

    , title =

    Clyde, Merlise and George, Edward I. , title =. Statistical Science , volume =

  23. [23]

    and Sillanp

    O'Hara, Robert B. and Sillanp. A Review of Bayesian Variable Selection Methods: What, How and Which , journal =

  24. [24]

    Shaffer, J. P. , editor =. Simultaneous testing. , booktitle =

  25. [25]

    Cournot, A. A. , title =. Reprinted 1984 as Volume 1 of Cournot's Oeuvres Completes , publisher =

    1984

  26. [26]

    Berry, D. A. , editor =. Multiple Comparisons, Multiple Tests and Data Dredging: A Bayesian Perspective , booktitle =

  27. [27]

    Irwin, J. O. , title =. Biometrika , volume =

  28. [28]

    Biometrika , volume =

    Student , title =. Biometrika , volume =

  29. [29]

    and Hotelling, H

    Working, H. and Hotelling, H. , title =. Journal of the American Statistical Association , volume =

  30. [30]

    Fisher, R. A. , title =

  31. [31]

    , title =

    Newman, D. , title =. Biometrika , volume =

  32. [32]

    Tukey, J. W. , title =. Biometrics , volume =

  33. [33]

    Duncan, D. B. , title =. Unpublished doctoral thesis, Iowa State College, Ames, Iowa , year =

  34. [34]

    Duncan, D. B. , title =. Virginia Journal of Science , volume =

  35. [35]

    Duncan, D. B. , title =. Virginia Polytechnique Institute Technical Report , year =

  36. [36]

    Duncan, D. B. , title =. Biometrics , volume =

  37. [37]

    Duncan, D. B. , title =. The Annals of Mathematical Statistics , volume =

  38. [38]

    Duncan, D. B. , title =. Technometrics , volume =

  39. [39]

    Duncan, D. B. and Bonner, R. G. , title =. Virginia Agri. Exp. Sta. Tech. Rept., Blacksburg, Va , year =

  40. [40]

    A method for judging all contrasts in the analysis of variance , journal =

    Scheff. A method for judging all contrasts in the analysis of variance , journal =

  41. [41]

    and Peritz, E

    Marcus, R. and Peritz, E. and Gabriel, K. R. , title =. Biometrika , volume =

  42. [42]

    Roy, S. N. , title =. The Annals of Mathematical Statistics , volume =

  43. [43]

    Roy, S. N. and Bose, R. C. , title =. The Annals of Mathematical Statistics , volume =

  44. [44]

    Armagan and D

    A. Armagan and D. B. Dunson and J. Lee , title =. Statistica Sinica , volume =

  45. [45]

    Cai, T. T. and Jin, J. and Low, M. J. , title =. The Annals of Statistics , volume =

  46. [46]

    and Jin, J

    Donoho, D. and Jin, J. , title =. The Annals of Statistics , volume =

  47. [47]

    van der Pas, S. L. and Salomond, J. B. and Schmidt-Hieber, J. , title =. Electronic Journal of Statistics , volume =

  48. [48]

    George, E. I. and Foster, D. P. , title =. Biometrika , volume =

  49. [49]

    Donoho, D. L. and Johnstone, I. M. , title =. Journal of the American Statistical Association , volume =

  50. [50]

    and Li, R

    Fan, J. and Li, R. , title =. Journal of the American Statistical Association , volume =

  51. [51]

    and Fan, J

    Antoniadis, A. and Fan, J. , title =. Journal of the American Statistical Association , volume =

  52. [52]

    and Benjamini, Y

    Abramovich, F. and Benjamini, Y. , title =. In Wavelets and Statistics. Lecture Notes in Statistics , volume =

  53. [53]

    and Braun, H

    Benjamini, Y. and Braun, H. , title =. The Annals of Statistics , volume =

  54. [54]

    and Benjamini, Y

    Abramovich, F. and Benjamini, Y. , title =. Computational Statistics and Data Analysis , volume =

  55. [55]

    and Birg

    Barron, A. and Birg. Risk bounds for model selection via penalization. , journal =

  56. [56]

    Gaussian model selection

    Birg. Gaussian model selection. , journal =

  57. [57]

    Bayesian Estimation of Sparse Signals with a Continuous Spike-and-Slab Prior , year =

    Ro. Bayesian Estimation of Sparse Signals with a Continuous Spike-and-Slab Prior , year =

  58. [58]

    and Datta, J

    Bhadra, A. and Datta, J. and Polson, N. G. and Willard, B. , title =. Bayesian Analysis , volume =

  59. [59]

    , title =

    Tibshirani, R. , title =. Journal of the Royal Statistical Society: Series B , volume =

  60. [60]

    Donoho, D. L. and Johnstone, I. M. , title =. Biometrika , volume =

  61. [61]

    Donoho, D. L. and Johnstone, I. M. , title =. Probability Theory Related Fields , volume =

  62. [62]

    Donoho, D. L. and Johnstone, I. M. , title =. Bernoulli , volume =

  63. [63]

    Donoho, D. L. and Johnstone, I. M. , title =. The Annals of Statistics , volume =

  64. [64]

    Pinsker, M. S. , title =. Problems Information Transmission , volume =

  65. [65]

    Efroimovich, S. Y. and Pinsker, M. S. , title =. Automat. Remote Control (translated from Avtomatika i Telemekhanika) , volume =

  66. [66]

    and Greenshtein, V

    Abramovich, F. and Greenshtein, V. and Pensky, M. , title =. The Annals of Statistics , volume =

  67. [67]

    and Angelini, C

    Abramovich, F. and Angelini, C. , title =. Sankhy

  68. [68]

    Brown, L. D. and Greenshtein, E. , title =. The Annals of Statistics , volume =

  69. [69]

    Zhang, C. H. and Huang, J. , title =. The Annals of Statistics , volume =

  70. [70]

    Zhang, C. H. , title =. The Annals of Statistics , volume =

  71. [71]

    , title =

    Zou, H. , title =. Journal of the American Statistical Association , volume =

  72. [72]

    and Hastie, T

    Zou, H. and Hastie, T. , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =

  73. [73]

    and Schmidt-Hieber, J

    Castillo, I. and Schmidt-Hieber, J. and van der Vaart, A. W. , title =. 2014 , eprint =

    2014

  74. [74]

    and Mizera, I

    Koenker, R. and Mizera, I. , title =. Journal of the American Statistical Association , volume =

  75. [75]

    and Benjamini, Y

    Abramovich, F. and Benjamini, Y. and Donoho, D. L. and Johnstone, I. M. , title =. The Annals of Statistics , volume =

  76. [76]

    Golubev, G. K. , title =. Problemy Peredachi Informatsii , volume =

  77. [77]

    Cai, T. T. and Jin, J. , title =. The Annals of Statistics , volume =

  78. [78]

    , title =

    Gelman, A. , title =. Bayesian Analysis , volume =

  79. [79]

    , title =

    Jin, J. , title =. Journal of the Royal Statistical Society: Series B , volume =

  80. [80]

    Ingster, Y. I. , title =. Mathematical Methods of Statistics , volume =

Showing first 80 references.