The Benjamini--Hochberg Procedure Can Fail to Control the FDR for Correlated Two-Sided Gaussian Tests

Edgar Dobriban

In a specified three-block Gaussian factor model with 96N true nulls and 4N nonnulls, ordinary Benjamini–Hochberg testing at level 0.01 has false discovery rate greater than 0.0104 for all sufficiently large N when applied to correlated two-sided Gaussian p-values.

Sources for this record

What was already known

Ordinary Benjamini–Hochberg control was known under independence and positive regression dependence. Its validity for correlated two-sided Gaussian tests had remained conjectural.

AI involvement AI-autonomous

Edgar Dobriban reports giving GPT-5.6 Pro the problem definition alone. The model produced the counterexample, analytic proof and certificate code; Dobriban checked and edited the result.

The source says the model produced the counterexample and proof from only the problem definition, with later human checking and editing.

Importance Notable

Refutes the conjectured general FDR guarantee with a concrete asymptotic counterexample.

Limitations

The theorem concerns the specified dependence model and asymptotic N, not every Gaussian dependence structure. Dobriban’s complete review and the exact interval certificate are reported by the paper; neither was rerun here.

Technical verification Formal (python-flint/Arb outward-rounded interval arithmetic + human)

The author reports a complete check of the analytic argument and associated certificate, and the final strict inequality is certified by outward-rounded Arb interval arithmetic over the specified finite grids. These are source-reported checks.

Formal checking

Careful check of the entire argument and numerical certificate; source-reported.

Checker
Edgar Dobriban, paper author
Scope
Theorem 1's FDR > 0.0104 counterexample in the specified Gaussian factor model for all sufficiently large N
Basis
Reported by the linked checker

Outward-rounded ball arithmetic with rational inputs and strict interval comparisons; source-reported.

Checker
Arb interval arithmetic through the Appendix B Python certificate, as reported by Dobriban
Scope
Strict finite inequalities yielding the certified FDR lower bound 0.010416829... on the specified grids
Basis
Reported by the linked checker
Evidence source

Checking a formal statement does not establish that it matches the original problem.

Human understanding Full

Dobriban explains why the opposed common-factor movements of the null and signal blocks create the failure, and lays out the threshold and certificate argument in Sections 1 and 3–6.

Correction and publication history

2026-10-18 · publication

Refutes the conjectured general FDR guarantee with a concrete asymptotic counterexample.

Read this published revision