Breakthrough Tracker record
Correlated two-sided Gaussian tests can defeat the Benjamini–Hochberg guarantee
An explicit correlated Gaussian factor model, backed by an interval-arithmetic certificate, has asymptotic false discovery rate above 0.0104 when Benjamini–Hochberg is run at nominal level 0.01. A follow-on paper extends the negative result.
← Back to the filtered Breakthrough Tracker
- Stable ID
math-benjamini-hochberg-correlated-gaussian-2026- Revision
math-benjamini-hochberg-correlated-gaussian-2026.v1- Field
- Mathematics · Mathematical statistics
- Evidence
- Tier 1 · Peer reviewed: No
- Record state
- Current · Expert-supported preprint with reproducible certificate
- Last checked
AI role
The paper states that GPT-5.6 Pro produced the proof and that Dobriban carefully checked it.
Record details
- Problem or result
- False-discovery-rate control for Benjamini–Hochberg under unrestricted Gaussian correlation
- Authors
- Edgar Dobriban
- Institutions
- University of Pennsylvania
- Result date
- Preprint submitted July 13, 2026
Why it matters
The result closes a precise theoretical question about one of the most widely used multiple-testing procedures.
Limits
It does not invalidate Benjamini–Hochberg under independence or accepted positive-dependence conditions. The counterexample uses a special unrestricted dependence setting, and its exhibited excess is small.
Sources
- Primary: Edgar Dobriban, The Benjamini--Hochberg Procedure Can Fail to Control the FDR for Correlated Two-Sided Gaussian Tests
- Independent: Lihua Lei, No Universal Multiplicative FDR Bound for the BH Procedure
- Independent: Independent technical explainer with scope caveats
Correction and revision history
- Added on 2026-07-22 with the dependence assumptions and small numerical excess made explicit to prevent overstatement.
Machine-readable: JSON v1 · CSV v1 · Schema v1
This individual record remains noindex until a story-specific featured image passes Kingy’s rendered-pixel visual review. The source-linked tracker hub remains the public index.