Search papers, labs, and topics across Lattice.
This paper reveals that the widely used Benjamini鈥揌ochberg procedure can fail to maintain the false discovery rate (FDR) at its intended level when applied to correlated two-sided Gaussian p-values. By constructing a specific factor model, the authors demonstrate that for a significance level of 伪=0.01, the FDR exceeds 0.0104 as the number of hypotheses increases, thus disproving a long-held conjecture. Monte Carlo simulations corroborate the theoretical findings, which were verified using GPT-5.6 Pro for proof generation.
The Benjamini鈥揌ochberg procedure may not control the FDR as expected, with implications for thousands of studies relying on its validity.
We show that the Benjamini--Hochberg procedure can fail to control the false discovery rate (FDR) at its nominal level for correlated two-sided Gaussian $p$-values. We construct a factor model for which, at level $伪=0.01$, a rigorous interval-arithmetic certificate proves $FDR>0.0104$ for all sufficiently large numbers of hypotheses. This disproves a conjecture widely believed to be true for twenty years. Monte Carlo experiments are consistent with the theoretical result. The proof was obtained by GPT-5.6 Pro and carefully checked by the author.
