← All problems

Benjamini–Hochberg under correlated Gaussian tests

19 years old · traced to

Does the Benjamini–Hochberg procedure always control false-discovery rate at its nominal level for correlated two-sided Gaussian p-values?

References

Progress summary

Refreshed
Open · solution reported

A 2026 paper gives a checked counterexample, so ordinary Benjamini–Hochberg does not always control false-discovery rates for correlated two-sided Gaussian tests.

The problem asked whether the Benjamini–Hochberg procedure always achieves its nominal false-discovery rate for arbitrarily correlated two-sided Gaussian tests. The conjecture is now disproved by an explicit Gaussian factor-model construction.

July 2026 counterexample and sharp asymptotics

Dobriban’s construction has q=0.01q=0.01 and FDR⁡>0.0104\operatorname{FDR}>0.0104, with an interval-arithmetic certificate for sufficiently many hypotheses. A companion result proves that, for arbitrary correlation structures, the worst-case FDR is at least (12π−δ)qlog⁡(1/q)\left(\frac{1}{2\sqrt{\pi}}-\delta\right)q\log(1/q) for small qq, disproving every universal bound FDR⁡≤Cq\operatorname{FDR}\le Cq. For a broad common-factor class, the matching upper order is O(qlog⁡(1/q))O(q\log(1/q)).

Current status (as of July 2026): The universal-control conjecture is settled negatively; correlated two-sided Gaussian tests can make ordinary BH exceed its nominal level, with worst-case inflation growing on the order of log⁡(1/q)\log(1/q) in the exhibited class.

Sources

Solutions 0

No solutions have been posted yet.