Search papers, labs, and topics across Lattice.
This paper clarifies the apparent contradiction in fixed-confidence best-arm identification by analyzing the role of the union bound in the context of multiple hypothesis testing. It reveals that while only one arm can be the best, the multiplicity of hypotheses leads to a Bonferroni-type correction factor of \(K-1\) due to two distinct orientations of the hypotheses. The findings bridge the gap between best-arm identification and strong familywise error rate (FWER) control, providing a clearer understanding of error rates in multi-armed bandit settings.
The apparent paradox of union bounds in best-arm identification is resolved, revealing that multiplicity issues manifest differently depending on hypothesis orientation.
In fixed-confidence best-arm identification, proofs often use a union bound across the competing arms. From a multiple-testing point of view this can look puzzling: if the best arm is unique, only one hypothesis of the form ``arm $i$ is best''can be true. Why then should there be a Bonferroni-type factor of $K-1$? The answer is that there are two natural ways to orient the hypotheses. In one orientation, best-arm identification is literally a strong familywise-error-rate (FWER) problem with $K-1$ true nulls. In the opposite orientation, exactly one null is true, but a pairwise implementation can falsely reject that one null through any of $K-1$ comparisons. Thus the multiplicity has not disappeared; it just pops up in different places. This note makes the equivalence explicit in the terminology of both communities.
