Bandit Multiclass PAC Learning - Corrected Lower Bounds, Exact Families, and a Confidence Direct-Sum Phenomenon

We study realizable multiclass PAC learning with bandit feedback: the learner observes an i.i.d. instance, predicts one of K labels, and learns only whether the prediction was correct. Hanneke, Meng, Moran, and Shaeiri (arXiv:2605.25678) characterized the optimal sample complexity via the bandit DS dimension BDS(H) up to logarithmic factors, and asked whether every class admits sample complexity O((BDS+log(1/δ))/ε). We give a fine-grained resolution of this landscape. First, we show that the published lower bound Ω((BDS+log(1/δ))/ε) is incorrect as stated: we exhibit explicit classes with BDS=K-1 whose sample complexity is exponentially smaller, and locate two independent gaps in the published proof. We repair the lower-bound theory around a new anchored dimension aBDS(H)(H), proving a constant-free three-part lower bound. On the upper-bound side we remove the ambient label count K entirely, proving O((Blog³ B+Blog(1/δ))/ε) for B=BDS(H), plus a constant-confidence bound via a new fiberization lemma; for two natural families we determine the sample complexity up to constant factors, with no logarithmic slack. Finally, we answer the open question in the negative under its uniform-constant reading and show the failure is intrinsic: for an explicit affine multiplexer class we establish the full confidence profile Θ((nminn,log(1/δ)+log(1/δ))/ε) — a confidence direct-sum regime in which a multiplicative log(1/δ) cost is information-theoretically necessary, followed by a rank-saturation phase transition. Moreover, two classes with identical dimension profiles can have polynomially different sample complexities, so no characterization by these dimensions alone can be accurate to polylogarithmic factors. We also explain why these results are consistent with additive-confidence list-PAC guarantees combined with the ListCascade bridge.