A Unified Optimism-Agnostic Framework for Linear Bandits over Spherical Action Sets
The results provide a modular framework that can be used to analyze linear bandit algorithms and explicitly connect quality of parameter estimation to optimal regret accumulation.