Unit-variance Gaussian bandits are 1-subgaussian
ProvedBanditAlgorithm.gaussianBandit_isSubgaussian_onebanditsgaussianprobabilitysubgaussian
Let be any real mean vector, and let arm have reward law . Then every centered reward has moment-generating function
Consequently the Gaussian bandit is -subgaussian: all arm rewards are integrable and their centered moment-generating functions are bounded by .
This packages the standard Gaussian MGF calculation in the stochastic-bandit interface for reuse by adaptive concentration theorems.
Preamble
import Definitions.Def_GaussianBandit open MeasureTheory ProbabilityTheory
Formal statement
namespace BanditAlgorithm
/-- Every unit-variance Gaussian bandit is 1-subgaussian. -/
theorem gaussianBandit_isSubgaussian_one {k : ℕ} (μvec : Fin k → ℝ) :
IsSubgaussianBandit 1 (gaussianBandit μvec) := by
sorry
end BanditAlgorithmSource
Lattimore and Szepesvári, Bandit Algorithms (CUP 2020), Example 5.6 and Gaussian moment-generating-function calculation, printed p. 75 / PDF p. 84; unit-variance Gaussian bandit class used in Theorem 36.3, printed p. 465 / PDF p. 474.