Expectation under a finite categorical outcome measure
ProvedBanditAlgorithm.integral_pmOutcomeMeasurefinite-distributionspartial-monitoringprobability
Let be a probability vector on a finite set and let be its categorical probability measure. Then every real-valued function has expectation
This is the integration rule that turns one-step action expectations under Algorithm 26 into finite weighted sums. The simplex hypothesis supplies nonnegative masses summing to one, including the finiteness needed by the Bochner integral.
Preamble
import Definitions.Def_PartialMonitoringStochastic import Mathlib.MeasureTheory.Integral.Bochner.SumMeasure open MeasureTheory open scoped BigOperators ENNReal namespace BanditAlgorithm
Formal statement
theorem integral_pmOutcomeMeasure
{d : ℕ} (p : Fin d → ℝ) (hp : p ∈ stdSimplex ℝ (Fin d))
(g : Fin d → ℝ) :
∫ i, g i ∂pmOutcomeMeasure p = ∑ i, p i * g i := by sorrySource
Finite categorical expectation identity; used implicitly throughout Lattimore and Szepesvári, Bandit Algorithms (2020), proof of Theorem 37.15, printed pp. 494–495, https://tor-lattimore.com/downloads/book/book.pdf