The GLR statistic Z_t is adapted to the natural filtration
ProvedBanditAlgorithm.trajGLR_measurable_filtrationbanditsbest-arm-identificationmeasurability
The generalised-likelihood-ratio statistic
of Lattimore--Szepesv'ari p.\ 409 is -measurable, where is the natural filtration.
This is what makes a stopping time, and it is not immediate: is defined through the empirical best arm , which is a Classical.choose maximiser and carries no measurability of its own. The point is that is tie-independent -- if two arms are tied for the empirical best, the pair term between them vanishes, so regardless of which is selected -- and therefore agrees with the value computed at the canonical least-index maximiser, which is -measurable.
Preamble
import Definitions.Def_TrackAndStop import Definitions.Def_GaussianBandit open MeasureTheory ProbabilityTheory InformationTheory NNReal ENNReal Filter
Formal statement
theorem BanditAlgorithm.trajGLR_measurable_filtration {k : ℕ} [NeZero k] (t : ℕ) :
Measurable[BanditAlgorithm.banditFiltration k t] (BanditAlgorithm.trajGLR (k := k) t) := by
sorrySource
Lattimore & Szepesvari, Bandit Algorithms (CUP 2020), Section 33.2.2, Algorithm 21 and p. 409.