Prefix consistency of the Bayesian adversarial-bandit law
ProvedBanditAlgorithm.bayesianAdversarialMeasure_prefix_marginalbanditsprobabilitythompson-sampling
For the Bayesian adversarial-bandit interconnection, the law constructed through any horizon has the recursively constructed law through every earlier time as its prefix marginal. Concretely, if one keeps the sampled reward matrix and truncates the length- history to its first observations, the resulting pushforward measure is exactly the joint law of the reward matrix and the length- history.
This projective-consistency result is the structural bridge that permits one-round conditional calculations to be performed under the recursively defined prefix law while the mutual-information definition uses the terminal law.
Preamble
import Definitions.Def_ThompsonSampling open MeasureTheory ProbabilityTheory
Formal statement
namespace BanditAlgorithm
/-- The terminal interconnection law has the recursively constructed shorter law as every prefix. -/
theorem bayesianAdversarialMeasure_prefix_marginal {k n : ℕ}
(Q : Measure (Fin n → Fin k → ℝ)) [IsProbabilityMeasure Q]
(pi : BanditPolicy k) (u : ℕ) (hu : u ≤ n) (t : ℕ) (ht : t ≤ u) :
Measure.map
(fun p : (Fin n → Fin k → ℝ) × BanditHistory k u ↦
(p.1, fun s : Fin t ↦ p.2 (Fin.castLE ht s)))
(bayesianAdversarialMeasure Q pi u hu) =
bayesianAdversarialMeasure Q pi t (ht.trans hu) := by
sorry
end BanditAlgorithmSource
Lattimore and Szepesvári, Bandit Algorithms (2020), §36.4, printed pp. 469–470 (free PDF pp. 478–479), https://tor-lattimore.com/downloads/book/book.pdf; purely formal projective-consistency bridge for the recursively defined interconnection law.