BanditRLlib
Lean gate passed before this site build; local proof declarations are shown as compiled.

Lean module · Tsallis-FTRL

BanditRLProof.TsallisFTRLFiniteHorizonSelection

# Canonical half-Tsallis finite-horizon minimizers This module applies the fixed half-Tsallis minimizer choice to cumulative loss vectors. It removes the caller-supplied minimizer certificates from the deterministic finite-horizon FTRL decomposition and records the successor indexing needed by importance-weighted updates. The selected minimizer uses `Classical.choose`. No measurability of the history-dependent selector, conditional expectation identity, or bound on the finite-horizon stability sum is claimed here.

Module map

Declarations
5
Placeholders
0

Imports

BanditRLProof.TsallisFTRLMinimizerExistence

Imported by

BanditRLProof, BanditRLProof.TsallisFTRLConditionalStability

Declarations

Open an item to read its exact compact statement and source link. Detailed teaching notes are linked when registered.

def BanditRLProof.Tsallis.halfTsallisCumulativeMinimizer Compiled

The fixed half-Tsallis minimizer for losses accumulated before round `t`.

noncomputable def halfTsallisCumulativeMinimizer {Action : Type u} [DecidableEq Action] (arms : Finset Action) (harms : arms.Nonempty) (eta : Real) (loss : Nat -> Action -> Real) (t : Nat) : Action -> Real
theorem BanditRLProof.Tsallis.halfTsallisCumulativeMinimizer_isRegularizedMinimizer Compiled

Every canonical cumulative selector carries the minimizer certificate required by the finite-horizon FTRL decomposition.

theorem halfTsallisCumulativeMinimizer_isRegularizedMinimizer {Action : Type u} [DecidableEq Action] (arms : Finset Action) (harms : arms.Nonempty) (eta : Real) (loss : Nat -> Action -> Real) (t : Nat) : FTRL.IsRegularizedMinimizer (FTRL.finiteSimplex arms) arms eta (negEntropyRegularizer arms (1 / 2 : Real)) (FTRL.cumulativeLoss loss t) (halfTsallisCumulativeMinimizer arms harms eta loss t)
theorem BanditRLProof.Tsallis.halfTsallisCumulativeMinimizer_succ Compiled

The selector at `t + 1` is the fixed minimizer after appending round `t`'s loss vector.

theorem halfTsallisCumulativeMinimizer_succ {Action : Type u} [DecidableEq Action] (arms : Finset Action) (harms : arms.Nonempty) (eta : Real) (loss : Nat -> Action -> Real) (t : Nat) : halfTsallisCumulativeMinimizer arms harms eta loss (t + 1) = halfTsallisMinimizer arms harms eta (fun action => FTRL.cumulativeLoss loss t action + loss t action)
theorem BanditRLProof.Tsallis.halfTsallisCumulativeMinimizer_succ_eq_updated Compiled

If the realized round loss is the importance-weighted estimator generated from the current selector, the successor selector is exactly the canonical one-step updated minimizer.

theorem halfTsallisCumulativeMinimizer_succ_eq_updated {Action : Type u} [DecidableEq Action] (arms : Finset Action) (harms : arms.Nonempty) (eta : Real) (loss : Nat -> Action -> Real) (rawLoss : Action -> Real) (chosen : Action) (t : Nat) (hloss : loss t = Exp3.importanceWeightedLoss (halfTsallisCumulativeMinimizer arms harms eta loss t) rawLoss chosen) : halfTsallisCumulativeMinimizer arms harms eta loss (t + 1) = halfTsallisUpdatedMinimizer arms harms eta (FTRL.cumulativeLoss loss t) rawLoss chosen
theorem BanditRLProof.Tsallis.cumulativeLinearLoss_sub_comparator_le_stability_add_powerSumPenalty_half_canonical Compiled

Finite-horizon half-Tsallis FTRL decomposition with the cumulative minimizer sequence selected internally. The remaining first term on the right is the pathwise stability sum.

theorem cumulativeLinearLoss_sub_comparator_le_stability_add_powerSumPenalty_half_canonical {Action : Type u} [DecidableEq Action] (arms : Finset Action) (harms : arms.Nonempty) (eta : Real) (loss : Nat -> Action -> Real) (q : Action -> Real) (T : Nat) (heta : 0 < eta) (hq : FTRL.finiteSimplex arms q) : let p := halfTsallisCumulativeMinimizer arms harms eta loss (Finset.range T).sum (fun t => FTRL.linearLoss arms (p t) (loss t) - FTRL.linearLoss arms q (loss t)) <= (Finset.range T).sum (fun t => FTRL.linearLoss arms (p t) (loss t) - FTRL.linearLoss arms (p (t + 1)) (loss t)) + ((powerSum arms (1 / 2 : Real) (p 0) - powerSum arms (1 / 2 : Real) q) / (1 - (1 / 2 : Real))) / eta