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

Lean module · Foundations

BanditRLProof.RegretCountBounds

# Deterministic regret/count bounds This module contains algorithm-neutral deterministic scaffolds that convert per-arm pull-count upper bounds into pseudo-regret upper bounds. It stays below probability, expectation, filtrations, concentration, and algorithm final theorem work.

Module map

Declarations
3
Placeholders
0

Imports

BanditRLProof.FiniteBanditModelInvariants, BanditRLProof.RegretDecomposition

Imported by

BanditRLProof, BanditRLProof.Algorithms.ETCRegretLemmas

Declarations

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

theorem BanditRLProof.pseudoRegret_le_finset_sum_gap_mul_count_bound Compiled

If every arm's pull count is bounded by `B`, then pseudo-regret is bounded by the corresponding gap-weighted count budget. This is the `REGRET-COUNT-BOUND` deterministic scaffold. It consumes only the compiled regret decomposition and model-derived gap nonnegativity.

theorem pseudoRegret_le_finset_sum_gap_mul_count_bound {K : Nat} (model : FiniteBanditModel K) (action : ActionTrace (Fin K)) (n : Nat) (B : Fin K -> Rat) (hB : forall a : Fin K, ((pullCount action a n : Nat) : Rat) <= B a) : pseudoRegret model action n <= (Finset.univ : Finset (Fin K)).sum (fun a : Fin K => model.gap a * B a)
theorem BanditRLProof.pseudoRegret_le_finset_sum_gap_mul_nat_count_bound Compiled

Nat-valued per-arm pull-count bounds imply the corresponding gap-weighted pseudo-regret bound after casting the count budget to `Rat`. This is the `REGRET-NAT-COUNT-BOUND` adapter. It is algorithm-neutral and keeps ETC/UCB-specific count facts out of this file.

theorem pseudoRegret_le_finset_sum_gap_mul_nat_count_bound {K : Nat} (model : FiniteBanditModel K) (action : ActionTrace (Fin K)) (n : Nat) (B : Fin K -> Nat) (hB : forall a : Fin K, pullCount action a n <= B a) : pseudoRegret model action n <= (Finset.univ : Finset (Fin K)).sum (fun a : Fin K => model.gap a * (((B a : Nat) : Rat)))
theorem BanditRLProof.pseudoRegret_le_sum_gap_mul_uniform_nat_count_bound Compiled

A uniform Nat-valued pull-count bound implies pseudo-regret is bounded by the sum of model gaps times that uniform count budget. This is the `REGRET-UNIFORM-NAT-COUNT-BOUND` adapter. It is still algorithm-neutral and does not prove any ETC/UCB-specific count fact.

theorem pseudoRegret_le_sum_gap_mul_uniform_nat_count_bound {K : Nat} (model : FiniteBanditModel K) (action : ActionTrace (Fin K)) (n B : Nat) (hB : forall a : Fin K, pullCount action a n <= B) : pseudoRegret model action n <= ((Finset.univ : Finset (Fin K)).sum (fun a : Fin K => model.gap a)) * (((B : Nat) : Rat))