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

Lean module · Finite-horizon RL

BanditRLProof.RL.FiniteHorizonNaturalCausalInverseSqrtThresholdUnboundedHittingAfterExpectedPositivePartConsistency

# Expected positive-part consistency at an uncapped inverse-sqrt hitting time The exact stopped average realized behavior regret can be negative. This file isolates its one-sided excess, proves that its expectation is bounded by the hit threshold, and sends that threshold to zero. The argument uses the exact at-hit inequality and finite-measure integration, not optional stopping.

Module map

Declarations
4
Placeholders
0

Imports

BanditRLProof.RL.FiniteHorizonNaturalCausalInverseSqrtThresholdUnboundedHittingAfterIntegrableExpectedUpperBound

Imported by

BanditRLProof, BanditRLProof.RL.FiniteHorizonNaturalCausalInverseSqrtThresholdUnboundedHittingAfterL1Consistency

Declarations

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

def BanditRLProof.FiniteHorizonRL.AdaptiveStochasticSampledEmpiricalOptimisticSource.selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart Compiled

Expected positive part of the exact stopped average realized behavior regret as a function of the threshold schedule index.

noncomputable def selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart (mdp : MDP State Action) (initialState : Measure State) [IsProbabilityMeasure initialState] (rewardSource : mdp.MeanCompatibleRewardKernel) (initialTable : DeterministicMarkovPolicyTable mdp) (defaultState : State) (varianceProxy : NNReal) (baseVisitFloor : Real) (scheduleIndex : Nat) : Real
theorem BanditRLProof.FiniteHorizonRL.AdaptiveStochasticSampledEmpiricalOptimisticSource.selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart_nonneg Compiled

The expected positive part of the exact stopped process is nonnegative.

theorem selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart_nonneg (mdp : MDP State Action) (initialState : Measure State) [IsProbabilityMeasure initialState] (rewardSource : mdp.MeanCompatibleRewardKernel) (initialTable : DeterministicMarkovPolicyTable mdp) (defaultState : State) (varianceProxy : NNReal) (baseVisitFloor : Real) (scheduleIndex : Nat) : 0 <= selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart mdp initialState rewardSource initialTable defaultState varianceProxy baseVisitFloor scheduleIndex
theorem BanditRLProof.FiniteHorizonRL.AdaptiveStochasticSampledEmpiricalOptimisticSource.selfConsistentScheduledCausalSource_inverseSqrtThresholdUnboundedHittingAfter_stoppedAverageRealizedBehaviorRegretPositivePart_integrable_and_integral_le_threshold Compiled

At every fixed threshold index, the positive part of the exact stopped average realized behavior regret is integrable and its expectation is bounded by the inverse-square-root hit threshold.

theorem selfConsistentScheduledCausalSource_inverseSqrtThresholdUnboundedHittingAfter_stoppedAverageRealizedBehaviorRegretPositivePart_integrable_and_integral_le_threshold (mdp : MDP State Action) (initialState : Measure State) [IsProbabilityMeasure initialState] [StandardBorelSpace State] [StandardBorelSpace Action] (rewardSource : mdp.MeanCompatibleRewardKernel) (varianceProxy : NNReal) (hvarianceProxy : 0 < varianceProxy) (law : rewardSource.UniformSubgaussianRewardLaw varianceProxy) (initialTable : DeterministicMarkovPolicyTable mdp) (defaultState : State) (support : ExploratoryPathSupport mdp initialState) (baseVisitFloor : Real) (hbaseFloor : ExploratoryPathUniformVisitFloor support 1 baseVisitFloor) (hrewardBound : forall state action, |mdp.reward state action| <= 1) (hhorizon : 4 < mdp.horizon) (hbaseVisitFloor : 0 < baseVisitFloor) (scheduleIndex : Nat) : let source := selfConsistentScheduledCausalSource mdp initialState rewardSource initialTable defaultState varianceProxy baseVisitFloor let stoppingPrefix := selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppingPrefix mdp initialState rewardSource initialTable defaultState varianceProxy baseVisitFloor let stoppedProcess := selfConsistentScheduledNaturalCausalStoppingTimeAverageRealizedBehaviorRegretProcess mdp initialState rewardSource initialTable defaultState varianceProxy baseVisitFloor stoppingPrefix scheduleIndex Integrable (fun trajectory => max (stoppedProcess trajectory) 0) source.trajectoryMeasure /\ integral source.trajectoryMeasure (fun trajectory => max (stoppedProcess trajectory) 0) <= selfConsistentScheduledNaturalCausalInverseSqrtFirstPassageThreshold scheduleIndex
theorem BanditRLProof.FiniteHorizonRL.AdaptiveStochasticSampledEmpiricalOptimisticSource.selfConsistentScheduledCausalSource_inverseSqrtThresholdUnboundedHittingAfter_stoppedAverageRealizedBehaviorRegretExpectedPositivePart_tendsto_zero Compiled

The expected positive part of the exact stopped average realized behavior regret tends to zero with the inverse-square-root threshold schedule.

theorem selfConsistentScheduledCausalSource_inverseSqrtThresholdUnboundedHittingAfter_stoppedAverageRealizedBehaviorRegretExpectedPositivePart_tendsto_zero (mdp : MDP State Action) (initialState : Measure State) [IsProbabilityMeasure initialState] [StandardBorelSpace State] [StandardBorelSpace Action] (rewardSource : mdp.MeanCompatibleRewardKernel) (varianceProxy : NNReal) (hvarianceProxy : 0 < varianceProxy) (law : rewardSource.UniformSubgaussianRewardLaw varianceProxy) (initialTable : DeterministicMarkovPolicyTable mdp) (defaultState : State) (support : ExploratoryPathSupport mdp initialState) (baseVisitFloor : Real) (hbaseFloor : ExploratoryPathUniformVisitFloor support 1 baseVisitFloor) (hrewardBound : forall state action, |mdp.reward state action| <= 1) (hhorizon : 4 < mdp.horizon) (hbaseVisitFloor : 0 < baseVisitFloor) : Tendsto (selfConsistentScheduledNaturalCausalInverseSqrtThresholdUnboundedHittingAfterStoppedAverageRealizedBehaviorRegretExpectedPositivePart mdp initialState rewardSource initialTable defaultState varianceProxy baseVisitFloor) atTop (nhds 0)