import Determinize.Spec.FiniteModel.Modelnamespace Determinize.Spec.FiniteModel/-- Expected reward returned within `steps` transitions; unfinished paths contribute zero. -/def Model.rewardWithin (model : Model) : Nat → Fin model.size → Rat | 0, state => match model.kind state with | .returned reward => reward | _ => 0 | steps + 1, state => match model.kind state with | .returned reward => reward | .rejected => 0 | .transient => ∑ next, model.transition state next * model.rewardWithin steps nextend Determinize.Spec.FiniteModelnamespace Determinize.Proof.FiniteModelopen Spec.FiniteModel MeasureTheory/-- Increasing the exploration horizon only adds returned output mass. -/theorem outputWithin_mono (model : Model) (state : Fin model.size) : Monotone (fun steps ↦ model.outputWithin steps state) := by apply monotone_nat_of_le_succ intro steps induction steps generalizing state with | zero => cases h : model.kind state <;> simp only [Model.outputWithin, h, Std.le_refl] exact bot_le | succ steps ih => cases h : model.kind state with | returned reward => simp [Model.outputWithin, h] | rejected => simp [Model.outputWithin, h] | transient => simp only [Model.outputWithin, h] apply Finset.sum_le_sum intro next _ exact smul_le_smul_left _ (ih next)theorem returned_outputWithin (model : Model) (state : Fin model.size) (reward : Rat) (returned : model.kind state = .returned reward) (steps : Nat) : model.outputWithin steps state = Measure.dirac (reward : ℝ) := by cases steps <;> simp [Model.outputWithin, returned]theorem rejected_outputWithin (model : Model) (state : Fin model.size) (rejected : model.kind state = .rejected) (steps : Nat) : model.outputWithin steps state = 0 := by cases steps <;> simp [Model.outputWithin, rejected]theorem returned_output (model : Model) (reward : Rat) (returned : model.kind model.initial = .returned reward) : model.outputMeasure = Measure.dirac (reward : ℝ) := by simp [Model.outputMeasure, returned_outputWithin model model.initial reward returned]theorem rejected_output (model : Model) (rejected : model.kind model.initial = .rejected) : model.outputMeasure = 0 := by simp [Model.outputMeasure, rejected_outputWithin model model.initial rejected]/-- Every finite-horizon output law has finite support. -/theorem integrable_outputWithin (model : Model) (steps : Nat) (state : Fin model.size) (f : ℝ → ℝ) : Integrable f (model.outputWithin steps state) := by induction steps generalizing state with | zero => cases h : model.kind state <;> simp only [Model.outputWithin, h, integrable_zero_measure] exact integrable_dirac (by simp) | succ steps ih => cases h : model.kind state with | returned reward => simpa [Model.outputWithin, h] using (integrable_dirac (by simp) : Integrable f (Measure.dirac (reward : ℝ))) | rejected => simp [Model.outputWithin, h] | transient => simp only [Model.outputWithin, h] apply integrable_finsetSum_measure.mpr intro next _ exact (ih next).smul_measure (by simp)/-- Exact rational reward iteration agrees with the real-measure interpretation. -/theorem rewardWithin_integral (model : Model) (steps : Nat) (state : Fin model.size) : (∫ value : ℝ, value ∂model.outputWithin steps state) = (model.rewardWithin steps state : ℝ) := by induction steps generalizing state with | zero => cases h : model.kind state <;> simp [Model.outputWithin, Model.rewardWithin, h] | succ steps ih => cases h : model.kind state with | returned reward => simp [Model.outputWithin, Model.rewardWithin, h] | rejected => simp [Model.outputWithin, Model.rewardWithin, h] | transient => simp only [Model.outputWithin, Model.rewardWithin, h] rw [integral_finsetSum_measure (s := Finset.univ) (μ := fun next ↦ ENNReal.ofReal (model.transition state next : ℝ) • model.outputWithin steps next) (fun next _ ↦ (integrable_outputWithin model steps next (fun x ↦ x)).smul_measure (by simp))] simp only [integral_smul_measure, ih, Rat.cast_sum, Rat.cast_mul, smul_eq_mul] apply Finset.sum_congr rfl intro next _ rw [ENNReal.toReal_ofReal (by exact_mod_cast model.nonnegative state next)]end Determinize.Proof.FiniteModel