Coordinate oscillations on finite product spaces
DefinitionInfoLossLAR_Native_AFT_FiniteConcentrationfinite-probabilityinfo-loss-lar-native-5c3becbstoc-2027
Let and let be real bounds. The bounded differences condition is
The accompanying expectation identity splits an independent product into its first coordinate and remaining coordinates. This is the finite product setup for concentration estimates.
Definition code
import Definitions.Def_InfoLossLAR_Native_AFT_BlockFock
import Definitions.Def_InfoLossLAR_Native_AFT_Hoeffding
import Mathlib
/-! # Dimension-free concentration on actual finite product laws -/
namespace AFT.FiniteConcentration
open scoped BigOperators Classical
noncomputable section
variable {Color : Type*} [Fintype Color]
theorem expectation_cons (count : ℕ) (law : Fin (count + 1) → FiniteDist Color)
(value : (Fin (count + 1) → Color) → ℝ) :
(FiniteDist.independent law).expectation value =
(law 0).expectation (fun first =>
(FiniteDist.independent (fun index => law index.succ)).expectation
(fun rest => value (Fin.cons first rest))) := by
unfold FiniteDist.expectation FiniteDist.independent
rw [← Equiv.sum_comp (Fin.consEquiv (fun _ : Fin (count + 1) => Color))]
rw [Fintype.sum_prod_type]
apply Finset.sum_congr rfl
intro first _
rw [Finset.mul_sum]
apply Finset.sum_congr rfl
intro rest _
change (∏ index, (law index).weight ((Fin.cons first rest : Fin (count + 1) → Color) index)) * value (Fin.cons first rest) = _
simp only [Fin.prod_univ_succ, Fin.cons_zero, Fin.cons_succ]
ring
/-- Coordinate oscillations, with all other sample values fixed. -/
def BoundedDifferences (count : ℕ) (value : (Fin count → Color) → ℝ) (bound : Fin count → ℝ) : Prop :=
∀ index first second, (∀ other, other ≠ index → first other = second other) →
|value first - value second| ≤ bound index
end
end AFT.FiniteConcentration
Source
Local manuscript repository commit 5c3becb254daaa8738368b474349766c1769c2a0; lean/AFT/FiniteConcentration.lean; migrated source SHA256 f873907aeff4ff23ad9d882a6430d2d9acf05ae68ec17df93c36f4bb2cfd86a7; byte-range skeleton from Lean AST, with original retained definitions and proofs.