Bousquet eq. (6): for
Provedbousquet_phi_convexity_per_summand_scalar_boundanalysisconcentrationconvexityentropy-method
Bousquet 2002 eq(6) per-summand scalar core. The function phi(x)=e^x-x-1 is convex with phi(0)=0, so by convexity through the origin, for 0<=y<=1, phi(-lamy) <= yphi(-lam). Applied to the modified-LSI summand phi(-lam*(Z-Z_k)) with 0<=Z-Z_k<=1 this separates the telescoping (Z-Z_k) part (which sums to E[Z] via condition (3)) from the variance part.
Preamble
import Mathlib.Analysis.SpecialFunctions.Exp import Mathlib.Analysis.SpecialFunctions.ExpDeriv import Mathlib.Analysis.Convex.SpecificFunctions.Basic import Mathlib.Tactic open Real
Formal statement
theorem bousquet_phi_convexity_per_summand_scalar_bound (lam y : ℝ) (hy0 : 0 ≤ y) (hy1 : y ≤ 1) :
(Real.exp (-(lam * y)) - (-(lam * y)) - 1)
≤ y * (Real.exp (-lam) - (-lam) - 1) := by
sorry