Scalar Gibbs inequality:
ProvedScalarGibbs.mul_log_sub_mul_log_sub_sub_nonneganalysisentropy-methodinformation-theory
Scalar Gibbs / Bregman inequality for . For and , (equivalently ). This is the pointwise variational ingredient behind the variational formula of entropy (Boucheron–Lugosi–Massart, Concentration Inequalities, OUP 2013, Theorem 4.13) and its use in the modified logarithmic Sobolev inequality (Theorem 6.6): applied conditionally with , it produces the per-coordinate summand with .
Preamble
import Mathlib.Analysis.SpecialFunctions.Log.Basic import Mathlib.Analysis.SpecialFunctions.Log.NegMulLog import Mathlib.Analysis.SpecialFunctions.Exp open Real
Formal statement
theorem ScalarGibbs.mul_log_sub_mul_log_sub_sub_nonneg {y u : ℝ} (hy : 0 ≤ y) (hu : 0 < u) :
0 ≤ y * Real.log y - y * Real.log u - (y - u) := by sorrySource
Boucheron, Lugosi, Massart, Concentration Inequalities (OUP 2013), Theorem 4.13 (variational formula of entropy) and Theorem 6.6 (modified logarithmic Sobolev inequality), Section 6.3.