Scalar Jensen inequality applied to an operator
ProvedQAlgorithms.Papers.Independent.Support50.P23.scalarJensenOperatorLet be real, let be convex on , let be Hermitian on a finite-dimensional nonzero complex Hilbert space with , and let . Functional calculus gives
Both matrix elements are real. The unit-norm condition repairs the printed preamble, which omits it, and matches the normalized-purification applications on pages 5 and 8. The uncorrected statement fails for , , . This is the scalar convexity tool used in the entropy estimates.
import Definitions.Def_QAlgorithms_Papers_Independent_Support50_P23
noncomputable section
open Filter
open scoped ENNReal ComplexConjugate ComplexOrder
namespace QAlgorithms.Papers.Independent.Support50.P23
attribute [local instance] Classical.propDecidable
-- Exact accepted bodies. Other similarly named catalogue entries use different
-- smoothing, conditioning order, logarithm base, or channel-level objects.
open QAlgorithms.Papers.Independent.Gap20.P03 (fixedMin)
open QAlgorithms.Papers.Independent.Dependency50.P36 (relativeEntropy)
open QAlgorithms.Papers.Independent.Information100.P67 (IsSubnormalized)
open QAlgorithms.Papers.Independent.Information100.P69 (matrixPower)
variable {A B C D I : Type*}
variable [Fintype A] [Fintype B] [Fintype C] [Fintype D] [Fintype I]
variable [DecidableEq A] [DecidableEq B] [DecidableEq C] [DecidableEq D] [DecidableEq I]
/-- Lemma 11, PDF p.6. Printed/changed/why: add norm-one phi, as used for the
normalized purifications on pp.5,8. The literal preamble allows f=1,X=0,phi=0,
which gives 0>=1. Hermiticity is implicit in the source's Loewner comparisons. -/
theorem scalarJensenOperator [Nonempty I]
(a b : ℝ) (hab : a ≤ b) (f : ℝ → ℝ) (hf : ConvexOn ℝ (Set.Icc a b) f)
(X : Matrix I I ℂ) (hX : X.IsHermitian)
(hLower : (X - (a : ℂ) • (1 : Matrix I I ℂ)).PosSemidef)
(hUpper : ((b : ℂ) • (1 : Matrix I I ℂ) - X).PosSemidef)
(phi : EuclideanSpace ℂ I) (hphi : ‖phi‖ = 1) :
f ((inner ℂ phi (act X phi)).re) ≤ (inner ℂ phi (act (cfc f X) phi)).re := by sorry
end QAlgorithms.Papers.Independent.Support50.P23
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For every nonempty finite type with decidable equality, real numbers , function convex on the closed interval , Hermitian matrix on satisfying and , and vector with Euclidean norm , one has
where matrix action is ordinary matrix–vector multiplication and is defined by the continuous functional calculus. The interval may be a singleton, since is allowed; no differentiability hypothesis on is imposed.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.