§5.1, variational identity and profile (41)–(42)
OpenQAlgorithms.Papers.Independent.Intakenext100.P45.classicalVariationalFor finite equivalent probability laws , define . For every ,
where ranges over probability laws. An optimizer exists: it is if ; otherwise it lies on and . The finite-valued function is convex and non-increasing, with and . This identifies the variational error profile.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P45
noncomputable section
namespace QAlgorithms.Papers.Independent.Intakenext100.P45
open QAlgorithms Filter
open scoped Matrix ComplexConjugate ComplexOrder ENNReal BigOperators Topology
open QAlgorithms.Papers.Independent.Information100.P07
(support positivePart positiveProjection IsProjection)
open QAlgorithms.Papers.Independent.Information100.P56 (IsLaw)
open QAlgorithms.Papers.Independent.Information100.P60 (Mat copies)
open QAlgorithms.Papers.Independent.Dependency50.P28 (logNonnegativeExtended relativeEntropy)
open QAlgorithms.Papers.Independent.Claimed50.P15 (positiveMatrixRpow)
attribute [local instance] Classical.propDecidable Classical.decEq
variable {I J : Type} [Fintype I] [DecidableEq I] [Fintype J] [DecidableEq J]
/-- arXiv:0708.4282v1, PDF pp.20,21: §5.1, variational identity and profile (41)–(42). -/
theorem classicalVariational (p q : I → ℝ) (hp : IsLaw p) (hq : IsLaw q) (heq : SamePositiveSupport p q) :
(∀ r : ℝ, 0 ≤ r →
classicalExponent p q r = sInf {v : EReal | ∃ x : I → ℝ,
IsLaw x ∧ kl x q ≤ (r : EReal) ∧ v=kl x p} ∧
∃ x : I → ℝ, IsLaw x ∧ kl x q ≤ (r : EReal) ∧
classicalExponent p q r=kl x p ∧
((kl p q ≤ (r : EReal) → x=p ∧ classicalExponent p q r=0) ∧
((r : EReal) < kl p q → kl x q=(r : EReal) ∧ 0 < classicalExponent p q r))) ∧
classicalExponent p q 0=kl q p ∧ classicalExponent p q (kl p q).toReal=0 ∧
AntitoneOn (classicalExponent p q) (Set.Ici 0) ∧
ConvexOn ℝ (Set.Ici 0) (fun r => (classicalExponent p q r).toReal) ∧
(∀ r : ℝ, 0 ≤ r → classicalExponent p q r ≠ ⊤ ∧ classicalExponent p q r ≠ ⊥) := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P45
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For every finite index type with decidable equality, let be nonnegative functions summing to and satisfying if and only if for every . Define , using total real powers, and let for and otherwise. Define the extended-real function
For real functions , define if some has , and otherwise define . For every real ,
and there exists such a probability law attaining this infimum. This attaining also satisfies: if , then and ; if , then and . In addition, , , is nonincreasing on , the real-valued function is convex there, and is neither nor for any ; here the subscript “real” denotes conversion from extended reals to reals. Common zero coordinates and are allowed, and the theorem does not assert uniqueness of the attaining law .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.