Concavity, tagged measurement affinity, and minimax
OpenQAlgorithms.Papers.Independent.Intakenext100.P63.accessible_concavity_affinity_minimaxFix normalized states on a nonzero finite-dimensional space , with a finite nonempty label set . For a prior and finite POVM , let for the law . Let be probability vectors, let , and let be POVMs. Their tagged mixture has disjoint outcome sets and effects and . Then
The extrema are attained, including the inner minimum for each prior. Entropies are in bits, with zero-probability terms contributing zero. Retaining the random branch tag is essential to the affinity assertion.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakenext100_P63
noncomputable section
namespace QAlgorithms.Papers.Independent.Intakenext100.P63
open QAlgorithms Filter
open QAlgorithms.Papers.Independent.Information100.P63 (Mat)
open QAlgorithms.Papers.Independent.Information100.P04 (Measurement)
open QAlgorithms.Papers.Independent.Information100.P27 (shannon labelInformation)
open QAlgorithms.Papers.Independent.Information100.P71 (channelPower dilationMap SupportIncluded)
open QAlgorithms.Papers.Independent.Bridge50.P47 (SymmetricIndex universalSymmetricDilation)
open QAlgorithms.Learning (IsDistribution)
open scoped Matrix ComplexConjugate ComplexOrder
attribute [local instance] Classical.propDecidable Classical.decEq
/-- PDF p.8, Lemma 5: concavity in priors, tagged-measurement affinity, and minimax.
The displayed max/min are attained; they are not replaced by unattained bounds. -/
theorem accessible_concavity_affinity_minimax {I E J K : Type}
[Fintype I] [Nonempty I] [Fintype E] [DecidableEq E] [Nonempty E]
[Fintype J] [Fintype K]
(rho : I → Mat E) (hrho : ∀ i, IsDensityMatrix (rho i))
(p0 p1 : I → ℝ) (hp0 : IsDistribution p0) (hp1 : IsDistribution p1)
(Q : J → Mat E) (R : K → Mat E) (hQ : IsPOVM Q) (hR : IsPOVM R)
(lambda : ℝ) (hl0 : 0 ≤ lambda) (hl1 : lambda ≤ 1) :
(lambda * ensembleConditionalBits p0 rho Q + (1 - lambda) * ensembleConditionalBits p1 rho Q ≤
ensembleConditionalBits (fun i => lambda * p0 i + (1 - lambda) * p1 i) rho Q) ∧
(ensembleConditionalBits p0 rho (taggedEffects lambda Q R) =
lambda * ensembleConditionalBits p0 rho Q + (1 - lambda) * ensembleConditionalBits p0 rho R) ∧
(optimizedEquivocation rho = sInf {h | ∃ M : Measurement E, h = measurementWorstPrior rho M}) ∧
(∃ p : I → ℝ, IsDistribution p ∧ accessibleEquivocation p rho = optimizedEquivocation rho) ∧
(∀ p : I → ℝ, IsDistribution p → ∃ M : Measurement E,
ensembleConditionalBits p rho M.effect = accessibleEquivocation p rho) ∧
(∃ M : Measurement E, measurementWorstPrior rho M = optimizedEquivocation rho) := by sorry
end QAlgorithms.Papers.Independent.Intakenext100.P63
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let be finite nonempty types, with decidable equality on , and let be finite types. Let be positive semidefinite complex matrices on with trace , let be nonnegative functions summing to , let and consist of positive semidefinite matrices summing to the identity, and let . Write
Define , where ranges over measurements with any positive finite number of outcomes, define , and define , with both suprema over all nonnegative priors on summing to . The tagged measurement , indexed by the disjoint union , has effects and . The declaration asserts all of
together with the existence of a probability prior satisfying , the existence for every probability prior of a measurement satisfying , and the existence of a measurement satisfying . All logarithms and quotients in are total real operations, so zero-probability terms use and division by zero equal to zero. The endpoints are included, retaining zero effects on the discarded tagged branch; although may syntactically be empty, their measurement hypotheses cannot hold on nonempty in that case.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.