§IV (42)–(44): expanded Lagrange dual and weak-dual bound
ProvedQAlgorithms.Papers.Independent.Support50.P11.expanded_dual_weak_boundLet be finite CPTP, , and select NS and/or PPTp with flags . Fix , Hermitian on , and PSD multipliers on . Require without NS and without PPTp. The expanded dual has objective and constraints
These constraints are equivalent to the source's unreduced Lagrangian being at most that objective for every PSD pair , without normalization imposed on . Each dual feasible tuple satisfies
for every primal feasible pair and also bounds . This is the genuine expanded Lagrange dual and weak-dual bound; remain separate variables.
import Definitions.Def_QAlgorithms_Papers_Independent_Support50_P11
noncomputable section
open scoped Matrix ComplexConjugate ComplexOrder ENNReal
namespace QAlgorithms.Papers.Independent.Support50.P11
open QAlgorithms.Papers.Independent.Information100.P34 (choi dephasing permutationMatrix)
open QAlgorithms.Papers.Independent.Information100.P21 (pt)
open QAlgorithms.Papers.Independent.Dependency50.P09 (tensorAction)
open QAlgorithms.Papers.Independent.Information100.P76 (swapMatrix)
open QAlgorithms.Papers.Independent.Information100.P69 (replacer)
open QAlgorithms.Papers.Independent.Gap20.P19 (wernerIsometry)
attribute [local instance] Classical.propDecidable
attribute [local instance] Classical.decEq
section Statements
variable {I O : Type} [Fintype I] [Nonempty I] [Fintype O] [Nonempty O]
/-- PDF pp.7--8, (42)--(44): every expanded-dual feasible point bounds every
primal point and the operational optimum. Y,V remain independent PSD multipliers;
without PPT they both vanish, and without NS the Hermitian W vanishes. -/
theorem expanded_dual_weak_bound
(N : Superop I O) (hN : IsChannel N) (K : ℕ) (hK : 0 < K)
(s p : Bool) (hflags : s = true ∨ p = true) (mu : ℝ) (W : Matrix O O ℂ)
(X Y V : Matrix (I × O) (I × O) ℂ)
(hW : W.IsHermitian) (hX : X.PosSemidef) (hY : Y.PosSemidef) (hV : V.PosSemidef)
(hzeroW : s = false → W = 0) (hzeroYV : p = false → Y = 0 ∧ V = 0) :
(ExpandedDualFeasible N K s p mu W X Y V ↔
∀ (rho : Matrix I I ℂ) (L : Matrix (I × O) (I × O) ℂ), rho.PosSemidef → L.PosSemidef →
lagrangian N K mu W X Y V L rho ≤ mu + W.trace.re / (K : ℝ)^2) ∧
(ExpandedDualFeasible N K s p mu W X Y V →
(∀ rho L, Primal K s p rho L → primalObjective N L ≤ mu + W.trace.re / (K : ℝ)^2) ∧
optimalCodeFidelity (flagsClass s p) N K ≤ mu + W.trace.re / (K : ℝ)^2) := by sorry
end Statements
end QAlgorithms.Papers.Independent.Support50.P11
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
Let be the finite input and output index types, and let be a channel, meaning that for some finite family satisfying . Fix a positive integer , Boolean flags with at least one true, a real number , a Hermitian matrix on , and positive semidefinite matrices on ; assume if is false and if is false. Write for the matrix with entries , where is a matrix unit; write for partial transposition on the second factor, , and use for the corresponding partial traces. Under these assumptions, the two inequalities
are equivalent to the following bound for every positive semidefinite on and on , without requiring either matrix to have trace one:
Moreover, these inequalities imply for every pair satisfying , , , additionally when is true, and additionally when is true. They also bound the real supremum of the following code fidelities by the same number: range over channels for which there exists a channel with for every ; if is true also require a channel with , and if is true also require to be a channel. For each such , close its indices through to form
and take the fidelity , where and . The quantifier includes ; here means that the difference is positive semidefinite.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.