Generalized formation equals the coordinate convex roof
ProvedQAlgorithms.Papers.Independent.Support50.P35.formation_coordinate_roofLet , let be a continuous unitarily invariant concave real function on density matrices on , and let be a normalized bipartite density matrix. Set for unit vectors. Choose any complete Hermitian orthonormal bipartite operator basis , let , and let be its normalized pure-projector quadratic variety. Define . Then
Both sides range over every positive finite length, nonnegative weights of total one, and normalized components. More precisely, their feasible sets of average costs are equal, and for each normalized finite pure ensemble,
This transfers the actual optimization to affine coordinates. It asserts no additivity or unconditional operational cost interpretation.
import Definitions.Def_QAlgorithms_Papers_Independent_Support50_P35
noncomputable section
open scoped Matrix ComplexConjugate Topology Manifold
namespace QAlgorithms.Papers.Independent.Support50.P35
/-- PDF p.19: generalized formation is exactly a finite coordinate roof;
feasible values and normalized finite barycenters correspond in both directions.
This asserts no unconditional operational interpretation or tensor additivity. -/
theorem formation_coordinate_roof (M N : ℕ) (hN : 0 < N) (hNM : N ≤ M)
(f : Matrix (Fin M) (Fin M) ℂ → ℝ) (hf : IsMarginalCost f)
(mu : Fin ((M * N) ^ 2) → Matrix (Fin M × Fin N) (Fin M × Fin N) ℂ)
(hmu : IsHermitianONBasis mu)
(rho : Matrix (Fin M × Fin N) (Fin M × Fin N) ℂ) (hrho : IsDensityMatrix rho) :
formationValues f rho =
roofValues (generalPureSet mu) (coordinateCost mu f) (hermitianCoordinates mu rho) ∧
QAlgorithms.Papers.Independent.Information100.P03.convexRoof (fun psi => f (traceRight (pureDensity psi))) rho =
lowerRoof (generalPureSet mu) (coordinateCost mu f) (hermitianCoordinates mu rho) ∧
(∀ e : QAlgorithms.Papers.Independent.Information100.P03.PureEnsemble (Fin M × Fin N),
ensembleDensity e.weight e.vector = rho ↔
(∑ j, e.weight j • hermitianCoordinates mu (pureDensity (e.vector j))) = hermitianCoordinates mu rho) := by sorry
end QAlgorithms.Papers.Independent.Support50.P35
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For all natural numbers with , let be a real-valued function on complex matrices that is continuous and concave on the convex set of positive semidefinite trace- matrices and satisfies for every such and every unitary . Let be Hermitian matrices on satisfying and spanning all Hermitian matrices over the reals, and let be positive semidefinite with trace . Define
and let consist of the coordinates satisfying
Then the set of values , over all finite families of norm- vectors and nonnegative weights summing to with , equals the set of values , over all finite families with a positive number of terms, , nonnegative weights summing to , and . The real infima of these two sets are also equal. Finally, for every finite family of norm- vectors and nonnegative weights summing to ,
Zero weights and repeated vectors or coordinate points are allowed; a family with no terms cannot have total weight .
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.