§5.8, Theorem 1 — Minimum-distance duality for a subspace
ProvedVectorSpaceOpt.minimum_distance_dualityLet be a real normed space, any real linear subspace, and . Write . There exists a continuous linear functional with and such that
Hence attains the dual maximum and certifies the primal infimum without closedness or attainment assumptions. If is an actual best approximant, the same witness also satisfies . The theorem is the mission's capstone strong-duality certificate and generalizes Hilbert projection geometry to arbitrary normed spaces.
import Mathlib
namespace VectorSpaceOpt
theorem minimum_distance_duality
{X : Type*} [NormedAddCommGroup X] [NormedSpace ℝ X]
(M : Submodule ℝ X) (x : X) :
∃ f : X →L[ℝ] ℝ,
‖f‖ ≤ 1 ∧
(∀ m ∈ M, f m = 0) ∧
f x = Metric.infDist x (M : Set X) ∧
(∀ g : X →L[ℝ] ℝ, ‖g‖ ≤ 1 → (∀ m ∈ M, g m = 0) →
g x ≤ Metric.infDist x (M : Set X)) ∧
(∀ m₀ ∈ M, (∀ m ∈ M, ‖x - m₀‖ ≤ ‖x - m‖) →
f (x - m₀) = ‖f‖ * ‖x - m₀‖) := by sorry
end VectorSpaceOptRead-back
What the Lean code literally says, in plain math · gpt-5
Let be any real normed vector space, let be any real linear subspace of , and let ; neither completeness, nontriviality, nor closedness of is assumed. There exists a continuous real-linear functional such that simultaneously: ; for every ; equals the metric infimum distance from to the underlying set of ; every continuous real-linear with and for all satisfies the one-sided inequality ; and, for every satisfying for every , one has . A subspace is nonempty, so this infimum distance is taken to a nonempty set, but it may be zero even when because need not be closed; the last clause is vacuous if no nearest point exists.
Confirmed by the mission captain (proposal self-audit).