§2.4.2: inverse and composition of partial-unitary approximations
OpenQAlgorithms.Papers.Independent.Intakethird100.P03.partial_precisionFor partial unitaries, inversion preserves the source precision relation, including both domain and range inclusions. For two compatible chains with component precisions δ_j ≥ 0, specify initial domains N ⊆ Ntilde on which both chains are valid. In addition, require that the image of N under the exact composite is contained in the image of Ntilde under the approximate composite. Then the composites satisfy the full precision relation with precision ∑_j δ_j: both inclusions and the uniform vector bound ‖V_comp x − U_comp x‖ ≤ (∑_j δ_j)‖x‖ on N. The empty chain is the identity. The final-range assumption corrects a defect in the source extension; compatibility and factorwise range inclusions alone do not imply it.
import Definitions.Def_QAlgorithms_Papers_Independent_Intakethird100_P03
noncomputable section
open scoped Matrix ComplexConjugate Kronecker
namespace QAlgorithms.Papers.Independent.Intakethird100.P03
open QAlgorithms.Papers.Independent.Renew46.P01 (cleanVector)
attribute [local instance] Classical.propDecidable
/-- PDF p.12: inversion preserves precision. Source correction for composition:
printed "Lemma 3 remains valid for partial unitary operators" omits a necessary
composite-range inclusion. We require that inclusion explicitly, while retaining
both component inclusions, valid chains and the full precision conclusion.
Reason: I on a line followed by I versus a small full-space rotation has close
factors but distinct composite range lines. Inversion is unchanged. -/
theorem partial_precision {I : Type*} [Fintype I] :
(∀ (U V : PartialUnitary I) (d : ℝ), 0 ≤ d → PartialApprox V U d →
PartialApprox V.inverse U.inverse d) ∧
(∀ (L : ℕ) (U V : Fin L → PartialUnitary I) (d : Fin L → ℝ)
(N Ntilde : Submodule ℂ (EuclideanSpace ℂ I)),
(∀ j, 0 ≤ d j) → (∀ j, PartialApprox (V j) (U j) (d j)) →
N ≤ Ntilde →
(∀ j : Fin L, j.val = 0 → N ≤ (U j).domain) →
(∀ j : Fin L, j.val = 0 → Ntilde ≤ (V j).domain) →
(∀ i j : Fin L, j.val = i.val + 1 → (U i).range ≤ (U j).domain) →
(∀ i j : Fin L, j.val = i.val + 1 → (V i).range ≤ (V j).domain) →
(∀ x ∈ N, ∃ y ∈ Ntilde,
partialRun (List.ofFn U) x = partialRun (List.ofFn V) y) →
PartialSequenceApprox (List.ofFn V) (List.ofFn U) N Ntilde (∑ j, d j)) := by sorry
end QAlgorithms.Papers.Independent.Intakethird100.P03
end
Read-back
What the Lean code literally says, in plain math · gpt-6.1-sol
For every finite index type , possibly empty, put with its Euclidean norm. Write a partial unitary as an isometric complex linear isomorphism , and similarly write as . Say explicitly that approximates with bound when , , and for every . The theorem asserts two claims together. First, for all partial unitaries and all real , this approximation implies the corresponding approximation of their inverses: , , and for every . Second, for every natural number , every two families , of partial unitaries, every family of real numbers , and every pair of subspaces , assume that each , each approximates with bound as just expanded, , and whenever , and and whenever . Also assume that for every there exists such that running the in increasing index order on equals running the in increasing index order on . Then , that same existential equality of runs holds, and
Here each run applies the relevant isomorphism after orthogonally projecting onto its domain at every step. The case is included: both runs are identities, the sum is zero, and the first-index and consecutive-index assumptions are vacuous.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.