§6.11, Definition — Minimum-norm least-squares pseudoinverse
DefinitionVectorSpaceOpt_pseudoinverse_predicateLet and be real normed spaces, let be continuous linear, and let be continuous linear. Define IsPseudoinverse A B to mean that for every target , the vector first minimizes residual norm and then has minimum norm among all residual minimizers:
The two-stage condition covers inconsistent as well as underdetermined equations. It is the shared interface for the existence theorem and all operator identities in the mission, while the continuous-linear-map type records the source's linearity and boundedness requirements.
import Mathlib
namespace VectorSpaceOpt
/-- `B` is the minimum-norm least-squares pseudoinverse of `A` when, for each target `y`,
`B y` minimizes the residual norm and has minimum norm among all residual minimizers. -/
def IsPseudoinverse
{G H : Type*} [NormedAddCommGroup G] [NormedSpace ℝ G]
[NormedAddCommGroup H] [NormedSpace ℝ H]
(A : G →L[ℝ] H) (B : H →L[ℝ] G) : Prop :=
∀ y : H,
(∀ x : G, ‖A (B y) - y‖ ≤ ‖A x - y‖) ∧
(∀ x : G, (∀ z : G, ‖A x - y‖ ≤ ‖A z - y‖) → ‖B y‖ ≤ ‖x‖)
end VectorSpaceOptRead-back
What the Lean code literally says, in plain math · gpt-5
For arbitrary real normed vector spaces and , a continuous real-linear map , and a continuous real-linear map , is defined to mean the following for every target : first, for every ; and second, for every , if for every , then . Thus is a global residual-norm minimizer and has norm no larger than any other global residual-norm minimizer. The definition quantifies over every target and every competitor, does not assert uniqueness of either kind of minimizer, and assumes neither inner products, completeness, nor nontriviality; continuity and linearity of the total assignment are built into the type of .
Confirmed by the mission captain (proposal self-audit).