Exact original-dual point recovery by rational coordinate rounding
ProvedInfoLossLAR.WorstCase.IntegerBasis.roundedPoint_eqbddexact-geometryinteger-latticeregev
If the full Euclidean query error times each original input-column norm is strictly less than one half, the actual rational-coordinate nearest-integer reconstruction equals the specified genuine dual-lattice point. The hypothesis controls the complete error norm, not merely individual coordinate approximations. This is an exact stopping/recovery identity and supplies neither a BDD oracle nor its success guarantee.
Preamble
import Definitions.Def_InfoLossLAR_BDD_Coordinates set_option autoImplicit false
Formal statement
namespace InfoLossLAR.WorstCase.IntegerBasis
open RegevLWE.GaussConv GaussianProjection GaussianProjection.CosetGeometry
theorem roundedPoint_eq {dimension : ℕ} (basis : Matrix (Fin dimension) (Fin dimension) ℤ) (hdet : basis.det≠0)
(query : Fin dimension → ℚ) (nearest : dual (inputLattice basis hdet))
(hsmall : ∀ index : Fin dimension,
‖rationalPoint query-(nearest : Space (dimension:=dimension))‖*
‖(integerBasis basis hdet index : Space (dimension:=dimension))‖ < 1/2) :
roundedPoint basis hdet query=nearest := by sorry
end InfoLossLAR.WorstCase.IntegerBasisSource
Complete existing checked theorem InfoLossLAR.WorstCase.IntegerBasis.roundedPoint_eq and its minimal proof-only helpers, extracted by official Lean declaration and syntax/info-tree oracles from the frozen BDD geometry modules. All helper and root proof bodies are exact source byte slices. Only the root binding is renamed to global solution and its surrounding dimension parameter is exposed explicitly. Shared objects use the pure original-basis coordinate definition provider. No new mathematical argument.