Law of a block of a Gaussian matrix under an orthonormal frame
ProvedGaussianMatrix.block_lawLet be a standard Gaussian matrix and let have orthonormal columns, . Then
i.e. is a standard Gaussian matrix.
In the analysis of the randomized SVD one writes the right singular vectors of the input as and works with ; this statement says is itself standard Gaussian, so the moment identities and singular-value tail bounds for Gaussian matrices apply to it directly.
Formalization Note. The statement is an equality of measures: the pushforward of under equals . The hypothesis forces .
import Definitions.Def_GaussianMatrix_basic open MeasureTheory ProbabilityTheory open scoped Matrix
namespace GaussianMatrix
theorem block_law {n k t : ℕ} (V₁ : Matrix (Fin n) (Fin k) ℝ) (hV₁ : V₁ᵀ * V₁ = 1) :
Measure.map (fun G : Fin n → Fin t → ℝ => Matrix.of.symm (V₁ᵀ * Matrix.of G))
(gaussianMatrix n t) = gaussianMatrix k t := by sorry
end GaussianMatrixRead-back
What the Lean code literally says, in plain math · claude-fable-5-1
Let be natural numbers (implicit; any may be ) and let be a real matrix with
(orthonormal columns). Let be the law of an array with independent entries (product of copies of ), and the analogous law for arrays. The statement asserts the equality of measures on the space of real arrays
i.e. the pushforward of under the map taking an array (viewed as a matrix) to the matrix (viewed as an array) equals . Mathlib's pushforward would be the zero measure for a non-a.e.-measurable map, but this map is linear, hence measurable, so that convention is not triggered. Edge cases and vacuity: the hypothesis forces (the rank of is at most ), so for no such exists and the statement is vacuous. If the hypothesis is the trivial identity of matrices, satisfied by the unique matrix, and is the point mass on the one-point array space. If both sides are point masses on one-point spaces. If then must be for the hypothesis to hold.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.