Rotational invariance of the standard Gaussian matrix: for orthogonal
ProvedGaussianMatrix.rotation_invarianceLet be a standard Gaussian matrix (independent entries, law ), and let and be orthogonal matrices, and . Then the matrix is again a standard Gaussian matrix:
This is the basic symmetry of the Gaussian ensemble. It lets one reduce computations involving fixed matrices to the case of diagonal and , and it is the engine behind every conditioning argument in the analysis of randomized sketching: multiplying a Gaussian sketch by an orthogonal factor of the input matrix does not change its law.
Formalization Note. The statement is an equality of measures: the pushforward of under equals .
import Definitions.Def_GaussianMatrix_basic open MeasureTheory ProbabilityTheory open scoped Matrix
namespace GaussianMatrix
theorem rotation_invariance {p m : ℕ} (U : Matrix (Fin p) (Fin p) ℝ) (V : Matrix (Fin m) (Fin m) ℝ)
(hU : Uᵀ * U = 1) (hV : Vᵀ * V = 1) :
Measure.map (fun G : Fin p → Fin m → ℝ => Matrix.of.symm (U * Matrix.of G * V))
(gaussianMatrix p m) = gaussianMatrix p m := by sorry
end GaussianMatrixRead-back
What the Lean code literally says, in plain math · claude-fable-5-1
Let be natural numbers (implicit; either may be ), let and be real matrices, and assume
(each has orthonormal columns; since they are square this means each is an orthogonal matrix). Let denote the law of a array with independent entries, i.e. the product measure on real arrays. The statement asserts the equality of measures on the space of real arrays
i.e. the pushforward (image measure) of under the map sending the array (viewed as a matrix) to the matrix product (viewed again as an array) is itself. Equality of measures means equality on every measurable set of the product sigma-algebra. Caveats on the pushforward: Mathlib's image measure is defined to be the zero measure if the map is not almost-everywhere measurable; here the map is linear on a finite-dimensional space, hence measurable, so this convention is not triggered. Edge cases: if or the array space is a single point, both sides are the unit point mass, and or is an empty matrix for which the orthogonality hypothesis holds automatically.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.