AcceptedCorollary · Statement with omitted proof
Monotonicity of rangefinder error in semidefinite order
Ordering the input Gram matrices orders the squared residual matrices for a fixed test matrix.
Martinsson & Tropp ↗
Source label: cor:rrf-monotone
Formal statement. The assumptions and conclusion are explicit; the theorem proof is omitted (sorry).
01 / MATHEMATICAL STATEMENT
Meaning & conditions
Mathematical summary
Ordering the input Gram matrices orders the squared residual matrices for a fixed test matrix.
Assumptions & context
- B*B ≼ C*C and both matrices are tested with the same X.
Source
Martinsson & Tropp. Randomized numerical linear algebra: Foundations and algorithms.
Source label: cor:rrf-monotone
02 / LEAN REPRESENTATION
Formal statement
Statement with omitted proof. Expand the complete file below to read imports, local definitions and surrounding context.
theorem rrf_monotone {m p n ℓ : ℕ}
(B : Matrix (Fin m) (Fin n) 𝕜) (C : Matrix (Fin p) (Fin n) 𝕜)
(hBC : PsdLE (Bᴴ * B) (Cᴴ * C))
(X : Matrix (Fin n) (Fin ℓ) 𝕜) :
sqAbs (rangefinderError B X) = schurComplement (Bᴴ * B) X ∧
PsdLE (schurComplement (Bᴴ * B) X) (schurComplement (Cᴴ * C) X) ∧
schurComplement (Cᴴ * C) X = sqAbs (rangefinderError C X) := by
sorry
Complete Lean file and local definitions
import RNLAPhaseB.B3Base
import RNLAPhaseB.Complexity
open MeasureTheory
open scoped Matrix
open scoped Matrix.Norms.L2Operator
open scoped BigOperators
open scoped ENNReal
open scoped ComplexOrder
namespace RNLAPhaseB.MT2020CorRrfMonotone
open Matrix
variable {𝕜 : Type*} [RCLike 𝕜]
/-- Moore–Penrose conditions: `G` is the pseudoinverse `M^†` of `M`. -/
def IsMoorePenroseInverse {r s : ℕ} (M : Matrix (Fin r) (Fin s) 𝕜)
(G : Matrix (Fin s) (Fin r) 𝕜) : Prop :=
M * G * M = M ∧ G * M * G = G ∧ (M * G)ᴴ = M * G ∧ (G * M)ᴴ = G * M
/-- The Moore–Penrose pseudoinverse `M^†`, selected by its defining conditions
(it exists and is unique, so the selection is the pseudoinverse). -/
noncomputable def pinv {r s : ℕ} (M : Matrix (Fin r) (Fin s) 𝕜) :
Matrix (Fin s) (Fin r) 𝕜 :=
Classical.epsilon (IsMoorePenroseInverse M)
/-- The Nyström approximation `A⟨X⟩ := (AX)(X^* A X)^† (AX)^*` (eqn:nys-def). -/
noncomputable def nystrom {n ℓ : ℕ} (A : Matrix (Fin n) (Fin n) 𝕜)
(X : Matrix (Fin n) (Fin ℓ) 𝕜) : Matrix (Fin n) (Fin n) 𝕜 :=
(A * X) * pinv (Xᴴ * A * X) * (A * X)ᴴ
/-- The Schur complement `A / X := A - A⟨X⟩` (eqn:schur-complement). -/
noncomputable def schurComplement {n ℓ : ℕ} (A : Matrix (Fin n) (Fin n) 𝕜)
(X : Matrix (Fin n) (Fin ℓ) 𝕜) : Matrix (Fin n) (Fin n) 𝕜 :=
A - nystrom A X
/-- The rangefinder approximation error `E(B, X) := (I - P_Y) B` with `Y = BX`. -/
noncomputable def rangefinderError {m n ℓ : ℕ} (B : Matrix (Fin m) (Fin n) 𝕜)
(X : Matrix (Fin n) (Fin ℓ) 𝕜) : Matrix (Fin m) (Fin n) 𝕜 :=
(1 - orthogonalProjector (B * X)) * B
/-- The squared error `|E|^2 := E^* E` (a psd matrix, not a scalar). -/
noncomputable def sqAbs {m n : ℕ} (E : Matrix (Fin m) (Fin n) 𝕜) :
Matrix (Fin n) (Fin n) 𝕜 :=
Eᴴ * E
/-- The semidefinite (Loewner) order `M ≼ N`: `N - M` is positive semidefinite. -/
def PsdLE {n : ℕ} (M N : Matrix (Fin n) (Fin n) 𝕜) : Prop :=
(N - M).PosSemidef
/-- mt2020 cor:rrf-monotone: if `B^* B ≼ C^* C`, then for each fixed test matrix `X`,
`|E(B,X)|^2 = (B^*B)/X ≼ (C^*C)/X = |E(C,X)|^2`. -/
theorem rrf_monotone {m p n ℓ : ℕ}
(B : Matrix (Fin m) (Fin n) 𝕜) (C : Matrix (Fin p) (Fin n) 𝕜)
(hBC : PsdLE (Bᴴ * B) (Cᴴ * C))
(X : Matrix (Fin n) (Fin ℓ) 𝕜) :
sqAbs (rangefinderError B X) = schurComplement (Bᴴ * B) X ∧
PsdLE (schurComplement (Bᴴ * B) X) (schurComplement (Cᴴ * C) X) ∧
schurComplement (Cᴴ * C) X = sqAbs (rangefinderError C X) := by
sorry
end RNLAPhaseB.MT2020CorRrfMonotone
How to read the formalization
- The squared error is E*E, a positive semidefinite matrix. It is not a scalar squared norm.