AcceptedCorollary · Statement with omitted proof
Gaussian power-rangefinder moment bound
A power rangefinder controls expected spectral error through a higher error moment.
Martinsson & Tropp ↗
Source label: cor:rrf-power-gauss
Formal statement. The assumptions and conclusion are explicit; the theorem proof is omitted (sorry).
01 / MATHEMATICAL STATEMENT
Meaning & conditions
Mathematical summary
A power rangefinder controls expected spectral error through a higher error moment.
Assumptions & context
- B is real, q ≥ 1, k < ℓ−1, Y = (BBᵀ)^qΩ, and Ω is a standard Gaussian m × ℓ matrix. Z is the spectral norm of a k × ℓ Gaussian pseudoinverse.
Source
Martinsson & Tropp. Randomized numerical linear algebra: Foundations and algorithms.
Source label: cor:rrf-power-gauss
02 / LEAN REPRESENTATION
Formal statement
Statement with omitted proof. Expand the complete file below to read imports, local definitions and surrounding context.
theorem cor_rrf_power_gauss
{Ω Θ : Type*} [MeasurableSpace Ω] [MeasurableSpace Θ]
(P : Measure Ω) (Q : Measure Θ)
(m n l k q : ℕ) (B : Matrix (Fin m) (Fin n) ℝ)
(Om : Ω → Matrix (Fin m) (Fin l) ℝ) (hOm : StandardGaussianTestMatrix P Om)
(hk : k < l - 1) (hq : 1 ≤ q)
(Γ : Θ → Matrix (Fin k) (Fin l) ℝ) (hΓ : StandardGaussianTestMatrix Q Γ)
(Γpinv : Θ → Matrix (Fin l) (Fin k) ℝ)
(hpinv : ∀ θ, IsMoorePenroseInverse (Γ θ) (Γpinv θ)) :
let Y : Ω → Matrix (Fin m) (Fin l) ℝ := fun ω => (B * Bᵀ) ^ q * Om ω
let err : Ω → ℝ := fun ω => ‖(1 - orthogonalProjector (Y ω)) * B‖
let EZ : ℝ := ∫ θ, ‖Γpinv θ‖ ∂Q
let σ : ℕ → ℝ := singularValue B
(∫ ω, err ω ∂P) ≤ (∫ ω, err ω ^ (2 * q) ∂P) ^ (1 / (2 * (q : ℝ))) ∧
(∫ ω, err ω ^ (2 * q) ∂P) ^ (1 / (2 * (q : ℝ))) ≤
((1 + Real.sqrt ((k : ℝ) / ((l : ℝ) - k - 1))) * σ (k + 1) ^ (2 * q)
+ EZ * Real.sqrt (∑ j ∈ Finset.Ioc k n, σ j ^ (4 * q))) ^ (1 / (2 * (q : ℝ))) := 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
namespace RNLAPhaseB.MT2020CorRRFPowerGauss
open RNLAPhaseB
/-- Moore–Penrose conditions: `X = A^†` (unique solution). -/
def IsMoorePenroseInverse {k l : ℕ} (A : Matrix (Fin k) (Fin l) ℝ)
(X : Matrix (Fin l) (Fin k) ℝ) : Prop :=
A * X * A = A ∧ X * A * X = X ∧ (A * X)ᵀ = A * X ∧ (X * A)ᵀ = X * A
/-- Singular values of `B` in weakly decreasing order: square roots of the eigenvalues
of `Bᵀ B`, which has `n` of them. -/
noncomputable def singularValues {m n : ℕ} (B : Matrix (Fin m) (Fin n) ℝ) : List ℝ :=
(List.ofFn fun i : Fin n =>
Real.sqrt ((Matrix.isHermitian_conjTranspose_mul_self B).eigenvalues i)).mergeSort
(fun a b => decide (b ≤ a))
/-- 1-based `σ_j(B)`, zero-padded; used only for `j ≥ 1`. -/
noncomputable def singularValue {m n : ℕ} (B : Matrix (Fin m) (Fin n) ℝ) (j : ℕ) : ℝ :=
(singularValues B).getD (j - 1) 0
/-- MT2020 cor:rrf-power-gauss, under the conditions of thm:rrf-gauss.
* `B ∈ ℝ^{m×n}` and `k < ℓ - 1`.
* `Z = ‖Γ^†‖`, where `Γ ∈ ℝ^{k×ℓ}` is standard normal.
* The test matrix `Ω` is standard normal and `m×ℓ`, as in alg:power-rangefinder.
* `Y = (BB^*)^q Ω`.
The norm is the spectral norm. The exponent `1/(2q)` is read with `q ≥ 1`. -/
theorem cor_rrf_power_gauss
{Ω Θ : Type*} [MeasurableSpace Ω] [MeasurableSpace Θ]
(P : Measure Ω) (Q : Measure Θ)
(m n l k q : ℕ) (B : Matrix (Fin m) (Fin n) ℝ)
(Om : Ω → Matrix (Fin m) (Fin l) ℝ) (hOm : StandardGaussianTestMatrix P Om)
(hk : k < l - 1) (hq : 1 ≤ q)
(Γ : Θ → Matrix (Fin k) (Fin l) ℝ) (hΓ : StandardGaussianTestMatrix Q Γ)
(Γpinv : Θ → Matrix (Fin l) (Fin k) ℝ)
(hpinv : ∀ θ, IsMoorePenroseInverse (Γ θ) (Γpinv θ)) :
let Y : Ω → Matrix (Fin m) (Fin l) ℝ := fun ω => (B * Bᵀ) ^ q * Om ω
let err : Ω → ℝ := fun ω => ‖(1 - orthogonalProjector (Y ω)) * B‖
let EZ : ℝ := ∫ θ, ‖Γpinv θ‖ ∂Q
let σ : ℕ → ℝ := singularValue B
(∫ ω, err ω ∂P) ≤ (∫ ω, err ω ^ (2 * q) ∂P) ^ (1 / (2 * (q : ℝ))) ∧
(∫ ω, err ω ^ (2 * q) ∂P) ^ (1 / (2 * (q : ℝ))) ≤
((1 + Real.sqrt ((k : ℝ) / ((l : ℝ) - k - 1))) * σ (k + 1) ^ (2 * q)
+ EZ * Real.sqrt (∑ j ∈ Finset.Ioc k n, σ j ^ (4 * q))) ^ (1 / (2 * (q : ℝ))) := by
sorry
end RNLAPhaseB.MT2020CorRRFPowerGauss
How to read the formalization
- The test matrix is in the row dimension m in this version. The power and moment exponent are 2q, unlike the 2q+1 form used in HMT.