Documentation

LeanPool.HighContrastHomogenization.Setup.SchurPositivity

High-contrast homogenization: Setup.SchurPositivity #

Imported from the Apache-2.0 HighContrastHomogenization development at commit 7a13dbcd8d6609264a713373f5c69ceeac870472.

Positivity of the Schur block and of the conjugated skew-corrected form #

The reference text reads |σ_*^{-1/2}(σ + (k - h)ᵗ σ_*⁻¹ (k - h))σ_*^{-1/2}|, the intrinsic contrast of the reference block, as the spectral norm of a symmetric positive semidefinite matrix. This file supplies that reading: on a symmetric positive definite doubled block matrix H, the conjugated skew-corrected form σ_*^{-1/2}(σ + (k - h)ᵗ σ_*⁻¹ (k - h))σ_*^{-1/2} is symmetric and positive semidefinite, for every skew parameter h — indeed for every h at all.

The substantive input is the positivity of the Schur block σ. It comes from the Schur complement identity: the doubled quadratic form of H evaluated at the state (p, k p) built from the Schur skew block is exactly the quadratic form of σ at p, so positive definiteness of H transfers to σ.

On real matrices, conjugate transpose agrees with transpose.

Symmetry of the blocks #

The upper-left block of a symmetric doubled block matrix is symmetric.

The off-diagonal blocks of a symmetric doubled block matrix are transposes of one another.

The Schur complement identity #

The Schur complement identity. Evaluated at the state (p, k p) built from the Schur skew block, the doubled quadratic form of a symmetric doubled block matrix with invertible lower-right block is the quadratic form of its Schur block σ.

Positivity #

The Schur block of a positive definite doubled block matrix is positive semidefinite.

The skew-corrected Schur form of a symmetric positive definite doubled block matrix is positive semidefinite, at every value of the skew parameter.

The conjugated skew-corrected form is symmetric positive semidefinite. This is what makes specBound of it the spectral norm |·| written in the intrinsic contrast of the reference block and in e.Theta.m.