Feature factorisation M = FFᵀ + L and the Schur complement of E #
SC01–SC07 record the weights q i, γ, the feature matrix F, the
weighted Laplacian L, the E-block L_EE and its inverse, and the
reduced data L_red, 𝒰 of docs/sol.tex §4.
After block elimination replaces L by diag(L_EE, L_red) and the feature
rows by (F_E, 𝒰), the Schur complement of the first block is
L_red + 𝒰 (I - F_Eᵀ (L_EE + F_E F_Eᵀ)⁻¹ F_E) 𝒰ᵀ. The bracket equals
(I + F_Eᵀ L_EE⁻¹ F_E)⁻¹ by the Woodbury companion identity
(paper (eq:Schur)).
Woodbury formula for a rank-κ update L + F Fᵀ.
SC08. Woodbury companion identity:
I - Fᵀ (L + F Fᵀ)⁻¹ F = (I + Fᵀ L⁻¹ F)⁻¹.
Expanding 𝒰 (I - F_Eᵀ (L_EE + F_E F_Eᵀ)⁻¹ F_E) 𝒰ᵀ as a Schur remainder.
After replacing L by diag(L_EE, L_red), the Schur complement of E in
FFᵀ + L is L_red + 𝒰 (I - F_Eᵀ (L_EE + F_E F_Eᵀ)⁻¹ F_E) 𝒰ᵀ.
SC08, reduced coordinates. The Schur complement of E in FFᵀ + L
after block diagonalization of L equals
L_red + 𝒰 (I + F_Eᵀ L_EE⁻¹ F_E)⁻¹ 𝒰ᵀ.
Configuration-dependent Schur data (SC01–SC07) #
SC01 — weights q i and γ #
qᵢ = sᵢ / dᵢ.
Equations
- BollobasNikiforov.configQ s t ρ x i = s i / BollobasNikiforov.configD t ρ x i
Instances For
γ = σ - ∑ qᵢ.
Equations
- BollobasNikiforov.configγ s t ρ x = BollobasNikiforov.configσ s - ∑ i : Fin k, BollobasNikiforov.configQ s t ρ x i
Instances For
SC02 — feature matrix F #
Feature inner products
SC03 — Laplacian L and M = FFᵀ + L #
Embeds the E-indices into ConfigIdx k p: the axis vector none ↦ idxZ0 and the left vectors
some i ↦ idxZ i.
Equations
Instances For
SC05 — Schur complement of the lower-right of L_EE is γ #
SC06 — inverse of L_EE #
The weight vector on the E-indices: 1 at the axis vector and (configD t ρ x i)⁻¹ at the
i-th left vector.
Equations
- BollobasNikiforov.configW t ρ x none = 1
- BollobasNikiforov.configW t ρ x (some i) = (BollobasNikiforov.configD t ρ x i)⁻¹
Instances For
The diagonal matrix on the E-indices with 0 at the axis vector and (s i * configD t ρ x i)⁻¹ at the i-th left vector; it is the diagonal part of (configLEE s t ρ x)⁻¹.
Equations
- BollobasNikiforov.configLEEInvDiag s t ρ x = Matrix.diagonal fun (x_1 : Option (Fin k)) => match x_1 with | none => 0 | some i => (s i * BollobasNikiforov.configD t ρ x i)⁻¹
Instances For
SC07 — L_red and 𝒰 #
The right-vector rows of the three-column factor configF s t ρ x.
Equations
- BollobasNikiforov.configFT s t ρ x = (BollobasNikiforov.configF s t ρ x).submatrix BollobasNikiforov.idxY id
Instances For
The Schur complement L_TT - L_TE L_EE⁻¹ L_ET of the E-block in the weighted Laplacian. The
positivity hypotheses are the conditions under which L_EE is invertible.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The reduced three-column factor F_T - L_TE L_EE⁻¹ F_E. The positivity hypotheses are the
conditions under which L_EE is invertible.
Equations
- One or more equations did not get rendered due to their size.
Instances For
SC10 — center row of L_EE⁻¹ F_E #
The product L_EE⁻¹ F_E of the inverse E-block with the E-rows of the three-column
factor.
Equations
- BollobasNikiforov.configInvFE s t ρ x = (BollobasNikiforov.configLEE s t ρ x)⁻¹ * BollobasNikiforov.configFE s t ρ x
Instances For
SC12 — rows of 𝒰 #
SC14 — off-diagonal of L_red #
SC15 — L 1 = 0 implies L_red 1 = 0 #
SC16 — L_red is a weighted Laplacian #
SC09 — elimSchurR equals the Woodbury remainder #
SC13 — C₁ is completely positive #
The matrix C₁ on the right-vector indices: ρ j * (U (x j) ⬝ᵥ 𝒦 *ᵥ U (x ℓ)) * ρ ℓ, a kernel
Gram matrix with parameters configQ and configγ.
Equations
- One or more equations did not get rendered due to their size.
Instances For
SC13. C₁ is completely positive.
Pad a CP matrix by zeros along the complement of an injection.
SC17 — assembly for γ > 0 #
SC17. M = elimC0 + extend C₁ + extend L_red.
SC18 — cor:laplacian on the assembled remainder #
The weights -(configLred j ℓ) on pairs j < ℓ of right-vector indices, and 0 on every other
pair of ConfigIdx k p.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Move the mass of coordinate j of p onto coordinate i: p i + p j at i, 0 at j, and
p elsewhere.
Instances For
Move the mass of coordinate i of p onto coordinate j: 0 at i, p i + p j at j, and
p elsewhere.
Equations
Instances For
The pairs (i, j) with i < j.
Equations
- BollobasNikiforov.offDiagLt = {p ∈ Finset.univ.offDiag | p.1 < p.2}
Instances For
The weighted Laplacian ∑_{i < j} ℓ i j • (eᵢ - eⱼ)(eᵢ - eⱼ)ᵀ.
Equations
- One or more equations did not get rendered due to their size.
Instances For
SC18. If γ > 0 then M(Xconfig) is completely positive.