Ordered elimination residuals #
The residual of symmetric Gaussian elimination of a principal index list is a
ratio of minors (EL01). Eliminating idxZ0 leaves the z-block and its
y-coupling unchanged (EL02). That z-block is a Gram matrix plus a
positive diagonal, hence PD, so later pivots stay positive (EL03). The
numerator minor expands in the I-diagonal as a sum of complementary Gram
minors (EL04); those principal cofactors have sign +1 (EL05). Factoring
s i > 0 and ρ > 0 leaves the unscaled kernel (1 + t_a t_b)² or a last
row (t_b - x)₊² (EL06). An algebraic identity rewrites (1 + s t)² as a
truncated square (EL07). Row arguments -1/t_a are increasing and lie
below any x ≥ 0 (EL08–EL09). Those leftover minors are nonnegative by
TN13 (EL10), so residual columns stay nonnegative (EL11). Scaled outer
products of those columns yield a completely positive C₀ (EL12). The
remainder is the T-Schur complement of E and is PSD (EL13–EL14).
EL01: the residual after eliminating the principal block A[I] is a ratio of
minors. Injectivity of e and α, j ∉ range e record the elimination setup.
Singleton Schur residual, before simplifying the 1 × 1 inverse.
Feature map i ↦ zᵢ ⊗ zᵢ on Fin 2 × Fin 2.
Equations
- BollobasNikiforov.zKron s t i p = BollobasNikiforov.zVec s t i p.1 * BollobasNikiforov.zVec s t i p.2
Instances For
A diagonal Schur residual of a PD matrix is a ratio of positive principal minors, hence positive.
EL03: every Schur pivot inside the z-block is positive.
EL03: leading principal submatrices of the z-block are PD.
EL04 — diagonal expansion of a numerator minor #
Equal index Finsets give equal principal minors.
EL04: det(A + diagonal d) expands over kept index sets S.
The product runs over the complementary (deleted) diagonal entries, and the
leftover minor is the principal submatrix of A on S.
If d is supported on I, the expansion runs over deleted sets S ⊆ I.
The z-z numerator minor is the Gram block plus the I-diagonal
(and the unexpanded last diagonal if α = j).
Row map for a last y-index after the I-block.
Equations
- BollobasNikiforov.elimRowY e ℓ a = if h : a = Fin.last m then BollobasNikiforov.idxY ℓ else BollobasNikiforov.idxZ (e (a.castPred h))
Instances For
The y-row numerator minor is the Gram/coupling block plus the I-diagonal.
EL04: expand a y-row numerator in the diagonal summands on I.
EL05 — principal cofactor signs are +1 #
EL05: deleting the same row and column subset contributes sign +1.
The leftover is the complementary principal minor, with no extra sign.
EL06 — factor positive s and ρ #
EL06: the same factoring on a complementary leftover of the z-z Gram.
EL10 — leftover kernel minors are nonnegative #
EL10: a z-z kernel minor with strictly increasing index maps is
nonnegative.
EL10: complementary leftover of a z-z kernel minor is nonnegative.
Row argument of the elimination step: the vector on Fin m.succ whose last entry is xval and
whose other entries are -1 / t along the indices selected by elimSnoc e jidx.
Equations
- BollobasNikiforov.elimRowArg t xval e jidx a = if a = Fin.last m then xval else -1 / t (BollobasNikiforov.elimSnoc e jidx a)
Instances For
EL10: a last truncated-square row still gives a nonnegative minor.
EL10: complementary leftover of a y-row kernel minor is nonnegative.
EL11 — residual columns are nonnegative #
EL11: a y-row numerator minor is nonnegative.
EL11: residual of a later z-index against a leading z-pivot is ≥ 0.
EL11: residual of a y-index against a leading z-pivot is ≥ 0.
Instances For
The E-block of M (Xconfig s t ρ x): rows and columns indexed by the axis vector and the left
vectors through elimEEmbed.
Equations
Instances For
The block of M (Xconfig s t ρ x) with E-rows (via elimEEmbed) and right-vector columns
(via elimT).
Equations
Instances For
The block of M (Xconfig s t ρ x) with right-vector rows (via elimT) and E-columns (via
elimEEmbed).
Equations
Instances For
Schur complement of the E-block in the T-block.
Equations
- BollobasNikiforov.elimSchurR s t ρ x = BollobasNikiforov.elimTT s t ρ x - BollobasNikiforov.elimTE s t ρ x * (BollobasNikiforov.elimEE s t ρ x)⁻¹ * BollobasNikiforov.elimET s t ρ x
Instances For
Extension by zero off the T-indices.
Equations
- BollobasNikiforov.extendByZeroT R α β = match (BollobasNikiforov.configIdxEquiv k p) α, (BollobasNikiforov.configIdxEquiv k p) β with | Sum.inr j, Sum.inr ℓ => R j ℓ | x, x_1 => 0
Instances For
EL11: every residual column on ConfigIdx is nonnegative.
EL12 — C₀ is completely positive #
The j-th elimination residual elimRes divided by the square root of its pivot elimPivot:
the j.succ-th factor of C₀.
Equations
- BollobasNikiforov.elimVec s t ρ x j α = BollobasNikiforov.elimRes s t ρ x j α / √(BollobasNikiforov.elimPivot s t ρ x j)
Instances For
The k + 1 factors of C₀: elimVec0 followed by the vectors elimVec s t ρ x j.
Equations
- BollobasNikiforov.elimC0Factor s t ρ x = Fin.cons (BollobasNikiforov.elimVec0 s t ρ x) (BollobasNikiforov.elimVec s t ρ x)
Instances For
Sum of scaled residual outer products.
Equations
- BollobasNikiforov.elimC0 s t ρ x = ∑ a : Fin (k + 1), Matrix.vecMulVec (BollobasNikiforov.elimC0Factor s t ρ x a) (BollobasNikiforov.elimC0Factor s t ρ x a)
Instances For
EL12: C₀ is completely positive.
Row α of M (Xconfig s t ρ x) restricted to the E-columns.
Equations
- BollobasNikiforov.elimColE s t ρ x α a = BollobasNikiforov.M (BollobasNikiforov.Xconfig s t ρ x) α (BollobasNikiforov.elimEEmbed a)
Instances For
P EE⁻¹ Pᵀ on configuration indices.
Equations
- BollobasNikiforov.elimGramE s t ρ x α β = BollobasNikiforov.elimColE s t ρ x α ⬝ᵥ (BollobasNikiforov.elimEE s t ρ x)⁻¹.mulVec (BollobasNikiforov.elimColE s t ρ x β)
Instances For
The leading m × m principal block of M (Xconfig s t ρ x) on the left-vector indices idxZ (Fin.castLE hm i).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Row α of M (Xconfig s t ρ x) restricted to the first m left-vector columns.
Equations
- BollobasNikiforov.elimColZLead s t ρ x hm α i = BollobasNikiforov.M (BollobasNikiforov.Xconfig s t ρ x) α (BollobasNikiforov.idxZ (Fin.castLE hm i))
Instances For
The matrix c_α ⬝ᵥ (elimZLead)⁻¹ *ᵥ c_β built from the columns elimColZLead: the part of M
explained by the first m left vectors.
Equations
- BollobasNikiforov.elimGramZLead s t ρ x hm α β = BollobasNikiforov.elimColZLead s t ρ x hm α ⬝ᵥ (BollobasNikiforov.elimZLead s t ρ x hm)⁻¹.mulVec (BollobasNikiforov.elimColZLead s t ρ x hm β)
Instances For
The block-diagonal matrix on Option (Fin k) with (M idxZ0 idxZ0)⁻¹ in the none corner and
the inverse of the left-vector block of M (Xconfig s t ρ x) on the some indices.
Equations
- BollobasNikiforov.elimEEInv s t ρ x none none = (BollobasNikiforov.M (BollobasNikiforov.Xconfig s t ρ x) BollobasNikiforov.idxZ0 BollobasNikiforov.idxZ0)⁻¹
- BollobasNikiforov.elimEEInv s t ρ x none (some val) = 0
- BollobasNikiforov.elimEEInv s t ρ x (some val) none = 0
- BollobasNikiforov.elimEEInv s t ρ x (some i) (some h) = ((BollobasNikiforov.M (BollobasNikiforov.Xconfig s t ρ x)).submatrix BollobasNikiforov.idxZ BollobasNikiforov.idxZ)⁻¹ i h
Instances For
EL14 — lem:elimination.