High-contrast homogenization: #
Provider.PolynomialHomogenization.Root.Corrector.PushforwardNormalizedSupply
Imported from the Apache-2.0 HighContrastHomogenization development at commit
7a13dbcd8d6609264a713373f5c69ceeac870472.
The normalized-gauge supply, from the certificate's own block row #
The stationary corrector family clause reduces to RootNormalizedSupply. This module
supplies it from the same physical block row that hhomogenized consumes,
through the available converter
exists_shiftedNormalizedReferencePowerTail_of_hasAllLaterPhysicalBlockRow.
Three things make the reduction work, and each is worth naming.
- The gauge matches. The converter's reference family satisfies
(aRef.coeffOn Q).toCoeffField = affineCoefficient (Selection.normalizedRoot (symmPart abar)) _ ↑↑(normalizedCenteredCoeff a abar hS), which is definitionallygaugeCoeff a abar hS— the two sides are the sameaffineCoefficientapplication, so the identification isrfl— hence almost everywhere⇑(normalizedSample abar hS a).1. This is exactly the identity the identity-gauge demand cannot have. - No smallness on the certificate's
deltais needed. The converter returns a power tail, andScalarIdentityPowerTail.rebasetrades reference scale for amplitude: pushing the base scale out drives the amplitude below any prescribed tolerance. That is why the identity-gauge row was stated∀ eps > 0and why nothing here demandsdeltabelow a private constant. - The translates are covered by the certificate itself.
hquenchedproduces the row on a setOmegaEndwith∀ z, translateCoeff z ⁻¹' OmegaEnd = OmegaEnd, so a certified sample and all of its integer translates carry the row. No per-zcertificate is assumed.
The one exponent condition is rho < 2 * s, where s is the good-tail order
the supply itself chooses. s is not a
frozen constant: it is carried by the block-row
supply and consumed only here, because the corrector datum records its order as
a field rather than an index. At the certificate's rho = (1 + 3 g) / 4 a
per-g order s := (1 + g) / 4 satisfies both 0 < s < 1/2 and rho < 2 * s
for every g ∈ [0, 1), so the admitted range is the whole of [0,1); the
frozen order would have forced g < 1/6.
Trading reference scale for amplitude #
A power tail is a good tail at every prescribed tolerance. Rebasing to
a larger reference scale divides the amplitude by a positive power, and the
geometric sum of the rebased tail is below any eps > 0 once the base is far
enough out.
The certificate-side block row #
The certificate-side input, in the shape hquenched produces it. A
certified sample lies in a translation-invariant set on which the physical block
row at the comparison matrix abar holds, with the row exponent below twice the
private order.
Every field is on RootAssembly's own binder surface: Omega is OmegaEnd
together with its ∀ z, translateCoeff z ⁻¹' OmegaEnd = OmegaEnd; rho is
(1 + 3 g) / 4; kappa, delta, S, X are the interface's; 1 ≤ X b and
S b ≤ X b are hquenched's own (∀ a, 1 ≤ X a) and (∀ a, S a ≤ X a); and
the row itself is the hypothesis hhomogenized consumes, at the same
Book.Ch02.constantBlockMatrix abar. Nothing new is introduced.
The good-tail order s is supplied here
together with the row, subject only to the corrector cone 0 < s < 1/2 and the
window rho < 2 * s. It is not a global constant, so the provider may choose
it after the contrast exponent g — and nothing downstream of the supply ever
sees it.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The supply #
The normalized-gauge datum, from the block row. The single input
correctorFamilyHole_of_normalizedSupply
needs is produced by the certificate's own row, in the certificate's own
gauge.
The hole, from the certificate's row alone #
the stationary corrector family clause of the root assembly, from the certificate's own
block row. This composes the inhabitant with the supply: no
gauge change, no per-z certificate, no smallness beyond the exponent window
rho < 2 * s at the supply's own order s.