Documentation

LeanPool.HighContrastHomogenization.Provider.PolynomialHomogenization.Root.CorrectorComposition.BlockRowSupplyProjection

High-contrast homogenization: #

Provider.PolynomialHomogenization.Root.CorrectorComposition.BlockRowSupplyProjection

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

RootBlockRowSupply against the ceiling-exporting smallness interface — what is carried and #

what is not

The stationary corrector family clause reduces to

def RootBlockRowSupply (d : ℕ) [NeZero d]
    (GoodScale : Mat d → CoeffSpace d → ℝ → Prop) : Prop :=
  ∀ (abar : Mat d), (symmPart abar).PosDef →
    ∀ (a : CoeffSpace d) (x : ℝ), GoodScale abar a x →
      ∃ (Omega : Set (CoeffSpace d)) (s rho kappa delta : ℝ)
        (S X : CoeffSpace d → ℝ),
        a ∈ Omega ∧
        (∀ z : Fin d → ℤ, translateCoeff z ⁻¹' Omega = Omega) ∧
        0 < s ∧ s < 1 / 2 ∧
        0 < rho ∧ rho ≤ 1 ∧ rho < 2 * s ∧
        0 < kappa ∧ 0 ≤ delta ∧
        ∀ b ∈ Omega, 1 ≤ X b ∧ S b ≤ X b ∧
          Quenched.HasAllLaterPhysicalBlockRow rho kappa delta
            (Book.Ch02.constantBlockMatrix abar) S X b

(order-free form: the order s is supplied here, not frozen upstream; the statement is restated below rather than imported, because importing it would pull in the whole normalized-supply chain at the free order.)

What rootGoodScaleAt literally carries #

Destructuring rootGoodScaleAt d g c kappaRate abar a x gives h : RowSupply.RowRetainingPrintOrderGoodScale d g c kappaRate abar a x, and that structure carries, for the sample a only:

demanded fieldcarried bystatus
rhoCertificate.printRowOrder g = (1 + 3 * g) / 4✓
0 < rho ∧ rho ≤ 1g ∈ Ico 0 1 arithmetic✓
kappa and 0 < kappa2 * kappaRate, positive from the hole's own 0 < κ✓
delta, 0 ≤ deltah.delta, h.delta_nonneg✓
S, Xh.sourceScale, h.activationScale✓
1 ≤ X a, S a ≤ X ah.activation_one, h.source_le_activation✓ at a
the row at ah.row, at Book.Ch02.constantBlockMatrix abar✓ at a
s with 0 < s < 1/2 and rho < 2 * sthe printed order (1 + g) / 4, chosen after g✓
Omega with ∀ z, translateCoeff z ⁻¹' Omega = Omega, a ∈ Omega, and the row **at every `b ∈
Omega`**—x the missing field

The single missing field is the translation-invariant event. The certificate is a per-sample predicate: it carries the row at a and at no other sample. The demanded Omega exists — it is hquenched's OmegaEnd, which carries MeasurableSet, P.real = 1, the literal ∀ z, translateCoeff z ⁻¹' OmegaEnd = OmegaEnd, and ∀ a ∈ OmegaEnd, HasAllLaterPhysicalBlockRow ((1 + 3 * g) / 4) kappa delta Abar S X a — but hhomogenized's binder surface applies one sample at a time (… (a : CoeffSpace d), HasAllLaterPhysicalBlockRow … S X a → 1 ≤ X a → S a ≤ X a → GoodScale g kappaRate abar a …), so the event is discarded before GoodScale is formed. No choice of Omega repairs this downstream: Omega = {a} is not translation invariant, Omega = univ does not carry the row, and the orbit of a carries the row only if the row is already known at every integer translate — which is the demand itself.

The second missing field was an order window, not a datum — and it is now gone. At the stationary corrector family clause's frozen rootCorrectorOrder = 3/16 the window rho < 2 * rootCorrectorOrder read (1 + 3 * g) / 4 < 3 / 8, i.e. g < 1/6 (still proved below, as the record of what the frozen order cost). The corrector datum records its order as a field rather than reading a global constant, so the supply chooses the order after g, at the printed (1 + g) / 4; the corrector cone s < 1/2 and the window rho < 2 * s are then both equivalent to g < 1, and the admitted range is the whole of Ico 0 1.

The repair, machine-checked: the event-retaining certificate #

Widen the certificate by the event and widen hhomogenized correspondingly. Both directions are proved here:

So the event-retaining certificate is a pure binder-surface change to hhomogenized: replace its per-sample row premise by the event-level one that the assembly already has in hand at the point of application.

The stationary corrector family clause's demand, restated #

The order window #

theorem HCPolySupport.HighContrast.CorrectorComposition.printRowOrder_mem {g : ℝ} (hg : g ∈ Set.Ico 0 1) :
0 < (1 + 3 * g) / 4 ∧ (1 + 3 * g) / 4 ≤ 1

On g ∈ Ico 0 1 the row order is a genuine exponent.

The per-g order, and the exact admitted range #

The order the supply chooses is the printed certificate order HighContrast.printCertificateOrder g = (1 + g) / 4, selected after g. At that order the two constraints the supply must meet — the corrector cone 0 < s < 1/2 and the window printRowOrder g < 2 * s — are both equivalent to g < 1, so the admitted range is exactly Ico 0 1: the whole root range, and nothing is admitted beyond it.

The corrector cone at the printed order is exactly g < 1.

The supply window at the printed order is exactly g < 1.

The printed order is positive on the whole root range (indeed for -1 < g).

The event-retaining certificate the event-retaining certificate #

The block-row event. Exactly the data hquenched produces, restricted to what RootBlockRowSupply reads: a translation-invariant set carrying the sample, and the physical block row on all of it.

Equations
  • One or more equations did not get rendered due to their size.
Instances For
    def HCPolySupport.HighContrast.CorrectorComposition.rootGoodScaleOn (d : ℕ) [NeZero d] (g c kappaRate : ℝ) (abar : Mat d) (a : CoeffSpace d) (x : ℝ) :

    The event-retaining certificate: the ceiling-exporting smallness interface with the producing event retained.

    Equations
    • One or more equations did not get rendered due to their size.
    Instances For

      The discharge #

      hhomogenized at the event-retaining certificate #

      theorem HCPolySupport.HighContrast.CorrectorComposition.exists_rootGoodScaleOn_of_quenchedEventRow (d : ℕ) [NeZero d] {g : ℝ} (hg : g ∈ Set.Ico 0 1) {c kappa delta : ℝ} (hc : c ∈ Set.Ioo 0 1) (hkappa : 0 < kappa) (hdelta : delta ∈ Set.Ioo 0 1) :
      ∃ (kappaRate : ℝ) (Lcert : ℝ) (pCert : ℝ), 0 < kappaRate ∧ kappaRate ≤ (1 + g) / 4 ∧ 1 ≤ Lcert ∧ 0 ≤ pCert ∧ ∀ (abar : Mat d), Matrix.PosDef (symmPart abar) → ∀ (Omega : Set (CoeffSpace d)) (S X : CoeffSpace d → ℝ), (∀ (z : Fin d → ℤ), translateCoeff z ⁻¹' Omega = Omega) → (∀ b ∈ Omega, 1 ≤ X b) → (∀ b ∈ Omega, S b ≤ X b) → (∀ b ∈ Omega, Quenched.HasAllLaterPhysicalBlockRow ((1 + 3 * g) / 4) kappa delta (Book.Ch02.constantBlockMatrix abar) S X b) → ∀ a ∈ Omega, rootGoodScaleOn d g c kappaRate abar a (X a * (Lcert * RowSupply.eccentricityFoldFactor abar pCert))

      hhomogenized survives the widening. From the event-level row — which is precisely what hquenched hands the assembly — the event-retaining certificate holds at every sample of the event, at every corrector smallness. No new hypothesis is introduced: the premises below are hquenched's own OmegaEnd data.