Documentation

LeanPool.TrilinearHilbertTransform.Auto.Sec163SummingUpdateLevels.RetainedNumericExp

Trilinear Hilbert transform: retained_numeric_exp #

theorem TrilinearHT.Auto.retained_numeric_exp {L₄ lW α θ₀ Ca : ℝ} {b m c₃ c cc Cl : ℕ} (hL₄ : 0 ≤ L₄) (hlW : 0 ≤ lW) (hα : 0 < α) (hα8 : α ≤ 1 / 9600) (hθ₀ : 0 < θ₀) (hθ₁ : θ₀ ≤ 1) (hCa : 0 < Ca) (hb : 1 ≤ b) (hbc : 30 * b < c) (hmc₃ : m < c₃) (hbc₃ : b < c₃) (hm1 : 1 ≤ m) (hcc : 1 ≤ cc) :
∀ᶠ (k : ℕ) in Filter.atTop, (Real.exp (2 * α * ↑k ^ θ₀) + 1) * (20 * (6000 ^ 30 * (L₄ + 2 * ↑k ^ b) ^ 30 + 1) + 60) + 1 ≤ Real.exp (↑k ^ c) ∧ 8 * (lW + (L₄ + 2 * ↑k ^ b) + 1) ≤ Real.exp ((2 * ↑Cl + 2) * ↑k ^ (cc + 1)) ∧ 4 * (6000 ^ 30 * (L₄ + 2 * ↑k ^ b) ^ 30 + 1) * ↑Cl * ((2 * ↑Cl + 2) * ↑k ^ (cc + 1)) ^ Cl * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl ≤ Real.exp (2 * α * ↑k ^ θ₀) ∧ ↑Cl * ((2 * ↑Cl + 2) * ↑k ^ (cc + 1)) ^ Cl + 2 ≤ Real.exp (2 * α * ↑k ^ θ₀) ∧ 4 * (6000 ^ 30 * (L₄ + 2 * ↑k ^ b) ^ 30 + 1) * ↑Cl * (2 + (↑Cl + 1) * (2 * α * ↑k ^ θ₀ + 1)) ^ 6 * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl ≤ Real.exp (α * ↑k ^ θ₀ / 4) ∧ 4 * (6000 ^ 30 * (L₄ + 2 * ↑k ^ b) ^ 30 + 1) * (↑Cl * ((2 * ↑Cl + 2) * ↑k ^ (cc + 1)) ^ Cl * (Real.exp (2 * α * ↑k ^ θ₀) + 1) ^ Cl * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl * ((Real.exp (2 * α * ↑k ^ θ₀) + 2) ^ 2 * Real.exp (-↑k ^ c₃) + Real.exp (↑Cl * ↑k ^ cc) * (Real.exp (-((2 * ↑Cl + 2) * ↑k ^ (cc + 1)) / 2) + Real.exp (-((2 * ↑Cl + 2) * ↑k ^ (cc + 1))))) + ↑Cl * Real.exp (-↑k)) + (Real.exp (Ca * ↑k ^ m) + 1) * (4 * (6000 ^ 30 * (L₄ + 2 * ↑k ^ b) ^ 30 + 1) * (9 * (Real.exp (L₄ + 2 * ↑k ^ b + 3) + 1) * Real.exp (-↑k ^ c₃) * 16)) ≤ Real.exp (-2 * ↑k ^ θ₀) ^ min (1 / 1200 / 2) (α / 4)

The exponential standing bounds of the retained-attempt estimate at accuracy k (blueprint lines 2870-2888: P = ⌈η₀^{-α}⌉ dominates every fixed power of k and of the unit G = e^{(1 + log 2k)²}, and the closure, mode and catalog errors are below η₀^{c_*}), with Y = L₄ + 2 k^b, F = 4 (6000^{30} Y^{30} + 1), k_D = (2 C_l + 2) k^{cc+1}, τ₃ = e^{-k^{c₃}} and η₀ = e^{-2 k^{θ₀}}.

theorem TrilinearHT.Auto.retained_arith {Cl P k cc : ℕ} {kD η₀ τ₃ δ CN Cq q R Nlv Er : ℝ} (hη0 : 0 < η₀) (hη1 : η₀ ≤ 1) (hCN : 0 ≤ CN) (hCq : 0 ≤ Cq) (hR : 0 ≤ R) (hNlv : 0 ≤ Nlv) (hkD : 0 ≤ kD) (hP : ↑P = ↑⌈η₀ ^ (-(1 / 1200 / (4 * (↑Cl + 1 + 1))))⌉₊) (hq2 : 2 < q) (hq3 : q < 3) (hq : 2 * (↑Cl + 1) * ((q - 2) / (q - 1)) ≤ 1 / 8) (h1 : Nlv * ↑Cl * kD ^ Cl * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl ≤ ↑P) (h2 : ↑Cl * kD ^ Cl + 2 ≤ ↑P) (h3 : Nlv * ↑Cl * (2 + (↑Cl + 1) * Real.log ↑P) ^ 6 * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl ≤ ↑P ^ (1 / 8)) (h4 : Nlv * (↑Cl * (kD * ↑P) ^ Cl * Real.exp ((1 + Real.log (2 * ↑k)) ^ 2) ^ Cl * ((↑P + 1) ^ 2 * τ₃ + Real.exp (↑Cl * ↑k ^ cc) * (√δ + Real.exp (-kD))) + ↑Cl * Real.exp (-↑k)) + Er ≤ η₀ ^ min (1 / 1200 / 2) (1 / 1200 / (4 * (↑Cl + 1 + 1)) / 4)) :
Nlv * leafBnd Cl kD P k cc (4 * η₀) η₀ τ₃ δ CN Cq q R + Er * R ≤ ((3 + Cq) * (2 ^ (↑Cl + 1) + 1) + 1) * η₀ ^ min (1 / 1200 / 2) (1 / 1200 / (4 * (↑Cl + 1 + 1)) / 4) * (1 + CN) * R

The final absorption on one root (blueprint lines 2851-2888, from eq:before-absorption to eq:active-final) for the explicit leaf bound summed over N_lv leaves plus a localization error E_r |R|. With ε = 1/1200, C₄ = C_l + 1, α = ε/(4(C₄+1)), P = ⌈η₀^{-α}⌉ and η = 4η₀: if P dominates the leaf count, the powers of k_D and the unit G = e^{(1 + log 2k)²} (h1-h3), and the remaining errors are at most η₀^{c_*} (h4), the total is at most ((3 + C_q)(2^{C₄} + 1) + 1) η₀^{c_*} (1 + C_N) |R| with c_* = min(ε/2, α/4).

theorem TrilinearHT.Auto.ActiveSetup.root_assembly {κ : LocalKernelClass} {c₀ : ℕ} (A : ActiveSetup κ c₀) (p : Fin 3) {δa : ℝ} (Tc FR : Finset (ℤ × ℤ)) (hTc : Tc = Finset.filter (A.active δa) A.cells) (Nl : ℕ) (Fl : ℕ → Fin 6 → Finset (ℤ × ℤ)) (Lf : ℕ → ℤ × ℤ → ℝ) {ρ Bnd Mp W : ℝ} (hFRsub : ∀ c ∈ FR, dyadicGridCell A.G c.1 c.2 ⊆ dyadicGridCell A.G A.R.1 A.R.2) (hFR : ∑ c ∈ FR, dyadicGridLength A.G c.1 ≤ 1 / 8 * dyadicGridLength A.G A.R.1) (hFlsub : ∀ s < Nl, ∀ (i : Fin 6), ∀ c ∈ Fl s i, dyadicGridCell A.G c.1 c.2 ⊆ dyadicGridCell A.G A.R.1 A.R.2) (hFl : ∀ s < Nl, ∀ (i : Fin 6), dyadicRootLength A.G (Fl s i) ≤ 2 * ρ * dyadicGridLength A.G A.R.1) (hNlρ : ↑Nl * (6 * (2 * ρ)) ≤ 1 / 8) (hloc : ∀ cs ⊆ Tc, (∀ I ∈ cs, ¬∃ e ∈ FR, dyadicGridCell A.G I.1 I.2 ⊆ dyadicGridCell A.G e.1 e.2) → ∑ I ∈ cs, dyadicGridLength A.G I.1 * min A.η ‖A.Hf I (A.z p.castSucc I)‖ ≤ ∑ s ∈ Finset.range Nl, ∑ I ∈ cs, dyadicGridLength A.G I.1 * Lf s I + (∑ I ∈ cs, dyadicGridLength A.G I.1) * W) (hleaf : ∀ s < Nl, ∀ cs ⊆ Tc, (∀ I ∈ cs, ¬∃ e ∈ FR, dyadicGridCell A.G I.1 I.2 ⊆ dyadicGridCell A.G e.1 e.2) → (∀ I ∈ cs, ∀ (i : Fin 6), ∀ e ∈ Fl s i, ¬dyadicGridCell A.G I.1 I.2 ⊆ dyadicGridCell A.G e.1 e.2) → ∑ I ∈ cs, dyadicGridLength A.G I.1 * Lf s I ≤ Bnd) (hMp : ∀ (y : ℝ), ↑{I ∈ A.cells | y ∈ dyadicGridCell A.G I.1 I.2 ∧ A.active δa I}.card ≤ Mp) (hW : 0 ≤ W) :
∃ (X : Finset (ℤ × ℤ)), (∀ c ∈ X, dyadicGridCell A.G c.1 c.2 ⊆ dyadicGridCell A.G A.R.1 A.R.2) ∧ (∀ c ∈ X, ∀ c' ∈ X, c ≠ c' → Disjoint (dyadicGridCell A.G c.1 c.2) (dyadicGridCell A.G c'.1 c'.2)) ∧ ∑ c ∈ X, dyadicGridLength A.G c.1 ≤ 1 / 4 * dyadicGridLength A.G A.R.1 ∧ ∑ I ∈ {I ∈ A.cells | ¬∃ c ∈ X, dyadicGridCell A.G I.1 I.2 ⊆ dyadicGridCell A.G c.1 c.2} with (A.active δa) I, dyadicGridLength A.G I.1 * min A.η ‖A.Hf I (A.z p.castSucc I)‖ ≤ ↑Nl * Bnd + Mp * dyadicGridLength A.G A.R.1 * W

Blueprint lines 2839-2841 on one root ("choose their thresholds so the union of exceptional roots has length at most |R|/4"), assembled with the localization to the search leaves and the leaf bounds: the failure roots (length at most |R|/8) and the six cut families of each of the N_l leaves (lengths at most 2ρ|R|, 12 N_l ρ ≤ 1/8) have maximal union cells X of length at most |R|/4, and the active minima of the pattern below no cell of X are at most N_l times the leaf bound plus the localization error M_p |R| W.

A basic projection without columns vanishes almost everywhere: its current projection lies in the span of no columns.

theorem TrilinearHT.Auto.ActiveSetup.Hf_z_eq_zero_of_isEmpty_vtx {κ : LocalKernelClass} {c₀ : ℕ} (A : ActiveSetup κ c₀) (bf : A.BasicFamilies) (hV : ¬Nonempty bf.Vtx) (p : Fin 3) {I : ℤ × ℤ} (hI : I ∈ A.cells) :
A.Hf I (A.z p.castSucc I) = 0

Without vertices of the basic families (no basic columns) all pattern forms vanish on the cells of the setup: every pattern has a basic slot, and the basic arguments vanish there.