The normalized negative-translation distance to geometric approximants vanishes
ProvedMasulloRH.translated_negative_gram_distance_to_geometric_approximantsfunctional-analysisharmonic-analysis
With w(x)=exp(|x|)(1+x²)², suppose a fixed f can be approximated arbitrarily closely by a class S in L²(exp(-x)dx), and all the residuals belong to L²(w dx). Then the infimum over S of the squared L²(w dx) error after translation by -L, divided by exp(L)L⁴, tends to zero. This bounds the actual spectral projection error when S is a translation-invariant geometric class. In the application the S00 packet supplies the approximation premise.
Preamble
import Mathlib open MeasureTheory Filter Topology
Formal statement
theorem MasulloRH.translated_negative_gram_distance_to_geometric_approximants (f : ℝ → ℂ) (S : Set (ℝ → ℂ))
(hf : AEStronglyMeasurable f volume)
(hs : ∀ u ∈ S, AEStronglyMeasurable u volume)
(hint : ∀ u ∈ S, Integrable (fun x : ℝ =>
(Real.exp |x| * (1+x^2)^2) * ‖f x - u x‖^2))
(happrox : ∀ ε : ℝ, 0 < ε → ∃ u ∈ S,
(∫ x : ℝ, Real.exp (-x) * ‖f x-u x‖^2) < ε) :
Tendsto (fun L : ℝ => sInf ((fun u : ℝ → ℂ =>
∫ x : ℝ, ((Real.exp |x-L| * (1+(x-L)^2)^2) / (Real.exp L * L^4)) *
‖f x-u x‖^2) '' S)) atTop (𝓝 0) := by sorrySource
Masullo (9.2.2.41); critical-width large-translation asymptotics for the global Gram form.