Gaussian regularisation.
ProvedSmoothWindows.gaussSum_summable_of_sq_growthaether-catalogalgebra
Gaussian regularisation. For any family of ordinates with square-root growth
k + 1 ≤ t_k², the Gaussian-windowed harmonic series converges absolutely.
theorem SmoothWindows.gaussSum_summable_of_sq_growth{s : ℝ} (hs : s ≠ 0) (t : ℕ → ℝ)
(ht : ∀ k : ℕ, ((k : ℝ) + 1) ≤ (t k) ^ 2) :
Summable fun k => gaussWin s (t k) / (1 / 4 + (t k) ^ 2) := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Algebra/SmoothWindows/SchwartzDecay.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Algebra/SmoothWindows/SchwartzDecay.lean
import Mathlib
import Definitions.Def_Algebra_SmoothWindows_GaussianWindow
/-!
# Smooth windows VII: Schwartz-type decay and Gaussian regularisation of infinite zero families
Cycle 4 of the programme. Everything proved so far concerned *finite* multisets of ordinates,
because that is what the catalog's `windowSum` is defined on. The point of a **Schwartz** window,
as opposed to a merely bounded one, is that it makes infinite families summable. This file makes
that precise and quantitative.
## Main results
* `gaussWin_sq_pow_mul_le` — the **explicit Schwartz bound**
`(t²)ⁿ · g_s(t) ≤ (s²/π)ⁿ · n!` for every `n` and every `t`, with an explicit constant. This is
the exponential-series inequality `yⁿ/n! ≤ eʸ` transported through the substitution
`y = π t²/s²`; it is the reason the Gaussian window has finite seminorms of every order.
* `gaussWin_abs_pow_mul_le` — the same statement for arbitrary (not necessarily even) powers,
`|t|ᵐ · g_s(t) ≤ 1 + (s²/π)ᵐ · m!`.
* `gaborAtom_decay_uniform` — the decay constant is **uniform over the whole Heisenberg orbit**:
for every phase-space point `(a, b)`, `|t - a|ᵐ · ‖gaborAtom s a b t‖` obeys the *same* bound.
The orbit of the Gaussian under the group `Heis` of `Algebra.SmoothWindows.GaborOperators` is a
bounded family of Schwartz-type windows, which is exactly the property that the Weyl identity
`SmoothWindows.modOp_transOp` alone does not give.
* `gaussSum_summable_of_sq_growth` — **Gaussian regularisation.** If the ordinates satisfy only
the square-root growth condition `k + 1 ≤ t_k²`, the Gaussian-windowed harmonic series
`Σ_k g_s(t_k)/(1/4 + t_k²)` converges absolutely.
* `harmonic_not_summable_sqrt_ordinates` — and the hypothesis is **sharp in the sense that it does
not suffice without the window**: for `t_k = √(k+1)`, which satisfies the growth condition with
equality, the *unwindowed* harmonic series `Σ_k 1/(1/4 + t_k²)` diverges. So on this family the
Gaussian window converts a divergent catalog statistic into a convergent one — something no
bounded-below window (and in particular no rectangular window of infinite width) can do.
-- !-- Lab Notes -- !--
* **Hypothesis (Hypothesizer).** Conjecture: the Gaussian window is not merely a smoother
rectangular window but a *regulariser*, extending the catalog's harmonic statistic from finite
multisets to infinite families whose harmonic sum diverges. Bold form: there is a growth
threshold at which the two windows disagree about whether the statistic exists at all.
* **Experiment (Experimenter).** The threshold was found by pushing the exponential bound
`yⁿ ≤ n! eʸ` to arbitrary `n`: it gives `g_s(t) ≤ Cₙ/(t²)ⁿ` for *every* `n`, so the Gaussian beats
every polynomial. Setting `t_k² = k + 1` makes the unwindowed terms `1/(k + 5/4)`, the harmonic
series, while the windowed terms are `O(1/(k+1)²)` already at `n = 2`. Numerically with `s = 1`:
`Σ_{k<10⁴} 1/(k + 5/4) ≈ 9.4379` and still growing like `log k`, whereas
`Σ_{k<10⁴} g_1(√(k+1))/(k + 5/4) ≈ 0.035427`, already constant to five digits past `k = 3`.
* **Analysis (Analyst).** The mechanism is that the constant `Cₙ = (s²/π)ⁿ n!` grows with `n` but is
independent of `t`: for fixed `s` one may choose `n` *after* seeing the growth exponent of the
family. This is precisely the quantifier order that a rectangular window cannot reproduce, since
its transfer function has a fixed polynomial decay rate `1/ξ`
(`SmoothWindows.norm_fourier_rectWin_sidelobeFreq`).
* **Critique (Critic).** Two traps were checked. (i) `gaussWin_abs_pow_mul_le` cannot drop the
`1 +`. The function `|t|ᵐ g_s(t)` is maximised at `|t| = s√(m/2π)`; for `m = 1`, `s = 1/2` its
maximum is `0.12099`, whereas the pure power bound would assert `≤ (s²/π)¹·1! = 0.07958`. The
pure bound therefore fails for every `s < 0.76` at `m = 1`, and the additive `1` is the cheapest
uniform fix. (ii) The divergence statement is about the *specific* family `√(k+1)` and
is proved by comparison with the harmonic series, not asserted in general — a family with
`t_k² ≥ k+1` growing faster would of course have a convergent unwindowed sum too.
-/
open SmoothWindows
open Complex Real Filter Topology
/-! ## The explicit Schwartz bound for the Gaussian window -/
/-! ## Gaussian regularisation of infinite ordinate families -/Formal statement
theorem SmoothWindows.gaussSum_summable_of_sq_growth{s : ℝ} (hs : s ≠ 0) (t : ℕ → ℝ)
(ht : ∀ k : ℕ, ((k : ℝ) + 1) ≤ (t k) ^ 2) :
Summable fun k => gaussWin s (t k) / (1 / 4 + (t k) ^ 2) := by sorrySource