First properties of the extremal test function and the rescaling #
The extremal test function is strictly positive on its support — which is what makes its square root smooth, and hence the whole kernel construction possible — and the rescaling sends a zero to a real point exactly when that zero lies on the critical line.
From those, the normalised cutoff test function η_ψ and its square f_ψ are shown to be smooth,
compactly supported in (-1/2, 1/2) and even — and then so is the self-convolution Q_ψ, on
(-1, 1), together with its second derivative.
The extremal test function is strictly positive on [-1/2, 1/2].
The normalised cutoff test function is smooth, compactly supported and even #
The delicate point is ±1/2: there √f₀ is not smooth, because f₀ jumps
from a positive value down to 0. What saves it is that a δ-cutoff VANISHES on a neighbourhood of
every such point, so the product is locally constant there.
Stated without 0 < δ < 1/4 and without positivity of A_ψ: neither is needed. When A_ψ = 0 the
division gives η_ψ = 0, which is smooth, supported anywhere and even.
Positivity of the normalising denominator of f₀, read off from extremalTest_pos at 0
rather than re-proved, so the library holds only one proof of it.
Off [-1/2, 1/2] the cutoff kills the product, so on ALL of ℝ the test function agrees with
the expression that has f₀'s if removed. That representative is smooth wherever
cos (√2 x) > 0, and the cutoff handles everywhere else.
η_ψ is smooth.
η_ψ has compact support -- the C_c half of C_c^∞.
f_ψ is smooth.
f_ψ has compact support.
The self-convolution Q_psi: smoothness, support, parity #
Q_psi is f_psi convolved with itself, so Mathlib's convolution theory applies once f_psi is
known smooth with compact support, which is cutoffTestSq_contDiff and
cutoffTestSq_hasCompactSupport.
The support claim doubles: f_psi lives in (-1/2, 1/2), so the convolution lives in (-1, 1).
The elementary reason is that the integrand f_psi t * f_psi (x - t) is identically zero when
1 <= |x|: either |t| >= 1/2, killing the first factor, or |t| < 1/2 and then
|x - t| >= |x| - |t| > 1/2, killing the second.
For Q_psi'' the support and parity are inherited rather than recomputed: derivatives cannot
enlarge the support (tsupport_deriv_subset), and iteratedDeriv_comp_neg turns evenness of
Q_psi into evenness of an even-order derivative.
Outside [-c, c] the absolute value is at least c. Used repeatedly to turn a
HasCompactSupport.intro obligation into the pointwise vanishing statement.
Q_psi is Mathlib's convolution of f_psi with itself. Definitional, but it is what makes the
convolution smoothness theorem applicable.
Q_psi is smooth.
Q_psi has compact support.
Q_psi is Lipschitz continuous at 0 -- Lipschitz with some constant on some
neighbourhood of 0, which is what "Lipschitz continuous at 0" means here.
Q_psi'' is smooth.
Q_psi'' is Lipschitz continuous at 0.