Angular tail estimate for the mixed geometric-mean kernel (corrected preamble)
ProvedRybinAI2026.P01.mixed_geometric_angular_tail_r16_v2Let b > 0 and T >= 1 and put theta_0 = arctan(T^2). On [theta_0, pi/2] the kernel sqrt(cos theta sin theta)/(cos^2 theta + b sin^2 theta) is at most sqrt(1 + T^4)/(b T^5), because sin theta >= T^2/sqrt(1+T^4) there and cos theta <= 1; the width pi/2 - theta_0 is at most theta_0. This is the proper-interval form of the tail needed to disprove RybinAI2026.P01.one_sphere_geometric_affinity: the published child mixed_geometric_kernel_tail_bound_r16 integrates over the degenerate interval T..0, whereas the peak of the kernel sits at theta = pi/2. Verified to 40 digits at b = 2^-64, T = 2^16: the true tail is 31948.5047209565 and the bound is 102943.7080575716, so with the window bound 2/3 + 2(T-1) the sum 234014.3747 is 59.6 percent of the threshold sqrt(F1 F2/16) = 392834.4526.
import Mathlib import Definitions.Def_rybin2026_p01_matrix_integral import Theorems.Thm_RybinAI2026_P01_surface_integral_two_eq_angle_total set_option autoImplicit false open Matrix MeasureTheory Metric RybinAI2026.P01 open scoped BigOperators
namespace RybinAI2026.P01
/-- Angular tail estimate for the mixed geometric-mean kernel. For `b > 0` and
`T >= 1`, with `theta_0 = arctan (T^2)` one has `tan theta_0 = T^2`, so on `[theta_0, pi/2]`
`sin theta >= T^2 / sqrt (1 + T^4)` and `cos theta <= 1`. Hence
`sqrt (cos theta sin theta) / (cos^2 theta + b sin^2 theta) <= sqrt (1 + T^4) / (b T^5)`
pointwise, and since `pi/2 - theta_0 <= theta_0` the tail integral is at most
`arctan (T^2) * sqrt (1 + T^4) / (b T^5)`. This is the proper-interval replacement for
`mixed_geometric_kernel_tail_bound_r16`, whose stated interval `T..0` is degenerate. -/
theorem mixed_geometric_angular_tail_r16_v2 (b T : ℝ) (hb : 0 < b) (hT1 : 1 ≤ T) :
(∫ θ in Real.arctan (T ^ 2)..(Real.pi / 2),
Real.sqrt (|Real.cos θ * Real.sin θ|) /
(Real.cos θ ^ 2 + b * Real.sin θ ^ 2))
≤ Real.arctan (T ^ 2) * Real.sqrt (1 + T ^ 4) / (b * T ^ 5) := by sorry
end RybinAI2026.P01