from mpmath import mp
mp.dps = 150
def IKM(a, b, k):
if not (0 <= a < b and k >= 0):
raise ValueError("Require 0 <= a < b and k >= 0")
return mp.quad(lambda t: t**k * mp.besseli(0,t)**a
* mp.besselk(0,t)**b, [0, 1, 4, 16, mp.inf])
print(IKM(0, 3, 1))Original 40 odd moments, retained unchanged:
Direct tanh-sinh quadrature at 150 and 200 working decimal digits, with Arb evaluations of the Bessel factors, gives 100 significant stored digits. The worst relative discrepancy between the two runs was 6.064e-152. Arb midpoints enter the quadrature, so this is agreement evidence, not a certified integration enclosure. The upper endpoint is truncated at T=3*(working digits+60); the generator documents an explicit exponentially small tail bound. The integrand tends to zero at the lower endpoint for every selected moment.
The nonzero controls IKM(0,1;1)=1 and IKM(0,2;1)=1/2 were tested before the table values. All 14 applicable recurrence relations and 11 closed-form comparisons, including exact multinomial counts, were tested on the computed values. A separate quadrature uses mpmath Bessel functions, Gauss-Legendre integration and t=exp(x) at 60 working digits; it compares all 40 entries with a relative tolerance of 1e-50. These numerical checks are reproduced in checks.py.
After submission, all 40 values were read back and found identical to the submitted decimal strings. The 25 identity checks and the independent quadrature were repeated on these stored values, and generator.verify(sample=None) matched all 40 entries.
Five even additions: the same Arb-Bessel/tanh-sinh method, ported from the Sage Arb interface to python-flint, runs at 150 and 200 working digits. For k=0 the substitution t=u^2 makes the transformed integrand tend to zero at the lower endpoint. The unchanged upper-tail estimate applies and is below 4e-254 and 2e-318 for the two runs. The largest relative 150/200-digit discrepancy is below 4.1e-152. Independent gamma and elliptic-integral calculations at 140 and 180 digits check all five additions, with relative differences below 2.2e-180 from the Bessel quadrature. These comparisons are not certified integration enclosures; the rigour label remains heuristic (agreement-checked). All 40 original values are reproduced within their stored last-place intervals, without altering their decimal strings. The nonzero controls IKM(0,1;0)=pi/2, IKM(0,2;0)=pi^2/4, IKM(0,1;1)=1 and IKM(0,2;1)=1/2 pass before the retained-row reproduction. The current default generator enumerates all 45 rows; even-checks.py and even-checks-evidence.json record the extension checks. The older checks.py describes the original odd slice.