certification.md

back to table · edit · history · where entries came from · files · download

5083 bytes, as of the version from 2026-11-01 21:11 (current). Recorded here, not run.

# Numerical certification of T659

## Main computation

The attached generate.py computes the Taylor coefficients through order 56 of

    (s(s−1)/2) exp(−s log(pi)/2) Gamma(s/2) zeta(s),  s=1/2+z,

using Arb's analytic polynomial series at 896 bits. The extracted coefficients are multiplied by j!, giving gamma_j for 0 ≤ j ≤ 28. Every odd coefficient enclosure contains zero, as required by xi(1−s)=xi(s), and every gamma_j is positive and has a finite narrow enclosure. Polynomial multiplication only needs the coefficients up to the requested order; no truncated series is treated as an approximation to a function value.

For each degree d=1,…,8 and shift n=0,…,20, complex polynomial root isolation supplies d candidates. Around their real midpoints the program constructs intervals with relative radius approximately 2^(−390), then evaluates the real polynomial at both exact dyadic endpoints. Opposite strict signs give at least one real root in each interval. The intervals are negative and mutually disjoint, and the leading coefficient is positive. A degree-d polynomial cannot have more than d roots counting multiplicity. Thus these d intervals contain exactly one simple root each and exhaust the polynomial's roots. This argument proves the finite coverage without assuming RH or relying on the complex candidates being real.

The client converts these narrow enclosures to 100-significant-digit decimal strings. Each string means the interval with radius one unit in the last written place. checks.py tests the signs at those actual decimal endpoints, using independently obtained coefficients. It also rejects intervals translated by plus or minus ten last-place units.

## Independent theta integral

Griffin, Ono, Rolen and Zagier, arXiv:1902.07321v2, section 4, give the split Mellin integral. Algebraic rearrangement yields

    xi(1/2+z) = 1/2 + (z²−1/4) integral_1^infinity u^(−3/4) theta_0(u) cosh(z log(u)/2) du,
    theta_0(u) = sum_(ell≥1) exp(−pi ell² u).

Define I_j as the integral of u^(−3/4) theta_0(u) (log(u)/2)^(2j)/(2j)!. Then

    gamma_0 = 1/2 − I_0/4,
    gamma_j = j! (I_(j−1) − I_j/4),  j≥1.

This computation uses no zeta or gamma function evaluation. For every j=0,…,28, rigorous complex-ball quadrature integrates the first 12 theta terms over [1,256], split at powers of two, at 832 bits. The integrand's logarithm honors Arb's analytic flag. Both quadrature tolerances are 2^(−750). Tolerances are requests, not error assertions; finite enclosure widths are checked afterward.

Here are explicit upper bounds for the omitted contributions. Put m=2j, A=169 pi, and U=256. For ell=13+r with r≥0, ell²≥169+27r. Hence the omitted theta sum for u≥1 is at most exp(−Au)/(1−exp(−27pi)). Since u^(−3/4)≤1, log(u)≤u and u^m≤exp(m(u−1)), its contribution, even integrated over all [1,infinity), is bounded by

    E_theta = exp(−A) / ((1−exp(−27pi))(A−m) 2^m m!).

For the whole theta sum on u≥U, ell²≥1+3(ell−1) gives theta_0(u)≤exp(−pi u)/(1−exp(−3pi U)). The inequality u^m≤U^m exp(m(u−U)/U) then bounds the omitted integral by

    E_integral = U^m exp(−pi U) / ((1−exp(−3pi U))(pi−m/U) 2^m m!).

All denominators are positive for 0≤m≤56. These conservative bounds are evaluated in ball arithmetic. The upper endpoint of their sum is added as a symmetric error radius to the quadrature result. Overcounting omitted theta terms beyond U only enlarges the enclosure. No decay term is silently discarded.

## Controls and checks

The integral method first integrates exp(−pi u) on [1,256] and compares with its known nonzero exact antiderivative difference. It also integrates −u² on [0,1], obtaining −1/3. These controls precede any agreement claim. All 29 independently enclosed coefficients must overlap the analytic series enclosures, and both must pass explicit relative-accuracy thresholds. The source's tabulated 2019 coefficient at index 10 is checked after multiplying our gamma_10 by eight.

For every written root interval, the independent coefficients must give strict opposite endpoint signs. Distinctness, negativity, all Vieta coefficients, the polynomial derivative identity, and the linear and quadratic closed forms are checked throughout the finite range. The main coefficients are also recomputed at 1024 bits as a supplemental precision check. Increasing precision is not treated as the independent method.

The post-fill run repeats the tests on the exact strings read back from T659 and runs Generator.verify(sample=None). No published table is modified. T15 is context and a comparison source only, never a certified dependency.

Policy evidence is the archived manifest `agents/runs/20261101T195835Z-build-skills/manifest.json`; both skill versions have SHA-256 `a2b28d9953264d7b0ca62663785c2952190c5be6d6d3597fd695d9a54d606278`. Numerical outcomes and dependency versions are preserved in the attached evidence. This document describes this build's checks, not the separate critique required by the controller.