certification.md

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

6260 bytes, as of the version from 2026-11-10 11:43 (current). Recorded here, not run.

# Certification of T704

The quantity is the raw discriminant of J^(d,n), using the completion
xi(s)=s(s-1) pi^(-s/2) Gamma(s/2) zeta(s)/2. In particular its leading
coefficient gamma_(n+d) is retained. Multiplying every gamma by eight, as
in the 2019 paper, multiplies this discriminant by 8^(2d-2).

## Primary computation

Arb analytic polynomial series expand the completed product at exactly
s=1/2 through order 56, at 896 bits. Even coefficients multiplied by m!
give gamma_0,...,gamma_28; odd coefficients contain zero. Truncation here
extracts Taylor coefficients and is not an approximation at a nonzero z.
All coefficient enclosures are finite, positive, and have at least 542 bits
of relative accuracy. No T15 decimal is used as an input to the generator.

For every degree 2<=d<=8 and shift 0<=n<=20 the program forms the Sylvester
matrix of J and J'. Ball Gaussian elimination chooses pivots excluding
zero, and computes its determinant. The resulting value is
(-1)^(d(d-1)/2) Res(J,J')/gamma_(n+d). Division is between ball values;
integer factorials and binomial coefficients never undergo Python float
division. The method propagates all coefficient errors through the
resultant, with no point approximations substituted for intervals.

The weakest discriminant enclosure retains 450 bits, at (d,n)=(8,20).
The total loss from 896 working bits is therefore at most 446 bits across
this finite block. 450 bits exceed the roughly 333 bits required for 100
significant decimal digits. The independent check tests containment in the
actual serialized decimal intervals, not merely overlapping enclosures.

## Independent coefficient route

The theta-moment implementation is adapted from the public T659 checks.py,
retrieved on 2026-11-10. Its derivation was checked against Griffin, Ono,
Rolen and Zagier, arXiv:1902.07321v2, section 4. Their coefficient sequence
is eight times the sequence used here; the split Mellin formula gives

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

Let I_j be the integral of u^(-3/4) theta_0(u)
(log(u)/2)^(2j)/(2j)!. Then gamma_0=1/2-I_0/4 and

    gamma_j = j! (I_(j-1)-I_j/4), j>=1.

This method evaluates no zeta or gamma function. It integrates the first
12 theta terms over [1,256], splitting at powers of two, with rigorous
complex-ball quadrature at 832 bits. The logarithm respects the analytic
flag. Absolute and relative tolerances are 2^(-750); measured enclosure
widths, rather than those requested tolerances, establish the accuracy.

Here are the explicit omitted-tail bounds. Put m=2j, A=169*pi and U=256.
For ell=13+r, ell^2>=169+27r. Therefore the omitted theta terms, integrated
over all u>=1, contribute at most

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

Indeed u^(-3/4)<=1, log(u)<=u, and
u^m<=exp(m(u-1)) for u>=1. For the whole theta sum beyond U,
ell^2>=1+3(ell-1), and u^m<=U^m exp(m(u-U)/U), giving

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

Both denominators are positive for 0<=m<=56. The bounds are evaluated as
balls and their summed upper endpoint is added as a symmetric error radius.
Double-counting omitted theta terms beyond U is conservative. The 29
resulting gamma enclosures retain at least 496 relative-accuracy bits and
agree with the analytic-series route. A 1024-bit series rerun supplies a
supplemental precision check. The 29 central Taylor coefficients are also
checked against the actual stored T15 intervals.

## Independent discriminant route

For each polynomial, roots are recomputed using only the theta coefficients.
Complex ball roots supply candidates. Around their real midpoints, a radius
2^(-390)*max(1,abs(midpoint)) gives exact dyadic endpoint brackets. Strict
opposite endpoint signs certify a real root inside each bracket. All d
brackets are negative and pairwise disjoint. Since the leading coefficient
is nonzero, a degree-d polynomial has exactly d roots with multiplicity;
these brackets therefore exhaust its roots and prove simplicity. This does
not assume RH or infer realness from a small imaginary part.

The leading coefficient to power 2d-2 times the product of squared root
differences gives an independent discriminant enclosure. Its least relative
accuracy is 380 bits. Every such enclosure lies strictly inside the interval
centered on the written 100-digit value with radius one last-place unit.
The comparison cannot be passed by a wide or nonfinite interval. The newly
certified roots also lie inside all 735 corresponding T659 decimal intervals;
existing root decimals are not precision inputs to the discriminants.

## Controls and identities

Before comparisons, the integral method checks the nonzero integral of
exp(-pi*u) from 1 to 256 against its elementary antiderivative, and the
integral of -u^2 from 0 to 1 against -1/3. The discriminant implementation
checks x^2-1 and x^3-x (both 4), 2*x^2-2 (16), 2*x^2+2*x+1 (-4), and
(x^2+1)(x^2+4) (5184). The last polynomial has no real roots, so a positive
quartic discriminant alone does not establish hyperbolicity.

The checks cover all Vieta coefficients, all 21 quadratic and 21 cubic
Turan identities, all 147 scaling-by-eight identities, and all 84
specializations of the stored universal discriminants T381 in degrees
2 through 5. Translations of every discriminant decimal interval by plus
or minus ten last-place units are rejected (294 controls). All of these
checks are repeated on the strings read back from T704 after filling.
Generator.verify(sample=None) independently checks storage against the
primary computation. Logs retain the individual coefficient enclosures,
tail bounds and per-entry accuracy measurements.

## Scope and policy

The argument-selection document explains both free indices and their finite
bounds. These checks validate this build and do not constitute the separate
campaign critique. No offering, reviewing or publication operation is made.

The run uses the archived manifest
agents/runs/20261110T024815Z-build-skills/manifest.json. Both archived skill
versions and the source snapshot's canonical skill have SHA-256
886390324dd1971e381a43cf3b46b822ab0bc67d62f31d8481e29ec79556746b.