# Certification of the xi derivative zero ordinates in T660

The instruction source is the coordinated archive
`agents/runs/20261101T203350Z-build-skills/manifest.json`. Its local and live
SHA-256 values both equal
`a2b28d9953264d7b0ca62663785c2952190c5be6d6d3597fd695d9a54d606278`.
The independently reproducible mathematical certificate is as follows.

## Analytic evaluator

Set Xi(t)=xi(1/2+it), with xi(s)=s(s-1) pi^(-s/2) Gamma(s/2) zeta(s)/2.
Arb analytic power series for the exponential, gamma and zeta factors enclose
all needed derivatives. The coefficient of h^j in Xi(t+h), multiplied by j!,
is Xi^(j)(t). These derivatives are real on the real axis. The s derivative
satisfies Xi^(j)(t)=i^j xi^(j)(1/2+it). Both directions of differentiation
are checked independently in checks.py at all selected ordinates.

## Counting all real roots in a finite interval

Every half-unit cell of [0,90] is covered by a degree-95 Taylor polynomial
at its midpoint c, computed at 256 bits. Its coefficients carry ball bounds.
For the disk |z-c|<=1, Cauchy's estimate bounds the Taylor coefficient a_n
by M, where M bounds |Xi(z)| on the disk. The product formula evaluated on a
containing complex rectangle supplies one such M whenever finite.

There is also an entire-function bound, valid even if that rectangle crosses
a removable pole of a factor. The Mellin representation (Griffin, Ono,
Rolen and Zagier, arXiv:1902.07321v2, section 4) implies

    Xi(z) = 1/2 - (z^2+1/4) I(z),
    I(z) = integral_1^infinity u^(-3/4) theta_0(u) cos(z log(u)/2) du,
    theta_0(u) = sum_{ell>=1} exp(-pi ell^2 u).

For u>=1, theta_0(u)<=2 exp(-pi u). On the disk, |Im z|<=1 and
|cos(z log(u)/2)|<=u^(1/2), hence |I(z)|<=2 exp(-pi)/pi. Therefore

    M <= 1/2 + ((|c|+1)^2+1/4) 2 exp(-pi)/pi

is a valid choice of upper bound (the right side is evaluated outward).
The generator may use the smaller of the two proven bounds.

For N=96, derivative order m<=5 and |x|<=h<=1/4, the omitted tail obeys

    |sum_{n>=N} (n)_m a_n x^(n-m)|
       <= M N^m m! h^(N-m) / (1-h)^(m+1).

Indeed, (N+j)_m <= (N+j)^m <= N^m (j+1)^m and
sum_{j>=0}(j+1)^m h^j <= m!/(1-h)^(m+1). The code adds this bound as a
radius to interval Horner evaluation of each differentiated polynomial.
This gives enclosures over the full cell, not only sampled points.

For every cell, one of the following certifies its root count:

* Xi^(m) excludes zero throughout the cell, so it has no root there.
* Xi^(m+1) excludes zero, so Xi^(m) is strictly monotone. Equal nonzero
  endpoint signs prove no root; opposite signs prove exactly one simple root.

Unresolved cells would be bisected. In the completed run no bisections were
needed. At odd m, evenness of Xi gives Xi^(m)(0)=0. A nonzero enclosure of
Xi^(m+1) on [0,1/1024] excludes another root there. The rest is covered
without gaps. The origin is not counted as a positive root.

| m | zero-free cells | monotone empty cells | simple positive roots | origin handled |
|---|---:|---:|---:|---:|
| 1 | 150 | 5 | 25 | yes |
| 2 | 153 | 2 | 25 | no |
| 3 | 149 | 6 | 25 | yes |
| 4 | 151 | 4 | 25 | no |

The 25 roots certified for each order are verification coverage. Only their
first 20 are table entries. Thus every selected k is certified as a positive
critical-line index, independently of the Riemann hypothesis. This proof
makes no statement about zeros off the critical line.

## Refinement and the decimal claim

Eight bisections and twelve Newton iterations suggest a root centre at 512
working bits. These iterations are not the certificate. An interval of
radius 10^(-114) is then checked by strict opposite endpoint signs and a
nonzero next derivative throughout it. The resulting balls retain at least
379 relative accuracy bits and are serialized to 100 significant digits.
The stored decimal denotes its centre plus or minus one last-place unit.
Every such interval is certified again at 640 bits. Both intervals shifted
by ten last-place units are rejected for every entry.

## Independent theta-integral check

checks.py differentiates the Mellin integral directly, without using the
generator's Taylor coefficients or its special-function evaluator. Put
L=log(u)/2 and I_j(t)=integral u^(-3/4) theta_0(u) L^j cos(tL+j pi/2) du.
Then, for m>=1,

    Xi^(m)(t) = -(t^2+1/4) I_m(t)
                -2 m t I_(m-1)(t) -m(m-1) I_(m-2)(t),

with the last term absent at m=1. At m=0 add the constant 1/2. The
quadrature uses 576-bit complex balls on [1,2], [2,4], ..., [64,128], with
absolute and relative requested tolerances 2^(-520). The logarithm passes
through the quadrature callback's analytic flag, as required for its branch.
Failure to achieve useful error bounds is rejected, never treated as agreement.

The theta sum retains ell=1,...,10. For u>=1,

    sum_{ell>=11} exp(-pi ell^2 u) <= 2 exp(-121 pi u).

For each j, u^(-3/4)(log(u)/2)^j <= u^j. The combined error from missing
theta terms and u>=128 is therefore bounded by

    B_j = 2 [E_j(121 pi,1)+E_j(pi,128)],
    E_j(a,U) = exp(-a U) sum_{r=0}^j j! U^(j-r)/((j-r)! a^(r+1)).

This is the exact integral of u^j exp(-a u) from U to infinity. The full
Xi^(m) error is at most

    (t^2+1/4) B_m + 2m |t| B_(m-1) + m(m-1) B_(m-2).

All bounds are evaluated in balls and their upper endpoints are added to
the final radius. Every independent enclosure must be finite with radius
less than 10^(-145). Opposite signs at the two endpoints of every written
decimal interval independently certify its digits. The same checks are
repeated on the records fetched from the draft.

Known nonzero quadrature controls are integral_1^2 u du=3/2 and
integral_1^2 exp(-pi u) du=(exp(-pi)-exp(-2pi))/pi. An intentionally wrong
answer 5/2 is rejected. The theta integral also reproduces Xi(0), Xi''(0)
and Xi''''(0), including their signs, from the separate analytic evaluator.
Parity and the s-versus-t chain rule are checked throughout the selected range.

No database coefficient approximation is used as a certified dependency.
No numerical agreement alone is described as a rigorous bound. The analytic
root cover and independently bounded quadrature supply the error claims;
separate working precisions provide additional implementation checks.
