computation.md

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

6248 bytes, as of the version from 2026-11-15 19:16 (current). Recorded here, not run.

# Certified computation of T716

Instruction provenance is skill-manifest.json, copied from
agents/runs/20261115T165134Z-build-skills/manifest.json. All archived resources
were hash-checked and the valid local/live bundles agree, digest
d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab.

## Convention and reduction

CGLXZ arXiv:2210.14704v1 Lemma 2.2 uses z with v=+/-z^2. Put
z=sqrt(abs(v)), theta=asin(2z) for the positive sign or asinh(2z) for the
negative sign. Let M_(m,n)(x) be minus the integral of
log^(m-1)(2 sin(t/2)) log^(n-1)(2 cos(t/2)) over [0,x], or the sinh/cosh
analogue. At a zero of either kernel the endpoint integral is improper.
Write L=log(2sin(theta)) or log(2sinh(theta)). Then (2.4)-(2.5) give

    B_p(v) = 1/(z (p-1)!) sum_{j=1}^p (-1)^(j-1) binom(p-1,j-1) L^(p-j)
               [ M_(j,2)(2theta)/2
                 - sum_{l=1}^j binom(j-1,l-1) M_(l,j-l+2)(theta) ].

At abs(v)=1/16, L=0 exactly; only j=p contributes (0^0 here is 1).
At abs(v)=1/4, L=log(2). No rounded trigonometric argument or inferred zero
is used. For the positive endpoint, M(pi)=M(pi/2)+M with swapped indices
at pi/2, by reflection. Thus every direct trigonometric segment has endpoint
at most pi/2. Every hyperbolic segment ends at most 2asinh(1)<1.763.

## Finite series with the full omitted tail

Set a to a segment endpoint and u=t/a, w=u^2. Put q=(a/(2pi))^2.
The sine/cosine products DLMF 4.22.1-2 give

    log(2sin(t/2)) = log(u)+log(a)-sum_{j>=1} zeta(2j) q^j w^j/j,
    log(2cos(t/2)) = log(2)-sum_{j>=1}(4^j-1) zeta(2j)q^j w^j/j.

The sinh/cosh products DLMF 4.36.1-2 give the same coefficient magnitudes,
with sign (-1)^(j+1) instead of minus. In all used segments 8q<1.
At rho=2, zeta(2j)<2 gives absolute coefficient majorants

    A=-2log(1-2q),   B=-2log(1-8q).

For r=m-1,s=n-1, expand in powers log(u)^i. After truncating each product
at degree N, every coefficient through N is unchanged by intermediate
truncation. The absolute sum of omitted coefficients, weighted by 2^j,
is bounded by binom(r,i)(abs(log(a))+A)^(r-i)(log(2)+B)^s.
For 0<=w<=1, multiply this bound by 2^(-N-1). Since

    integral_0^1 u^(2j) log(u)^i du = (-1)^i i!/(2j+1)^(i+1),

an upper bound for the entire integrated omitted tail is

    a 2^(-N-1) (log(2)+B)^s
        sum_{i=0}^r binom(r,i) i! (abs(log(a))+A)^(r-i).

It includes cross-products of omitted terms and the logarithmic endpoint.
Arb encloses pi, exact-rational arguments, square roots, logarithms, zeta
values, polynomial arithmetic, moments and the final linear combinations.
The upper-rounded tail bound is added to each segment's ball radius before
combination. 100 requested digits use 557 working bits and N=605;
140 requested digits use 690 bits and N=738. The measured minima are in
numerical-evidence.json. Runtime assertions require finite native RealBall
results with at least bits(digits+20) relative accuracy. The final decimal
interval must contain the complete computed enclosure, checked separately.

## Independent Mellin quadrature and its tails

From CGLXZ (2.8), define s=sqrt(1-4q) and
G(q)=log((1+s)/(2s^2))/s. Its series is sum_{n>=1} binomial(2n,n)H_(2n)q^n.
The moment identity integrates exp(-(2n+1)t)t^(p-1)/(p-1)! to
1/(2n+1)^p. Thus

    B_p(v)=1/(p-1)! integral_0^infinity t^(p-1) exp(-t) G(v exp(-2t)) dt.

checks.py implements this formula, without importing generate.py, with
t=exp(x), x in [-800,log(160)], using 500-bit Arb analytic quadrature.
It computes 1-exp(-2t) as 2exp(-t)sinh(t) to prevent catastrophic cancellation
near t=0,v=1/4 (ComplexBall.expm1 is absent in this runtime).
The principal square root and logarithm are positive real on the integration
path; the quadrature's analytic flag checks branch crossings in complex
neighborhoods. The result must be finite and its imaginary part contain zero.

For 0<t<=1, absolute coefficients imply abs(G(v exp(-2t)))<=G(exp(-2t)/4).
With s^2=1-exp(-2t)>=t/2 and s<=1,
G<=sqrt(2/t)log(2/t). Put alpha=p-1/2, delta=exp(-800). The omitted lower
integral is bounded by

    sqrt(2) delta^alpha [ (log(2)+800)/alpha+1/alpha^2 ]/(p-1)!.

For t>=1, H_(2n)<=2n and binomial(2n,n)<=4^n imply
abs(G)<=2exp(-2t)/(1-exp(-2t))^2<3exp(-2t). The omitted upper integral at
T=160 is at most

    3^(1-p) exp(-480) sum_{k=0}^{p-1}480^k/k!.

Both positive bounds are added to the quadrature enclosure. Each stored
decimal's exact last-place interval must contain that full narrow enclosure.
All 16 interior rows also receive a direct coefficient sum through n=259,
using exact QQ recurrences for binomial and harmonic coefficients. Its tail
is bounded by 2 q^N[N/(1-q)+q/(1-q)^2], q=1/4,N=260.

The checker requires a narrow nonzero control B_1(1/4)=2 Catalan, obtained
independently as twice Im(Li_2(i)), and the two proved Sun identities:

    9B_3(1/16)+41U_4(1/16)=40 beta(4)+5pi zeta(3)/12,
    5B_2(-1/16)+17U_3(-1/16)=14 zeta(3).

U is recomputed by direct coefficients, not read from T715. beta(4) is
Im(Li_4(i)). The first identity follows from (1.1) by H_(2n+1)=H_(2n)+1/(2n+1),
the second from (1.2). B_1(1/4)=2 Catalan also follows from integrating
2[log cos(theta/2)-log cos(theta)] over [0,pi/2]. Checks are repeated using
stored intervals. Wrong sign, adding one, adding five last-place units,
a vacuous-radius ball and NaN must all be rejected. Generator and checker
share Arb and the generating-function/moment identity; the checker is an
independent numerical implementation, not an independent actor's verdict.

## Reused code provenance

The polynomial/log-moment method and reflection structure adapt
https://numberdb.org/files/T699/generate.py?raw=1, inspected alongside its
certification.md. The hyperbolic coefficient signs were checked against
DLMF and https://numberdb.org/files/T714/generate.py?raw=1. The T714 bound
restricted its arguments to q_cos<1/10; that restriction would not cover
2asinh(1), so it was not reused. The uniform weighted-majorant bound above
covers both signs and is rederived here. checks.py adapts decimal interval
parsing and controls from the T715 checker, but its Mellin integrand and two
tail bounds are new. Original downloaded sources and hashes are retained in
the run artifacts. No sibling data or review establishes mathematical truth.