certification.md

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

4051 bytes, as of the version from 2026-11-15 19:16 (current). Recorded here, not run.

# Enclosures and checks

## Interior arguments

Let q=|u|/4<1. Since the central binomial coefficient is maximal in row 2n,
4^n <= (2n+1) binomial(2n,n). Also H_(n-1)^(r)<=n for r>=1. Thus for s>=1
an absolute summand is at most (2n+1)q^n. The entire tail beginning at N is
bounded by q^N((2N+1)/(1-q)+2q/(1-q)^2). All partial sums and that bound are
computed in Arb. The selected N makes this bound <10^(-digits-35).
H is updated after adding a term, so the first term uses H_0=0. Binomial
reciprocals use the exact ratio n/(2(2n-1)), converted into balls before
any division. Inputs are exact QQ values; no binary float is introduced.

## Positive boundary

Write R(n)=4^n/binomial(2n,n)=sqrt(pi*n) E(n). From DLMF 5.11.1 and its
positive-real remainder bound in 5.11(ii),

log E(n) = L(1/n)+delta(n),
L(x)=sum_(k=1)^M (2-2^(1-2k)) B_(2k) x^(2k-1)/(2k(2k-1)),
|delta(n)| <= d n^(-m),
m=2M+1, d=(2+2^(-m)) |B_(2M+2)|/((2M+2)m).

The plus sign in d is intentional: bound the two log-Gamma remainders
separately, rather than assuming cancellation. Let P be the Taylor polynomial
of exp(L) before degree K. All its coefficients are exact rationals. For
rho>1/N, define E_rho=exp(sum_j |L_j|rho^j), a=1/(N*rho),
D=d N^(-m), L_N=sum_j |L_j|N^(-j). Cauchy's coefficient bound gives, for n>=N,

|E(n)-P(1/n)| <= E_rho*a^K/(1-a) + exp(L_N+D)D = epsilon_E,
|P(1/n)| <= E_rho/(1-a) = P_bound.

For r=1, DLMF 5.11.2 gives
H_(n-1)=log n+gamma-1/(2n)-sum_(k=1)^M B_(2k)/(2k n^(2k)).
The remainder is at most |B_(2M+2)|/((2M+2) N^(2M+2)) = epsilon_H.

For r>1, H_(n-1)^(r)=zeta(r)-zeta(r,n). Euler--Maclaurin gives the polynomial
zeta(r)-n^(1-r)/(r-1)-1/(2n^r)
-sum_(k=1)^M B_(2k) (r)_(2k-1)/((2k)! n^(r+2k-1)).
The periodic-Bernoulli integral remainder is bounded by the magnitude of
the last retained Bernoulli term at N, used conservatively as epsilon_H.
Indeed |B_(2M)({x})|<=|B_(2M)| follows by the triangle inequality in
the cosine Fourier series DLMF 24.8.1 (https://dlmf.nist.gov/24.8.E1),
whose absolute coefficient sum equals the value in magnitude at zero.
Thus integration of
(r)_(2M)|B_(2M)|/(2M)! x^(-r-2M) from N to infinity yields that bound.
This is also obtained directly by repeated integration by parts of the
Euler--Maclaurin representations in DLMF 25.11(iii); it does not treat an
asymptotic big-O symbol as a numeric bound.

Multiply P by this harmonic polynomial without truncating the product.
Each n^(-p-j) tail is zeta(p+j,N), where p=s-1/2. When r=1, the additional
log(n) term is -d/dp zeta(p+j,N), evaluated by Arb's Hurwitz zeta series
through degree one with an exactly real half-integer centre. Thus the sum
of the approximating expression is enclosed in ball arithmetic.

For all r>=1, 0<=H_(n-1)^(r)<=1+log n. Integral tests at N>=256 bound
J0=sum_(n>=N)n^(-p) by N^(-p)+N^(1-p)/(p-1), and
Jlog=sum_(n>=N)n^(-p)(1+log n) by
N^(-p)(1+log N)+N^(1-p)((1+log N)/(p-1)+1/(p-1)^2).
The final added radius is sqrt(pi)(epsilon_E*Jlog+P_bound*epsilon_H*J0).
This controls the whole infinite tail, including both product errors.
The finite part is the original series through n=N-1.

At 100 digits: K=128, M=64, N=256, rho=1/16. Working precision is
numberdb.bits(digits,losing=200) in the interior and losing=220 at the
endpoint. The generator scales these choices upward for higher precision,
and refuses a boundary enclosure lacking the requested accuracy.

## Scope and provenance

The 70 entries and their full balls are preserved in computed.json. Native
return types are RealBall; deterministic dry_run.py/check.py and numerical
comparisons have separate actual logs/results. The exact 100-digit encoding
is a decimal interval, with one unit of last place uncertainty. Read-back
checks must compare that interval, not treat its centre as exact.

Skill manifest: ../20261115T174305Z-build-skills/manifest.json, matching bundle
SHA-256 d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab.
The derivation is builder evidence. A fresh scientific critic must independently
assess it before the controller offers the draft.