back to table · edit · history · where entries came from · files · download
9967 bytes, as of the version from 2026-11-18 19:41 (current). Recorded here, not run.
# T759 numerical evidence
Quantity: E_n=det[mu_(i+j+1)]_(i,j=0)^n, size n+1, with ordinary moments
mu_k=(1/4) integral_0^infinity exp(-t/2) Psi(t)t^k dt and Psi=-g. Every
n=0,...,20 is stored; the mathematical domain is all integers n>=0.
Neither moments nor determinants are divided by factorials or replaced by
absolute values. E_0=mu_1; input moments are recomputed through mu_41.
Inspected primary sources: Suzuki arXiv:2206.03682v4 (1.13),(1.15),
Theorem 1.1(3), Theorem 1.8, (8.1),(8.3)-(8.6); DLMF 25.5.13-14; and
FLINT acb_poly Taylor-series documentation, including deflate=True.
The moment integral converges absolutely unconditionally. The source's RH
criterion requires BOTH D_n>=0 and E_n>=0 for EVERY integer n>0. A positive
finite prefix does not establish that criterion or a positive-measure theorem.
Instruction manifest, archived and hash-verified:
/home/ubuntu/numberdb-campaign-20261116-sol/repo/agents/runs/20261118T193048Z-build-skills/manifest.json.
Both valid local/live bundles match SHA-256
b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2;
core SHA-256 b79b7fbc9fa3ce63f2d642cccdac12b0d214757ab01db66c03a2d84264a58e5f.
The same manifest is attached as skill-manifest.json. The builder issues no
independent scientific or argument verdict. argument-selection.md covers
all free parameters and the practical, noncanonical finite bound.
## Production method and measured precision
For s=1+u, the helper encloses Z(u)=zeta(1+u)-1/u with the native deflated
zeta series, then forms
xi(1+u)=(1+u)(1+u Z(u)) Gamma((1+u)/2) pi^(-(1+u)/2)/2.
This removes the pole algebraically. With h(u)=log xi(1+u), the analytic
branch at xi(1)=1/2 gives sigma_(k+1)=(-1)^k(k+1)[u^(k+1)]h(u).
Suzuki (1.15) yields
mu_n=n! sum_(k=0)^n (n-k+1)2^(n-k)sigma_(k+1).
All finite factors are exact integers. The coefficient truncation through
degree 42 suffices for every requested moment; it omits no contributing
Taylor term. Every arithmetic operation carries native Arb/FLINT bounds.
Reality follows from conjugation symmetry of the defined coefficients;
finite real and imaginary components and imaginary-zero containment are
also checked in code. The logarithmic derivative is independent of the
choice of local logarithm branch.
The generator uses fresh moment enclosures, then native unscaled acb matrix
determinants. It starts at 1024 bits and doubles if retained relative bits
fall below bits(digits,losing=96). This selection needed 1024 bits, with
at least 997 moment bits and 830 determinant bits, before serialization to
100 significant decimal digits. Every returned value is a native RealBall;
no Python float, exact-looking decimal or zero-width approximation is used.
production.json records balls, serialized values and per-order accuracy.
The largest stored string is 105 characters; the 21-row Numbers block is
2975 JSON bytes. E_20 is approximately 2.152758565e430, stored in scientific
notation. The last decimal place carries one unit of interval uncertainty.
Existing agents/table-build/dry_run.py and check.py were copied unchanged
and run through agents/sage.sh. Imports, exactness/native enclosure types,
suspicious decimals, prose and sizes pass. Their native working-ball block
is 6.9 KB; those extra working digits are not the stored precision claim.
The displayed standalone Sage program was executed at n=20 and also retains
830 relative bits. Software: Sage 10.9, Python 3.12.3, numberdb 0.1.10,
FLINT 3.4.0; Arb is integrated and not separately reported by pkg-config.
versions.json and the executable versions_probe.py preserve this evidence.
## Separate analytic representation and determinant implementation
independent_check.py imports neither production generator nor moment helper.
Its theta/Li components were inspected and reused from family work, then all
quadratures and transforms were executed anew. No sibling table's numerical
values serve as input or independent evidence. The methods share completed-xi
theory, Suzuki's relations and Arb/FLINT arithmetic; library defects and the
shared analytic foundations are remaining limitations of this builder check.
DLMF gives xi(s)=1/2+s(s-1)/2 integral_1^infinity
(x^(s/2)+x^((1-s)/2)) omega(x) dx/x,
where omega(x)=sum_(m>=1)exp(-pi m^2 x). Hence
A_k=integral_1^infinity (sqrt(x)+(-1)^k)(log(x)/2)^k/k! omega(x) dx/x,
and xi(1+u)=1/2+(u+u^2)sum A_k u^k/2.
Differentiation is justified locally by polynomial factors times the
exponentially decaying theta sum. For x>=1 and every k>=0,
(log(x)/2)^k/k! <= sqrt(x) and |sqrt(x)+(-1)^k|<=2sqrt(x), so the absolute
coefficient integrand is at most 2 omega(x), uniformly in k.
The checker evaluates all k=0,...,41 with 1152-bit rigorous quadrature,
absolute/relative goals 2^-1000, on 1<=x<=L=256, using m=1,...,16.
The infinite integration tail satisfies
error_x <= 2 exp(-pi L)/(pi(1-exp(-3 pi L))) < 3.34e-350,
using m^2>=1+3(m-1). For omitted summands m>=a=17,
error_m <= 2 exp(-pi a^2)/(pi a^2(1-exp(-pi(2a+1)))) < 1.10e-397,
using m^2>=a^2+(2a+1)(m-a) and replacing 1/m^2 by 1/a^2.
These proven bounds are added as radii to EVERY coefficient. Their overlap
outside the finite rectangle only overcounts error. Positive-real square
root/log branches are used; analytic quadrature callbacks reject boxes
crossing their cut. Finite quadrature components and reality are checked.
The checker substitutes u=w/(1-w) using exact binomial coefficients and
obtains Li coefficients by an independent logarithmic-derivative recurrence,
without the production logarithmic-series routine. Suzuki (8.3) gives
mu_n/n!=sum_(j=1)^(n+1)(-1)^(j+1)lambda_j
sum_(k=1)^(n-j+2)k 2^(k-1)binomial(n-k+2,j).
All 42 separately computed moments agree with production enclosures.
The matrix entry mu_(i+j+1) is scaled by 1/(i!j!); a separately implemented
Bareiss recurrence is used and the exact factor product_(i=0)^n(i!)^2
restores the raw determinant. Every interval pivot must exclude zero.
Theta quadrature ran once, about 22 seconds, saving theta-checkpoint.json.
Preupload and stored-value checks resume the preserved outward-rounded
coefficients. The initial in-memory determinants retained at least 841
relative bits; checkpoint readback conservatively retains at least 963
moment bits and 839 determinant bits. Every independent ball lies strictly
inside the exact rational endpoints of its STORED NumberDB decimal interval,
with radius less than 1/1000 of its width. Decimal parsing uses exact
fractions and last-place uncertainty, never floating-point parsing.
check-results.json reports checked_entries=total_entries=21, failures=[],
both control booleans true, and all actual per-entry/moment comparisons.
## Checker controls and identities
Executed nonzero known-answer controls: integral exp(-pi x) on [1,L] against
its antiderivative; mu_0=1+(EulerGamma-log(4pi))/2; E_0=mu_1; and
E_1=mu_1 mu_3-mu_2^2. Exact factorial moments m_k=k! give
D_n=product_(j=0)^n(j!)^2 and E_n=product_(j=0)^n j!(j+1)!, checked for
every n=0,...,8. A row-swapped exact matrix gives determinant -6.
The signed-measure moments m_k=1-2^k/16 are 15/16,7/8,3/4,1/2 at
k=0,...,3, all positive, but E_1=-1/8. This rejects any presumption that
positive entries force a positive determinant. Deliberately wrong sign,
factor 4, size, unshifted offset, factorial-normalized moments and omitted
rescaling are rejected, together with NaN and a radius-1e100 ball.
The displayed shifted Vandermonde integral is an algebraic deduction, not
a claimed source transcription: expand the two Vandermonde determinants,
integrate termwise with w_1(t)=t exp(-t/2)Psi(t)/4, and sum over
permutations to obtain (n+1)! det[mu_(i+j+1)]. Every monomial is absolutely
integrable by existence of all absolute moments. This identity holds for a
signed weight and supplies no positivity without a separate sign argument.
## Bounded proof-facing experiment
basis_experiment.py changes the SHIFTED moment matrix by the exact Li
polynomial basis from Suzuki Section 8. Its lower-triangular coefficients
are B_(j,k)=(-1)^k/k! sum_(a=0)^2 binomial(2,a)
binomial(j+3-a,j-k-a), k<=j. These are coefficients of
(1+w)^2/(1-w)^4 exp(-t w/(1-w)). The exact diagonal is (-1)^j/j!,
so det(B)^2=product(j!)^-2. Development orders 0..10 and withheld 11..20
all satisfy the congruence determinant relation with over 400 relative
bits (minimum 877 for transformed determinants). Exact basis matrices and
all results are preserved in basis-results.json. The signed control has
transformed matrix [[7/8,9/2],[9/2,23]], positive diagonal but determinant
-1/8. This refutes diagonal positivity as a mechanism. An invertible basis
change merely preserves inertia and rescales the original hard inequality.
No new positivity theorem or RH conclusion is claimed.
## Reproduction and final readback
Run every Sage command with agents/sage.sh and mount adjacent dependencies.
Initial computation: offline.py generate.py moment_series.py; initial check:
independent_check.py production.json; later preupload_check.py and readback.py
resume theta-checkpoint.json. The latter pipes the NumberDB key on stdin,
fetches the authenticated table, runs verify(sample=None), compares the
stored document and executes the displayed Programs example. Logs retain
all commands' output and checkpoints. Credentials appear in no artifact.
Authenticated package verification matches 21/21 stored rows, none missing,
differing or extra. The stored-value checker, authenticated audit, rendered
page, reference/companion links and visible Similar tables records pass.
Screenshots include the largest entries; no math error or horizontal
overflow is reported. Every attachment will be read back byte-for-byte;
history must preserve the initial #344 contributes-to #274 message, engine,
model and this run. Those final results are in final-validation.json.
The draft remains private in_progress, unoffered. Fresh independent
scientific and argument-selection sessions are controller responsibilities.