back to table · edit · history · where entries came from · files · download
9928 bytes, as of the version from 2026-11-18 20:02 (current). Recorded here, not run.
# T758 numerical evidence
## Defined quantity and instructions
D_n=det[mu_(i+j)]_(i,j=0)^n, n>=0, has matrix size n+1. Its ordinary
moments are mu_k=(1/4) integral_0^infinity exp(-t/2) Psi(t) t^k dt, with
Psi=-g in Suzuki's convention. There is no factorial normalization of
moments or determinants, no absolute value and no stored scaling. This
run selects every n=0,...,20 and uses fresh mu_0,...,mu_40. It reads no
rounded companion values as numerical inputs.
Inspected primary passages: Suzuki arXiv:2206.03682v4, (1.13), (1.15),
Theorem 1.1(3), Theorem 1.8, (8.1), (8.3), (8.4)-(8.6); DLMF 25.5.13-14;
and FLINT's acb_poly Taylor-series documentation, including deflate.
The source's determinant criterion requires BOTH unshifted and first-shifted
nonnegative determinants for ALL integers n>0. The finite positive D-prefix
is not an RH proof, a positive-measure theorem or evidence about all orders.
The integral and its absolute convergence do not require RH.
Coordinated instruction manifest:
/home/ubuntu/numberdb-campaign-20261116-sol/repo/agents/runs/20261118T190114Z-build-skills/manifest.json.
All local/live archived resource hashes were checked; both complete bundles
have SHA-256 b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2.
The manifest is attached as skill-manifest.json. argument-selection.md
justifies the only parameter and bounds. No independent verdict is written
by this builder; a separate critic must evaluate these checks and sources.
## Production computation
For s=1+u, Z(u)=zeta(1+u)-1/u is evaluated with the native ball-polynomial
zeta series and deflate=True. Form
xi(1+u)=(1+u)(1+u Z(u)) Gamma((1+u)/2) pi^(-(1+u)/2)/2.
The pole is removed algebraically, not subtracted numerically. Compute its
logarithm as a ball series through degree 41. The logarithm is the analytic
branch at xi(1)=1/2; its derivative does not depend on branch. Put
sigma_(k+1)=(-1)^k(k+1)[u^(k+1)]log xi(1+u). Equation (1.15) then gives
mu_n=n! sum_(k=0)^n (n-k+1)2^(n-k)sigma_(k+1).
All finite factors are exact integers; there is no Python integer division.
The finite series truncations preserve every requested coefficient exactly
in the formal-series sense; no unbounded analytic series tail is discarded.
moment_series.py implements this fresh moment computation. generate.py
forms unscaled native complex-ball matrices and returns their real determinant
balls. The defined coefficients are real by conjugation symmetry; imaginary
balls must be finite and contain zero as an additional implementation check.
Working precision starts at 1024 bits and doubles until every determinant
retains at least bits(digits,losing=96) relative bits. This run needed only
1024 bits and retained at least 833 bits on all 21 determinants. production.json
records all input balls, determinant balls, serialized values and per-order
accuracy. Native returns are RealBall, never floats or disguised exact values.
The largest stored string is 105 characters; the 21-row stored block is
2366 JSON bytes. Decimal serialization uses 100 significant digits and one
unit in the last place of uncertainty. Scientific notation is preserved for
large entries, including D_20 approximately 1.664396899e404.
The existing agents/table-build/dry_run.py and check.py were copied unchanged
and executed under agents/sage.sh. Imports, native types, finiteness, suspicious
exact decimals, prose and size pass. The dry-run native ball block is 7 KB;
its extra working digits are not additional stored precision. The displayed
Sage program was executed at n=20 and retains 833 relative bits.
## Separate analytic and determinant check
independent_check.py imports neither generate.py nor moment_series.py.
It recomputes all moments, rather than using the sibling table or production
inputs. The analytic theta and Li routines were reused from the prior family
build, inspected against the sources, and executed anew. This is independent
builder evidence, not an independent actor's scientific verdict.
DLMF 25.5.13-14 gives
xi(s)=1/2+s(s-1)/2 integral_1^infinity
(x^(s/2)+x^((1-s)/2)) omega(x) dx/x,
with omega(x)=sum_(m>=1)exp(-pi m^2 x). Thus
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.
Analytic differentiation under the integral is justified on a neighborhood
of u=0 by a polynomial-in-x multiple of the exponentially decaying theta
series. For every k>=0 and x>=1, (log(x)/2)^k/k! <=sqrt(x) and
|sqrt(x)+(-1)^k|<=2sqrt(x), so the coefficient integrand is bounded by
2 omega(x). This uniform majorant controls both omitted parts independently
of any observed digit agreement.
The checker evaluates k=0,...,40 in rigorous 1152-bit complex-ball quadrature,
x from 1 through L=256, using m=1,...,16 and absolute/relative goals 2^-1000.
For the infinite integration tail, m^2>=1+3(m-1) gives
error_x <=2 exp(-pi L)/(pi(1-exp(-3 pi L))) <3.34e-350.
For omitted theta summands m>=a=17, the bound
m^2>=a^2+(2a+1)(m-a) gives
error_m <=2 exp(-pi a^2)/(pi a^2(1-exp(-pi(2a+1)))) <1.10e-397.
Both bounds are added as radii to EVERY A_k. Their overlap outside the
finite rectangle overcounts error. The integral is real on the real domain,
with the analytic square-root/log branches agreeing with their positive-real
values. Quadrature finiteness and imaginary-zero containment are checked.
The checker substitutes u=w/(1-w) using exact binomial coefficients, obtains
Li coefficients from a logarithmic-derivative coefficient recurrence without
calling FLINT's logarithmic-series routine, and applies Suzuki (8.3),
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 41 separately computed moments agree with the production balls. It
scales the moment matrix entry by 1/(i!j!), uses a separately implemented
fraction-free Bareiss algorithm, and reverses the scaling by the exact
factor product_(i=0)^n(i!)^2. Production uses the native unscaled acb
determinant instead. Every pivot used must exclude zero.
The quadrature ran ONCE (about 23 seconds) and saved theta-checkpoint.json;
subsequent preupload and stored checks resumed its coefficients. Original
in-memory determinants had at least 845 relative bits. Re-parsing the
outward-rounded checkpoint gives at least 963 moment bits and 843 determinant
bits, the conservative figures used in the stored table. Every independent
ball is finite, nonzero, and strictly inside the exact rational endpoints
of the corresponding stored NumberDB interval, with radius less than 1/1000
of its width. check-results.json records checked_entries=total_entries=21,
failures=[], and all per-entry comparisons. Stored decimals are parsed with
exact fractions and last-place-unit semantics, never Python floats.
Shared foundations: both routes use completed-xi theory, Suzuki's integral/
coefficient identities and Arb/FLINT arithmetic. They differ in analytic
representation (zeta/gamma product versus bounded theta integral), coefficient
transformation (sigma versus Li), matrix scaling and determinant implementation.
These dependencies and possible common library errors remain honest limits.
## Controls, identities and experiment
Executed known-answer controls: the nonzero integral of exp(-pi x) on [1,L]
against its elementary antiderivative; mu_0=1+(EulerGamma-log(4pi))/2;
D_0=mu_0; D_1=mu_0 mu_2-mu_1^2; and exact factorial moments m_k=k!.
For n=0,...,8, the latter give D_n=product_(j=0)^n(j!)^2 and shifted
E_n=product_(j=0)^n j!(j+1)!, checking size, offset and reversed scaling.
A row-swap matrix has determinant -6. The signed moments m_k=1-2^k/8
have m_0=7/8, m_1=3/4, m_2=1/2 but D_1=-1/8; the checker preserves this sign.
Deliberately wrong sign, factor 4, matrix size, first shift, factorial
normalization and missing rescaling are rejected, as are NaN and a vacuous
radius-1e100 ball. Both required control booleans are true in executed JSON.
The table's Vandermonde integral is a checked algebraic deduction: expand
both Vandermonde determinants in permutations, integrate termwise, and use
permutation symmetry to obtain (n+1)! det[mu_(i+j)]. Every monomial is
absolutely integrable because all absolute moments exist. This proves the
identity for a signed weight as well; it supplies no positivity unless the
weight's sign is established separately. The small determinant identities
are also checked numerically with the fresh moments.
basis_experiment.py executes the proposed Li polynomial basis change, with
B_(j,k)=(-1)^k/k! sum_(a=0)^2 binomial(2,a)
binomial(j+3-a,j-k-a), for k<=j, and zero otherwise. It is the coefficient
matrix from (1+w)^2/(1-w)^4 exp(-t w/(1-w)) in Suzuki's Section 8. Its
exact diagonal is (-1)^j/j!, so det(B)^2=product(j!)^-2. All orders 0..10
and withheld 11..20 verify the resulting congruence relation at over 400
relative bits. Exact matrices, group labels and outputs are preserved in
basis-results.json. The signed control transforms to [[7/8,9/2],[9/2,23]],
which has positive diagonals and determinant -1/8. Thus diagonal positivity
in this basis fails as a mechanism, and invertible congruence merely
preserves inertia and rescales the same hard determinant inequality.
No new positivity theorem or RH conclusion was obtained.
## Readback and handoff
Authenticated package verify(sample=None) matches 21/21 rows, with no
missing, different or extra entries. The separate checker consumes the
actual authenticated stored document. The authenticated audit is clean.
The rendered page, references, Similar tables records and companion links
are checked in presentation-results.json and screenshots. Attachments are
read back byte-for-byte and history is inspected for model/run and exact
proposal #344 contributes-to #274 citations. None of this substitutes for
the controller's fresh independent scientific and argument-selection reviews.
The draft remains private, in_progress, and unoffered by this builder.