back to table · edit · history · where entries came from · files · download
5977 bytes, as of the version from 2026-11-18 08:37 (current). Recorded here, not run.
# Computation and check evidence for T748
Skill: ../20261118T072550Z-build-skills/manifest.json,
bundle SHA-256 d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.
Both archived bundles and every resource hash were verified equal before work.
Environment: versions.json. All Sage computations used agents/sage.sh.
## Exterior enclosure (k>8)
Let x=64/k^2 and a_n=(binomial(2n,n)/4^n)^3. Independence of the
three angles and vanishing odd cosine moments give
M(k)=log(k)-sum_{n>=1} a_n*x^n/(2n).
The recurrence a_n/a_(n-1)=((2n-1)/(2n))^3 is evaluated by Arb balls
with exact rational inputs. For a_n<=1, the omitted positive tail after N
is at most x^(N+1)/(2(N+1)(1-x)). The returned symmetric ball adds
this full bound to the finite-sum and logarithm enclosure. This is conservative;
the true one-sided interval would be narrower. No floating point rational input
or unbounded/NaN interval is used.
## Endpoint enclosure (k=8)
Write alpha=(1-sqrt(1-64/k^2))/2 and F(u)=2K(u)/pi, with K using the
parameter convention. The published elliptic Green formula and the logarithmic
moment expansion imply
M(k)=log(k)-integral_0^alpha (F(u)^2-1)(1-2u)/(2u(1-u)) du.
Indeed x=4u(1-u) gives dx/(2x)=(1-2u)du/(2u(1-u)). This also proves
the finite limiting formula at k=8 by convergence/continuity.
Let b_j=(binomial(2j,j)/4^j)^2, d_n=sum_{j=0}^n b_j*b_(n-j).
These are exact dyadic rationals, d_0=1 and 0<=d_n<=n+1. Multiplication
by (1-2u)/(1-u) gives coefficient
e_n=d_n-sum_{j=1}^{n-1}d_j, n>=1,
so the integrated sum is (1/2)sum_{n>=1} e_n*alpha^n/n. All coefficients
are computed with Python integer binomial convolution and exact QQ division.
For n>=1, |e_n| <= (n+1)^2 and (n+1)^2/(2n)<=n+1. Hence the omitted
absolute tail after N is bounded by
alpha^(N+1)*((N+2)/(1-alpha)+alpha/(1-alpha)^2).
At alpha=1/2 it is (N+3)*2^(-N), used directly in generate.py.
Absolute convergence at alpha=1/2 justifies termwise integration and products.
The removable integrand limit at zero is 1/4.
## Precision and deterministic checks
The working field is bits(100, losing=96), with N=p+32 in the endpoint
branch and exterior tail target 2^(-p+24). compute.log and computed.json retain
the native RealBall values, serialized 100-significant-digit decimals and
measured accuracy. The weakest ball has 406 relative accuracy bits (over 122
decimal digits). The table's bare decimal represents an interval of one final
printed unit on either side; the interval must contain the computed ball.
The existing dry_run.py and check.py report 25 finite enclosing values,
no float/exactness/import/prose warnings, longest 145 characters, block 4.4 KB.
## Independent implementations used by the builder
check_elliptic.py uses mpmath elliptic K and tanh-sinh integration of the
transformed elliptic integral at 140 and 185 decimal digits. This does not use
the generator's Taylor convolution or exterior moment recurrence. It checks
every row against its stored decimal interval, not merely agreement between
two generator precisions. Its quadrature is heuristic evidence, not the basis
for the generator's proven label. It shares the published elliptic identity
with the endpoint derivation and is not a separate critic session.
A second method integrates out the third original angle analytically, then
uses SciPy adaptive two-dimensional quadrature of
log((k+sqrt(k^2-64*cos(s)^2*cos(t)^2))/2).
Symmetry reduces the region to [0,pi/2]^2 with factor 4/pi^2. Seven values
k=8,9,10,12,16,24,32 cover the endpoint, nearest exterior case and range.
This low-precision corroboration starts with the original Laurent kernel and
is independent of the elliptic transformation. Error estimates are heuristic.
The nonzero zero-kernel control is log(9). Wrong factor 8, wrong z power and
modulus/parameter swap are deliberately rejected. Detailed actual outcomes
are in check-results.json and elliptic-check.log, generated by execution.
The paper's equation (31) intermediate z^l is a misprint: even moments force
z^(2l), in agreement with the final hypergeometric expression. Source equation
(32) also differs in its final printed endpoint digit; record the measured
discrepancy and check independently, without using those printed digits to
certify the 100-digit result. No modular L-series is evaluated here.
## Actual readback outcomes
verify.log reports 25/25 matching, zero differing/missing/extra entries. Its
separate containment test confirms every stored decimal interval contains the
computed Arb ball; the largest displacement including enclosure uncertainty is
0.493132 of the final printed unit. stored-check-results.json has 25 checked
entries, 25 total, no failures, both known-answer and negative controls passed.
The two precision quadratures agree far beyond 100 digits (the exported low
precision check value retains 130 significant digits). The sampled original
kernel integrals differ by at most 8.89e-16 from stored values, with heuristic
quadrature estimates no larger than 2.36e-12.
identity-results.json records exact direct step counts 0,8,0,216 for lengths
1,2,3,4. Finite differences at k=9,10,12,16,24,32 verify k*M'(k)=W(k)
with step 1e-35 and discrepancies below 1.90e-72. This is corroboration of the
differentiated integral, not a proof for all arguments. Corrupted stored values,
NaN and a vacuous interval were deliberately rejected by the checker controls.
The printed endpoint discrepancy is 6.580009795...e-50, established independently
by the enclosing endpoint series and elliptic quadrature. No disagreement among
the checked implementations remains.
Authenticated audit.json is clean. Browser inspection found all four intended
Similar tables relations, 56 MathJax nodes, zero math errors and no page errors.
A two-line derivative formula fixes a genuine horizontal-layout issue; no
mathematical selection or value changed. Screenshots are retained locally.
The independent critic's numerical and argument-selection reviews are pending.