method.md

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

5284 bytes, as of the version from 2026-11-18 12:45 (current). Recorded here, not run.

# Exact computation and independent checks for T751

Skill archive: ../20261118T114154Z-build-skills/manifest.json. Both archived
versions are ready and match bundle SHA-256
d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.
The builder does not issue an independent scientific verdict.

## Definition and electrical normalization

Use the vertex/edge sets specified in the table. Their positive Laplacian
symbols are 2(3-cos u-cos v-cos w), 8(1-cos u cos v cos w),
and 4(3-cos u cos v-cos v cos w-cos u cos w), respectively.
For source minus sink the energy is the Fourier integral of
2(1-cos(r dot theta))/Delta(theta). Independent sign symmetries turn the
numerator cosine into a product of coordinate cosines. Thus, for the Green
functions G, B, F in Glasser--Boersma (2000), equations (1), (26), (16),
the resistances are g0-G, (b0-B)/4, (f0-F)/2.

These factors also follow from finite Foster trace: on a connected
edge-transitive periodic quotient with N vertices and degree q,
sum_edge b^T Delta^+ b=N-1, so each edge has resistance 2(N-1)/(qN).
The limit is 2/q, giving 1/3, 1/4, 1/6. The FCC and BCC papers' tabulated
Green differences are twice and four times the physical resistance here.
They are used only after conversion, never as unscaled resistance checks.

## Exact finite linear algebra

On the three-dimensional angular torus let D=c0-sum_{s in S}d exp(i s.theta),
where (c0,d)=(3,1/2), (1,1/8), (3,1/4) for SC, BCC, FCC.
Define H(r)=average exp(i r.theta)/D. The origin constants h=H(0) are
computed from gamma products, independently compared with elliptic formulas.

Two exact identities hold:

1. q H(r)-sum_{s in S}H(r+s)=(q/c0) delta_{r,0}.
2. sum_{s in S}(r_i s_j-r_j s_i)H(r+s)=0 for i<j.

For identity 2 integrate the divergence of
exp(i r.theta)(D_j e_i-D_i e_j)/D. The differentiated denominator cancels
because the vector is tangent to D's level sets; the mixed second derivatives
cancel as well. Excise small balls about D's isolated quadratic zeros before
integration. D_j/D=O(1/|theta|), so their boundary flux tends to zero.
All original H integrals converge in dimension three. This supplies the
boundary justification rather than assuming an arbitrary finite truncation.

In derive.py, reduce by independent coordinate signs and permutations. Include
all valid representatives with maximum coordinate at most 8. Retain an equation
only when every nonzero term belongs to this set. Add the stated source seeds.
Solve all three rational coefficient columns in the basis (h,1/(pi^2 h),1).
Verify A*X=B exactly. Compute a full rational right-kernel basis of A and verify
that every selected coordinate vanishes in it. Thus the true integrals and the
computed solution agree at the selected coordinates, even though the finite
system leaves some remote helpers undetermined. No zero boundary condition
and no numerical fit or PSLQ are used. Nullities are 3 SC, 1 BCC, 2 FCC.

SC seeds are Glasser--Boersma Table 1 at 000,110,111,220,221,222;
FCC seeds are its equations (23),(24) and Table 2 at 000,200,220,222,400.
BCC seeds at 000,200,220,222 are converted from Asad et al., Table 1,
arXiv:1302.4376, p.8. Additional seeds are redundant consistency constraints
where exact row rank makes them redundant; they are not new free inputs.
The SC 111 coefficient is -1/8 (not -1/2).

## Precision and return types

All recurrence coefficients and exact adjacent values use Sage QQ. Arb
evaluates gamma products and all transcendental arithmetic at
numberdb.bits(digits,losing=256) bits. Nonexact output is a finite positive
real ball, with measured radius below 10^(-digits-20); exact rows return
native Sage rationals. The generator asks for 100 significant digits.
The mathematical error claim uses source seeds and exact identities, then
Arb enclosures. Quadrature agreement does not certify its own truncation error.

## Check scope

- dry_run.py and check.py run unchanged; their output is retained.
- Breadth-first search of the declared step sets independently checks every
  selected orbit and graph-distance formula, including both BCC parity classes.
- Exact harmonic identities are checked at all 87 rows, with helper sites
  outside the selected ball, including neighbours of cutoff rows.
- All 29 BCC coefficient triples are separately transcribed from the primary
  paper and compared against the exact solve (allowing the factor four).
- Gamma and elliptic origin formulas are evaluated independently in Arb.
- quad_check.py independently integrates out the third original angle using
  J_n=(b/(a+sqrt(a^2-b^2)))^n/sqrt(a^2-b^2), with its signed b and rho.
  Four angular subsquares, each split into two Duffy triangles, remove corner
  singularities. Jacobians and the factors 1,1/4,1/2 are explicit.
  Double precision checks cover all 87 rows at orders 64 and 96. High precision
  order comparisons cover all 41 rows in shells 1,2,6. Arithmetic-ball radii
  in these quadrature sums do not include Gauss truncation error.
- Controls include all three nonzero adjacent exact values; deliberately wrong
  scaling factors, invalid vertex parities, a decorated step set, and the FCC
  modulus/parameter swap are rejected. Detailed executed results and bounds
  belong in numerical-evidence.json, not inferred from this plan.