mathematical-checks.md

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

5084 bytes, as of the version from 2026-11-20 00:19 (current). Recorded here, not run.

# Mathematics and precision for T768

The exact definition and normalization were inspected in Rogers 0704.2438v4 Eq. (2.6), Prop. 2.2 Eq. (2.9), Cor. 2.6; Samart 1205.4803v2 Prop. 1.3(iii) and Theorem 1.4 Eqs. (1.9)–(1.12). DLMF 16.12.2 and 15.8.4 supply Clausen and the Gauss connection formula. The archived skill manifest is ../20261120T000421Z-build-skills/manifest.json (matching complete bundle b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2). All assertions below are builder derivations/checks, not an independent-critic verdict.

Divide the polynomial by kxyz. The four exponent vectors are (3,-1,-1), (-1,3,-1), (-1,-1,3), (-1,-1,-1). A constant term in their m-th power requires all four multiplicities equal, hence m=4n and the coefficient (4n)!/(n!)^4. The log expansion has negative sign at every surviving power because m is even. Thus B(k)=log k - sum (4n)!/(4n(n!)^4 k^(4n)) for k>4. Samart explicitly includes k=4. No analytic continuation is used to define an uncomputed interior ordinary measure. Direct Laurent convolution in definition_checks.py verifies m=1..24, including vanishing other powers and first coefficient 24. The first correction to B is 6/k^4.

Write z=256/k^4 and c_n=(1/4)_n(1/2)_n(3/4)_n/(n!)^3. For 0<z<1, I(z)=sum c_n z^n/n, B=(log(k^4)-I)/4. The exact ratio c_n/c_(n-1)=(4n-3)(2n-1)(4n-1)/(32 n^3)<1. After n terms, omitted I is <=c_n z^n*z/((n+1)(1-z)). The entire summation is rational, and Arb encloses the log and this explicit tail. The algebraic root cases compute directly from their exact fourth powers; log(k)=log(k^4)/4 and z=256/k^4 hold exactly, so no rounded algebraic root enters the computation and no root error is omitted. The isolating intervals satisfy 5^4<648<6^4 and 6^4<2304<7^4; the other roots are exactly 12 and 28.

At z=1 split I=integral_0^1 (F(t)-1)/t dt at t=1/2. Clausen says F(t)=g(t)^2, g(t)=2F1(1/8,3/8;1;t). For s=1-t in [0,1/2], the Gauss connection formula gives

g(1-s)=A U(s)+B sqrt(s) V(s),
A=Gamma(1/2)/(Gamma(7/8)Gamma(5/8)),
B=Gamma(-1/2)/(Gamma(1/8)Gamma(3/8)),
U=2F1(1/8,3/8;1/2;s), V=2F1(7/8,5/8;3/2;s).

All U,V coefficient ratios are positive and <1. Hence U,V<=2, their N-term truncations obey the same bounds, and each remainder is <=2s^N. With C=|A|+|B|<4 (checked by Arb), |g-g_N|<=2Cs^N and |g|,|g_N|<=2C, so |g^2-g_N^2|<=8C^2 s^N. Dividing by 1-s costs at most 2. The geometric approximation W_N=sum_(j=0)^(N-1)s^j has remainder <=2s^N, and |g_N^2-1|<=4C^2+1. Therefore the whole high-half integrand error is <=(24C^2+2)s^N<400s^N. The integrated error is <=400*2^(-N-1)/(N+1). Polynomial products U_N^2 W_N, s V_N^2 W_N and U_N V_N W_N are rational and integrated exactly term by term, including the sqrt(s) weight; the final sqrt(1/2) and gamma constants use Arb balls. The lower half uses the geometric rational series above. N=426 at 100 digits, and the ball computation uses 433 bits. The resulting radii are recorded in entries.json; all are <1e-120 before 100-digit serialization. Exact zero is not claimed: B(0)=7 zeta(3)/(2 pi^2)>0.

Independent numerical checks:

* Every one of the 15 k>4 values: mpmath integration of the Gauss square at 135 and 175 decimal digits, without importing the generating recurrence. Finite values, two-precision discrepancies and containment in stored one-ulp intervals are retained. These are numerical checks, not certification of quadrature.
* B(0): independent mpmath zeta versus Arb zeta. The torus map x->x^4, y->y^4, z->z^4 preserves Haar measure, giving the known Smyth linear-polynomial evaluation; T283 is related context, not the source of new digits.
* k=4: Rogers Cor. 2.6 gives B(4)=40 sqrt(2) L(f,3)/pi^3, f=eta(tau)^2 eta(2tau) eta(4tau) eta(8tau)^2. PARI eta-quotient L evaluation at 140/180 digits is independent of the continuation algorithm.
* The four CM identities from Samart, after division by 4, are checked with independent PARI primitive uncompleted L-functions. B(648^(1/4))=40 L(h,3)/pi^3+5 L(chi_-4,2)/(4pi), h=eta(4tau)^6. B(2304^(1/4))=20 sqrt(6)L(g24^(2),3)/pi^3+5 sqrt(3)L(chi_-3,2)/(4pi). B(12)=20 sqrt(10)L(g40,3)/pi^3+8 sqrt(2)L(chi_-8,2)/(5pi). B(28)=200 sqrt(2)L(f,3)/(3pi^3)+5 sqrt(3)L(chi_-3,2)/(2pi). Newforms are identified by the printed Fourier coefficients, especially a_2=-2 for g24^(2). No derivative-at-zero conversion or inferred root-number sign is used.
* Nonzero polynomial quadrature, 2F1(1,1;2;1/3)=-3log(2/3), Catalan's constant and zeta(2)=pi^2/6 control the checker. Adding 1, using the wrong level-24 form, and multiplying B(4) by 4 are rejected. Separate Arb 5F4 checks cover all 15 exterior entries. Reproducibility is checked separately by verify(sample=None).

At k>=0 the homogenization x^4+y^4+z^4+w^4+kxyzw is the Dwork quartic pencil up to the coefficient sign (changed by a fourth-root torus coordinate change). The generic smooth quartic gives the K3 context; the k=4 contact fibre is exceptional. The period relation k B'(k)=3F2(1/4,1/2,3/4;1,1;256/k^4) follows by termwise differentiation only on k>4. A period ratio such as T731 is a different quantity.