research.md

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

6040 bytes, as of the version from 2026-11-18 01:45 (current). Recorded here, not run.

# Regularized Bessel moments: checked research

Primary source inspected: Chuang–Yu, arXiv:2306.15216v2, https://arxiv.org/html/2306.15216v2. Equation (10) gives I0*K0's coefficients ((2s-1)!!)^3/(2^(3s)*s!). Lemma 21's n=4r+4 limit gives the exact power-only counterterms. Reversing its counterterm index gives S_(n,j) in document.json. Proposition 25 at a=n/2=2m gives (-1)^m*2^(n-2j)*pi^(2m). Corollary 28 gives the signed convergent-moment sum on m<=j<=2m-1. These formulas and the parameter-basis slice were checked in the source; no sibling table is numerical evidence. Formal gamma coefficients equal 4^(-ell) times coefficients of (sum a_s w^s)^(2m).

DLMF 10.32.1 and 10.32.9 inspected through web retrieval: I0(t)=pi^(-1)*integral_0^pi exp(t*cos(theta))dtheta and K0(t)=integral_0^infinity exp(-t*cosh(u))du for positive t. The latter gives K0(t)<=E1(t/2)<=log(2/t)+1 for t<=1, and K0(t)<=sqrt(pi/(2t))*exp(-t). Together with I0(t)<=exp(t), these give explicit endpoint majorants in generate.py. They justify the generator's proven enclosures without using a divergent asymptotic series as an equality.

Live discovery: catalogue fetched once, including signed-in draft existence. screen.already_here and already_asked refreshed; only proposal #335 names the proposed family. Nearby definitions inspected: T521 (45 ordinary convergent moments), T647 (odd-power full determinants), T648 (three vacuum determinants), T724/T725 (odd-power untwisted critical minors), T740 (full even-power rapid-decay determinant, currently offered). None tabulates the balanced finite parts. T196/T199 define the Bessel factors; T7 is pi. Numeric lookup of F_(4,1), G_(4,1), and ordinary B_4(0,2) finds only the last in T521. Existing tags integral, period, moments, special values, number theory are present; Bessel moments is absent. No new tag was created.

Provenance checked in live GitHub issues: #335 is this proposal; it explicitly contributes to #263 (arithmetic integral contexts). #323/#334 are family context, not direct origins of this table. The exact initial message from proposal.json was used for private draft T741 and is to be preserved on all amendments. No whole wanted issue is resolved by this table.

Selection: argument-selection.md. Full ten-pair triangle n=4,8,12,16 and n/4<=j<=n/2-1, each with bessel and period. No continuous coupling. The cutoff is computational, not canonical. n=4r+2, odd powers, logarithmic insertions and unequal scales remain outside this table. G beyond the stored basis slice is explicitly defined as a rescaling, not claimed to be a basis coordinate.

Numerical development finding: the first broad quadrature panels returned rigorous but unusably wide enclosures for n>=12. Their bounds were not silently interpreted as 100 digits; run-numerics.log and diagnose.log preserve the failure. Fixed width-four panels resolve the loss. The generator explicitly refuses results with fewer than 105 supported decimal digits (plus its bit guard). The separate finite-part checker uses two cutoff/truncation pairs; its asymptotic remainder is heuristic, independent of the generating sum rule, and labelled accordingly.

Skill manifest: ../20261117T154138Z-build-skills/manifest.json; both bundles verified, matching bundle SHA-256 889383d95f882353fc09d210af003abe6edcf0331f27c377d7acdabafa344208. Larger local artifacts: source.html/source.txt, source-passages.tex, catalogue.json, candidates.json, discovery.json, neighboring T*.json, and original issue JSON. Numerical evidence and file hashes are added after completion. Independent scientific and argument-selection verdicts are reserved for the controller's fresh critique actor.

Known-answer checker controls were derived from DLMF 10.32.10 (the Schläfli representation of K0), not a sibling table: squaring K0(t)=1/2*integral_0^infinity exp(-u-t^2/(4u))*du/u, integrating t^(2r+1), then substituting u=s*x, v=s*(1-x), gives integral_0^infinity t^(2r+1)*K0(t)^2 dt = 2^(2r-1)*(r!)^4/(2r+1)!. Thus the two positive controls are 1/2 and 1/3 for r=0,1. Positive integrands justify Tonelli; the remaining integrals are gamma and beta integrals.

Checkpoint serialization: the client form='ball' serializer rounds its center and radius separately without enlarging the radius for center rounding. The raw log preserves those original strings; values.json conservatively adds one exact decimal ulp of the center and one of the radius before using the enclosure. The table itself uses ordinary 100-digit NumberDB decimals. This artifact-only workaround does not modify the immutable client/toolchain. The final generator accepts these source-fingerprinted checkpoints; --recompute runs the full certified integration. Post-upload verify() uses these retained computation checkpoints, and is explicitly a stored-value/reproducibility check, not another independent calculation.

Independent-check runtime probe: at 560 bits, the returned enclosure for (I0(t)*K0(t))^2*t^2 loses almost all accuracy near t=192 and becomes huge by t=256, while t=384 switches back to a narrow enclosure. This is a special-function enclosure precision regime, not evidence of disagreement in the finite-part formula. probe-check-panels-v2.log preserves the actual point/panel results. A separate 1800-bit probe gives 1422, 1330, 1238, 1053 and 683 relative accuracy bits at t=128,160,192,256,384, respectively. The contour-quadrature attempt then used 1800-bit arithmetic with 2^(-480) tolerance and passed both four-factor conventions, but was too expensive to complete the whole family within the builder limit. Its partial results remain in checks-v2.log and code in checks-arb.py. The final checker uses the same defining finite parts and tail orders, tanh–sinh quadrature at 150/190 decimal digits, and cached 1800-bit Arb midpoint Bessel evaluations. It checks all stored 100-digit intervals with an explicit heuristic agreement guard. This corroboration is labelled heuristic; the generator remains proven and unchanged. No table precision downgrade was made.