Your free trial has ended.
Try the live demo or log in .
By: ValentinHirschi <valentin.hirschi@gmail.com>
Add the native rank-four bubble family TensorIntegral::B1111 NativeTensorEvaluator gains B1111 = [B0, B1, B11, B111, B1111] (arity 4, 5 coefficients, 15 outputs) for reductions with numerator rank n+2, on every NativeFloat type, with a binary64 specialization. It is OneLOop's bub1111 with the corrected mass-swap reconstruction B1111 = B0' + 4B1' + 6B11' + 4B111' + B1111': OneLOop 3.4 and 3.6 negate this sum in all three swap branches, not only at zero momentum. Binary64: at 53 bits rank four needs 9 and 17 series coefficients at the 0.01 and 0.2 radii of the 8- and 16-coefficient tables, so rank four keeps the tables with its own radii 0.0085 and 0.17. Its length is chosen apart only when RANK == 4, the per-rank arrays of the B11/B111 kernels keep their lengths (width parameters of Binary64::bubble), and the rank-four kernel is out of line: the B11 and B111 kernels compile to the same instructions as before, and their outputs are bitwise unchanged on every route. Tests: tests/native_tensor_b1111.rs checks every route (binary64 kernels, generic binary64, DoubleFloat, Float) against the exact b0..b1111 expressions, direct Feynman-parameter integrals per sector, a dyadic p² ladder 2^-4..2^-24, the swap identity in every branch and both mass orders, and B0..B111 bitwise equal to B111. tensor_feynman.txt gains a B1111 column (the 595 existing rows unchanged) and 133 rank-four rows. NATIVE_BACKEND.md and STATUS.md document the OneLOop sign bug, the small-p² loss of the exact b1111 in fixed precision, and the use of the native family for fixed-precision carriers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>