nx_bigfloat120_erfc.nx
buildroot/runtime/nx_bigfloat120_erfc.nx
about
nx_bigfloat120_erfc.nx -- ME2-ERF-002 FOUNDATION rung: the SOVEREIGN 120-bit erfc
oracle for the 1.75..6 range the erf Taylor series cannot reach (the LOUD-NAN gap the
erf rung-1 contract named). No python, no minimax, no external oracle -- and it is
PROVABLE by algebra alone (erf(x)+erfc(x)=1), so no trusted reference is imported.
Method = the Laplace continued fraction (the classic erfc CF, converges for x>0, fast
for x>=1.5):
erfc(x) = (exp(-x^2)/sqrt(pi)) * 1/(x + (1/2)/(x + (2/2)/(x + (3/2)/(x + ...))))
partial numerators a_k = k/2, partial denominators all = x. Evaluated by backward
recurrence from depth N (t=0; for k=N..1: t = a_k/(x+t); CF = 1/(x+t)).
exp(-x^2) = 1/exp(x^2); exp(x^2) by DIRECT Taylor (all-positive, x^2<=36 -> the
N=160 tail is < 2^-150 relative, no range-reduction needed); sqrt(pi) by Newton.
bf_erf here is the SAME 44-term Taylor the trusted erf gate uses, but returns a
bigfloat (not f64-rounded) so the gate can check erf+erfc==1 at full precision.
All values POSITIVE-magnitude (the bf core convention); erfc and erf are positive on
the tested domain so no sign bookkeeping is needed. Composes only proven core ops.
module: nishi-core.math.bigfloat120_erfc
depends: nishi-core.math.bigfloat120, nishi-core.math.bigfloat120_div, nishi-core.math.bigfloat120_trig
capability: SOVEREIGN_ERFC_ORACLE_120BIT
license_tier: ORIGINAL
dependencies 4 imports · 4 importers
imports: nx_syscalls.nxnx_bigfloat120.nxnx_bigfloat120_div.nxnx_bigfloat120_trig.nx
imported by: _bf_erfc_oracle_gate.nx_f64_erf_full_gate.nx_f64_erfc_gate.nx_f64_igammaq_gate.nx
structs
| none |
consts
| none |
functions
| 29 | func bfc_exp(out: *i64, v: *i64) -> i64 |
| 53 | func bfc_sqrt(out: *i64, a: *i64, seed_raw: i64) -> i64 |
| 70 | func bf_erfc(out: *i64, xb: *i64) -> i64 |
| 110 | func bf_erf(out: *i64, xb: *i64) -> i64 |