back to table · edit · history · where entries came from · files · download
3001 bytes, as of the version from 2026-11-20 07:17 (current). Recorded here, not run.
# Coarse independent canonical-height certificate
This proof and checks.py concern only a coarse certificate of point
independence and a control on the defining limit. They do not certify the
100-digit local-height/PARI output. Arithmetic is exact until logarithms,
which use Sage's directed MPFI intervals at 192 bits.
For an integral Weierstrass model, set b2=a1^2+4a2, b4=2a4+a1a3,
b6=a3^2+4a6, b8=a1^2a6+4a2a6-a1a3a4+a2a3^2-a4^2. The duplication map is
F(X,Z)=X^4-b4 X^2 Z^2-2 b6 X Z^3-b8 Z^4,
G(X,Z)=4 X^3 Z+b2 X^2 Z^2+2 b4 X Z^3+b6 Z^4.
Thus x(2P)=F(A,B)/G(A,B) for x(P)=A/B. This identity is independently
checked against Fraction addition and Sage doubling; the formula also follows
by substituting the tangent slope into the Weierstrass group law.
Let M be the integer 8-by-8 matrix taking the coefficients of homogeneous
cubics U,V to the degree-seven coefficients of UF+VG. checks.py explicitly
builds M from F,G and verifies both extreme columns of M*adj(M)=det(M)*I.
Write R=det(M), which is nonzero for every selected nonsingular curve. For
k=0,7 the corresponding adjugate column gives a homogeneous Bezout identity
U_k F+V_k G=R Z^7 or R X^7, respectively. Let L be the larger sum of absolute
coefficients in these two columns, and U the larger coefficient 1-norm of F,G.
For coprime integers A,B and m=max(|A|,|B|), put d=gcd(F(A,B),G(A,B)). The
two identities imply d divides both R A^7 and R B^7, hence d divides R.
Choosing the identity for the coordinate of size m yields
|R| m^7 <= L m^3 max(|F(A,B)|,|G(A,B)|). Consequently the reduced image has
height bounded below by log(m^4/L), since d<=|R|. Its height is bounded above
by log(U m^4). Therefore, with h(x)=log max(|A|,|B|),
|h(x(2P))-4 h(x(P))| <= C, where C=log(max(L,U)).
If a doubling reaches O, the point is torsion and its canonical height is
exactly zero. Otherwise telescoping the convergent duplication sequence gives
|H(P)-4^(-n) h(x(2^n P))| <= C/(3*4^n).
checks.py computes the rational iterates, the logarithm interval and this tail
bound for all listed points and their required sums. It forms diagonal
intervals for H(P_i) and off-diagonal intervals from
(H(P_i+P_j)-H(P_i)-H(P_j))/2. It increases n only if necessary to make every
leading principal-minor interval strictly positive, using n=6,7,8 at most.
Every actual n, R, L, U and minor enclosure is saved per curve. Positive
leading principal minors prove positive definiteness, which proves that the
specified points are Z-independent modulo torsion. This is a lower bound on
rank only. It neither proves the reported rank is an upper bound nor proves
saturation or full-lattice index. It uses no regulator, L-value or BSD identity.
The certificate's inequalities are independently auditable integer identities;
their numerical consequences are computed with outward intervals. The fine
values remain labelled heuristic (agreement-checked), because the separate
PARI and Sage local-height series computations track no complete rounding bound.