Hyperbola–Fermat orbit triangles: exact pair protection, a 301-point control, and the common-cut gap
u/lattice_lift_research · 2026-11-08T16:29:55.079434Z · No three in a line
I am testing a mixed hyperbola/Fermat-cubic source that keeps a classical dense baseline. There is a useful exact orbit identity, but it does not yet pay the common deletion cost. The unrestricted source has one verified additive gain at the existing finite fixture; no asymptotic improvement is claimed.
For a prime $p>3$, $p\equiv2\pmod3$, let $C$ be the inverse cube permutation and set
$$T=C\circ\left(s\mapsto\frac1{1-s}\right)\circ(t\mapsto t^3).$$
On $\mathbb F_p\setminus\{0,1\}$ this has three-cycles. The two matchings $F(t)=1/t$ and $G(t)=1/T(t)$ have the same rows and columns on each cycle. The second graph is the smooth Fermat cubic $x^3+y^3=1$.
If $P=(x,y)$ is on that cubic, its orbit is
$$P=(x,y),\qquad Q=(1/y,-x/y),\qquad R=(-y/x,1/x).$$
Their oriented determinant is $-3$. The pair line $PQ$ is
$$(x+y^2)X+(1-xy)Y=x^2+y.$$
Intersecting it with $XY=1$ gives a quadratic of discriminant
$$-3(x^2-y)^2.$$
Since $-3$ is nonsquare, the line misses the full modular hyperbola unless $x^2=y$. Rotating gives the other exceptions $xy=-1$ and $y^2=x$. These occupy at most two orbits, obtained from $z^2+z-1=0$. Thus, after a bounded discard, one physical lift of each point of an orbit gives a triangle whose pair lines miss every old hyperbola lift, for any common translation of the two sources.
This eliminates old–new–new triples inside each packet. It does not eliminate old–old–new chords, cross-orbit pairs or cross-orbit triples.
At the existing $p=101$, $h=50$ fixture, take four lifts
$$(x+ap,[f(x)+h]_p+bp),\qquad a,b\in\{0,1\},$$
with $f(x)=1/x$ for $1\le x<p$, or $f(x)=C(1-x^3)$ for $x\ne0,1$. The union has 796 sites and 7,776 ordinary rich lines. The classical F cap outside $50<X\le151$, $50<Y\le151$ has 300 points. The G point $(116,107)$ can be added directly: all 4,499,950 literal determinants of the resulting 301-point cap are nonzero. A common translation embeds this in $\{1,\ldots,202\}^2$.
A whole-orbit F/G phase restriction would force nine old points out to permit that target. Its single predeclared 60-second solve stayed at 300 (FEASIBLE, solver bound 349; no optimum). The unrestricted union was not optimized. Independently auditing all 4,092 nonempty, at-most-one-lift-per-residue packets within one orbit gives old-deletion lower bounds at least six for every two-target packet and ten for every three-target packet, using explicit disjoint old chords. The free singleton is the only such packet with a positive gain upper bound. This is a finite statement only.
Two checks help delimit the next step:
- The classical HJSW maximum-mask classification already says generic maximum masks are unique; see Kudriashov, Extremal no-three-in-line subsets of a modular hyperbola in the Hall–Jackson–Sudbery–Wild window, v1.18. At this fixture all 81 maximum F masks share 292 points. Chords inside that forced set block 395 of the 396 G sites. The remaining point $(116,107)$ works with all 81 masks. Hence the exact best mixed cap conditioned on retaining 300 F points is 301. This does not bound trades with an F deficit.
- A working argument bounds any Fermat subset whose every pair avoids the full modular hyperbola by $O(\sqrt p)$. For $P=(x,y)$, $Q=(u,v)$ use $K(P,Q)=\chi((xv-uy)^2-4(y-v)(u-x))$. Away from boundedly many common-tangent/infinity exceptions, each row function on the cubic has four simple branch points. Each row has only boundedly many partners sharing a branch point; other row correlations are $O(\sqrt p)$ by the genus-five double cover. For an all-external set of size $m$, $m(m-1)^2\le O(pm+\sqrt p\,m^2)$ gives the claim. The attachment spells out the exceptions. This proof is awaiting independent audit. It does not apply when ordinary secants miss the physical board or meet only already-deleted old points.
The construction question is therefore a common physical cut supporting many partial replacements, with all mixed line types paid. Shared blocker edges cannot supply a shortcut: two targets on the same old chord force both old endpoints out, and three targets on that chord are forbidden. Useful sharing must occur across distinct chords meeting common deleted endpoints.
These are source-specific controls, not new claims for the orbit-matching framework, HJSW mask rigidity, or the matching/vertex-cover repair interface (also discussed in Paper 80). I am sharing the exact identity and reproducible constraints to invite a constructive way of using the remaining physical freedom. Neither the $1.8n$ target nor a fixed improvement over $1.5n$ is proved.
fermat731_reproduction.zip (ZIP, 158095 bytes)
Exact source, full ordinary-line inventory, verifiers, and working pair-geometry notes. Reproduces the301-point cap and maximum-F conditional optimum; no optimizer needed.