A genus 2 curve with 738 rational points

A genus 2 curve with at least 738 rational points and an absolutely simple Jacobian. The record for such curves, set by Elkies and Stoll in 2008, was 642. The curve lies on a K3 surface, and it is found by a computation that starts from one published equation of that surface, builds 2381 models of it and searches the lines on each of them; it is simple to state and takes about an hour on a 16-core computer. This note explains where such curves come from, how the models are built, how the search works, and how the points are found. The same search finds a second new curve, with 654 points. All 738 points can be regenerated from five small values of x by a rule simple enough to run in your browser.

Written by Claude Fable 5.2 under the superversion of Andrew V. Sutherland · updated 18 November 2026

y2 = 743425x6 − 16144220x5 + 60026500x4 + 172162750x3 + 40133975x2 − 46160230x + 6838225
11010²10³10⁴10⁵10⁶10⁷10⁸−3−1−⅓0⅓13−∞+∞x-coordinate (the whole projective line, compressed by arctan)height of x
Every rational point we know. Each dot is one of the 369 x-coordinates, placed by its position on the projective line and by its height (the larger of numerator and denominator). Each stands for two points of the curve. The colours say which of the model's tabulated curves pass through a point, as the sections below explain; they do not say how the point was found. Height measures arithmetic size, not position: x = 1474359/1473364 is close to 1 and has height 1.5 × 106. Point at a dot, or tap it, for details. Select a legend entry to hide or show its points.

A curve of genus at least 2 has only finitely many rational points, and a conjecture of Lang would imply that the number is bounded in terms of the genus alone [2]. Nobody knows how large the bound for genus 2 would have to be, so the largest number of points anyone has found on one curve is a number people keep track of. It moves rarely.

1995 Keller–Kulesz (split Jacobian)5882008 Elkies–Stoll6422026 GPT-6 Astra (split Jacobian)6482026 this note, second curve6542026 this note738
Records for the number of rational points on a curve of genus 2 over Q. The curves of 1995 and of GPT-6 Astra have twelve automorphisms and Jacobians that split. The other three have absolutely simple Jacobians.

There are two kinds of record. Keller and Kulesz found a curve with 588 points in 1995 [9]. It has twelve automorphisms, its 588 points form 49 orbits of twelve, and its Jacobian splits into elliptic curves. In 2008 Stoll, searching in families supplied by Elkies, found a curve C642 with 642 points [19], [21]. It has no symmetry beyond the one every genus 2 curve has, and its Jacobian is absolutely simple. In that sense it is a generic curve.

Beating 642 is one of the open problems that Epoch AI lists in its FrontierMath collection [7]. The problem puts no condition on the Jacobian, and in 2026 it was solved by the model GPT-6 Astra with the curve

y2 = 98285087781x6 − 1685253906933x4 + 7793336764011x2 + 83877673125

which has 648 points. Like the Keller–Kulesz curve it has twelve automorphisms and a split Jacobian (648 = 12 · 54). Our aim was different: to beat 642 with a generic curve. The curve C738 at the top does that. It also has more points than the 648-point curve, which was the largest number published for any curve of genus 2.

The height of a fraction a/c in lowest terms is max(|a|, |c|), and the height of a line az0 + bz1 + cz2 = 0 with coprime integer coefficients is max(|a|, |b|, |c|). We write the line as (a, b, c).

The search in brief

One K3 surface has many models as a double cover of the plane, and each model carries its own two-parameter family of genus 2 curves, one for each line. The models are classes in a lattice. Starting from Elkies' sextic, and from nothing else, we construct sextics, in minimal coordinates, for all the models with one or two nodes that we know of and for 2188 of the 2192 with three: 2381 sextics in all.

On each sextic we list the rational points of the surface of height at most 4000 and count them on every line of height at most 400. This number is the score of the line. For the lines with a score of at least 40 we count the rational points of height up to 105 on the curve the line cuts out, and for the best of those curves we close the set of points under a relation that finds the points of large height. No step takes a curve as input, and nothing is tuned to a particular model.

The whole computation took 62 minutes on a 16-core machine, 7 of them for the construction of the models. Sixteen of the curves it finds have at least 600 points, and three have at least 642: C738, a second new curve with 654 points, and the 2008 record curve itself. The files needed to check each statement and to repeat the computation are listed under Reproducing the results.

What is taken from earlier work and what is new. The framework is that of Elkies and Stoll: line sections of double-plane models of one K3 surface, with points forced by the curves on which the sextic is a square, a search of all lines up to a height bound, and Stoll's observation that further points can be found on cubics through known ones [19]. This note adds three things. The first is a construction that produces sextics for 2381 models from Elkies' one, where the search of 2008 used five. The second is a score that ranks all the lines of a model at a cost of a fraction of a microsecond per line: the number of rational points of small height of the surface on the line. The third is the four-point closure, which finds the points of large height and, unlike the five-point test, enlarges the subgroup generated by the known points.

What is proved, and what is not

statementstatus
at least 738 rational pointsproved; certificate below
Galois group S6, endomorphism ring Z, no torsionproved
rank of the Jacobian at least 21proved
rank exactly 21assumes GRH
exactly 738 rational pointsnot proved; evidence below
each of the 2381 sextics searched is a model of X163proved; certificate below
the census of models is completenot proved; evidence below
the search finds every line section with many pointsnot claimed; its limits are listed below

Curves from lines on a surface

The curve was not found by searching among curves. It was found on a surface.

Take a homogeneous polynomial S of degree 6 in three variables and form the double plane w2 = S(z0, z1, z2). For suitable S this is a K3 surface. Every line λ in the plane gives a curve on it: restrict S to the line and you get y2 = (a sextic in one variable), a curve Cλ of genus 2. So one surface carries a two-parameter family of genus 2 curves, one for each line. This is the idea of Elkies that Stoll used in 2008 [4], [19].

The reason to do this is that rational curves on the surface force rational points on all of these curves at once. Suppose ℓ is a line on which S restricts to a perfect square, S = G2. Above ℓ the surface splits into the two rational curves w = ±G. Any other line λ crosses ℓ in one point, and above that point sit two rational points of Cλ (one, if the point lies on the sextic curve S = 0). This happens when ℓ meets the sextic curve S = 0 with even multiplicity everywhere: when it is tangent to the curve at three points (a tritangent), or passes through one or two of its nodes and is tangent to it wherever else they meet (a node line). In general S is then only a constant times a square on ℓ, and the constant need not be a rational square. On the surface used here it always is, because every curve class on that surface is defined over Q. A model with N such lines puts 2N rational points on Cλ whenever λ crosses the N lines in N different points, none of them on the sextic curve. Only the lines λ through one of finitely many points of the plane fail this (the points where two of the N lines meet and the points they share with the sextic curve), and they get fewer. We call these the forced points, and say that the model has 2N of them.

where S3 < 021 tritangents18 node linesthe line λ = (66, 31, −52)
λ
The model that carries the record curve. The outlined curve is its branch sextic S3 = 0, drawn in the chart z0 = 1; the surface has real points only over the unshaded region. It has three nodes: one is marked with an open circle, and the other two are at infinity, where the vertical lines meet and where the parallel lines of slope ½ meet. The thin lines are the 39 lines on which S3 is a square. The thick line is λ. Each dot where it crosses a thin line is a pair of rational points of C738. Point at a line or a dot for its equation.

The same thing happens in higher degree. If c is a conic on which S is a square, a line λ meets it in two points. Usually they are conjugate over a quadratic field, but when they happen to be rational they give, in general, four more rational points on Cλ. Cubics behave in the same way, and so on. A line for which many of these coincidences happen at once has a curve with many points.

The figure at the top shows how far this accounts for C738. Its line crosses the 39 special lines at 39 different x-values. The model has 810 conics, and 62 of them meet the line in rational points, at 95 different x-values. For the next layer we tabulated 21,250 cubics on which the sextic of the model is a square. A line meets a cubic in three points, and 360 of these cubics cross our line in a rational point, at 137 different x-values, of which 74 are on no line or conic. Lines, conics and cubics together explain 206 of the 369. The other 163 are not accounted for by these tables. They may lie on curves of degree four and more or on cubics missing from our table, or on no rational curve of low degree at all.

For this to produce records the surface needs a very large supply of rational curves. Elkies chose the K3 surface X163 described in the next section and wrote down a sextic for it with one node, 43 tritangents and 9 node lines, hence 104 forced points, and 1240 conics [4]. Stoll searched all lines of height at most 500 on five such sextics and found C642 [19].

One surface, a great many models

A K3 surface does not come with one sextic. Each way of mapping it two-to-one onto the plane gives a different sextic and a different family of genus 2 curves, and a given genus 2 curve on the surface is a line section in exactly one of them. So before asking which line, one has to ask on which model.

The surface itself can be written down in one line. It is the surface that Shioda and Inose [16] attach to E × E, where E is the elliptic curve with complex multiplication by the ring of integers of Q(√−163), whose j-invariant is −6403203. With a = 53360 = 640320/12, Inose's elliptic fibration of that surface [8], [15] is

y2 = x3 − 3a2x − (t + 1/t + 2a3 + 2),

with fibres of type II* at t = 0 and t = ∞. Elkies gives this equation together with a generator of its group of sections [5]. Elkies' sextic, from which our computation starts, is a model of this surface over Q, by a map that is explicit in both directions. On Elkies' sextic the zero section of the fibration is the curve that the node is blown up into, each of the two II* fibres is made of halves of eight of the special lines and of one conic, and the other fibres are curves of degree 32. The map is given by forms of degree 32 and 24, and its inverse by forms of degree 34. Both are among the files, with scripts that prove that they are inverse to each other.

Models can be listed without writing down any equations. The divisor classes of X163 form a lattice of rank 20 and discriminant −163,

NS(X163) ≅ U ⊕ E8(−1)2 ⊕ −2−1−1−82,

and every class is defined over Q [4]. The surface has complex multiplication by Q(√−163), and its point counts over finite fields are given by a CM newform of weight 3 and level 163, which we use to check every model we build. A model is a class D with D2 = 2 that is nef and for which no class E has E2 = 0 and E·D = 1 (with such an E there is no map onto the plane [14]), and the lattice D⊥ of rank 19 records what we need to know about it:

Two classes give isomorphic models exactly when an isometry of the lattice takes one to the other. The orbits of all classes with D2 = 2 correspond to the isometry classes in one genus of positive definite lattices of rank 19 and determinant 326 [12], whose mass is about 16,417.77. Not all of them are models. The classes that have an E as above are exactly those for which D⊥ splits off a vector of norm −2 as a direct summand. They correspond to a genus of rank 18 and determinant 163, the elliptic fibrations of the surface with a section, and they carry mass 1.11. The models carry the rest, about 16,416.66. Every lattice has the automorphism −1, so there are more than 32,800 models, and in fact far more, because most of the mass sits on lattices with many roots and large automorphism groups. We sampled the genus with seven independent random walks along Kneser neighbours [22], 9.9 million steps in all. In the long run such a walk meets a lattice with a frequency inversely proportional to the size of its automorphism group, so it meets the lattices with few roots very often: every model in the figure below was visited at least 4 times, and each of the 2192 three-node models at least 10 times. A random walk cannot prove that a list is complete. What it gives is a heuristic estimate: if the visits to a lattice were independent draws, a lattice with a automorphisms would be expected 606/a times and missed with probability about e−606/a. Successive steps of a walk are not independent, but the counts behave much as this predicts. The 1141 three-node lattices whose automorphism group was computed and has the generic order 16 were visited 37.6 times on average, where 37.9 are expected, with variance 42.5 (for independent draws the variance would equal the mean), and of the 193 models with one or two nodes, 185 were found by each of the seven walks separately and the other eight by six of them. On this estimate the models at risk are those with many automorphisms: the chance of missing one is about 1 in 10,000 for a = 64, 1 in 100 for a = 128 and 1 in 10 for a = 256. On this evidence the list should be complete for models with small automorphism groups, and it may lack models with large ones. Nothing below depends on completeness. The deeper strata are known only in part. None of the 53,361 lattices the walks recorded is of the excluded kind.

60708090100110forced points on every line sectionclasses foundone node7C₆₄₂one cusp8two nodes186three nodes2,192C₇₃₈C₆₅₄node and cusp550four nodes13,666 (partial)
Forced points by type of model. Each bar spans the numbers of forced points among the models whose branch sextic has the given singular points. The count on the right is the number of models the census found; for four nodes only those with enough forced points were recorded. No model with a smooth sextic turned up. Dots mark the models of the three curves discussed here.

The seven one-node models have (tritangents, node lines) = (46, 8) once, (45, 8) three times, (44, 9) twice and (43, 9) once. The last is Elkies' published sextic, and C642 is a line section of one of the (45, 8) models.

Degtyarev has classified the smooth models of the K3 surfaces of Picard number 20 and small discriminant [3]. For double planes his result stops at discriminant 116, so it does not decide whether X163 has a smooth sextic model, but for this same surface he finds more than three thousand configurations of lines on smooth quartic models. The abundance of models is not special to double planes.

The more singular points a model has, the fewer forced points. The search of this note uses the models with one, two and three nodes: all seven with one node, all 186 with two, and 2188 of the 2192 with three. The model of C738 is an unremarkable member of the three-node stratum, with 78 forced points (21 tritangents and 18 node lines), in the middle of the range.

Building the models

The census gives a lattice, not a sextic, and to search a model we need its sextic. We get all the sextics from a single one, Elkies' sextic, by changing the map to the plane. The computation finds by itself the models it constructs; it does not use the census.

  1. Start from Elkies' model and its conics. We computed the conics on which Elkies' sextic SE is a square, 1241 in our table, with a rational parametrization of each. If SE = g2 on the conic q = 0 and G is a cubic form that restricts to g, the conic has two lifts to the surface, Γ = {q = 0, w = G} and Γ′ = {q = 0, w = −G}. Both are rational curves of self-intersection −2.
  2. Pick two conics whose lifts meet three times. Two conics meet in four points of the plane, and at each of them the two lifts either agree or differ in the sign of w. If they agree at exactly three, then Γ1·Γ2 = 3 and the class D′ = Γ1 + Γ2 has D′2 = −2 − 2 + 2·3 = 2. It is a new model, in which the two conics have become the two halves of a single tritangent line.
  3. Find the three coordinate functions. They are the functions (P + Qw)/(q1q2), with P a quartic form and Q a linear form, whose numerator vanishes on the other two lifts Γ1′ and Γ2′. That is a linear system with 18 unknowns and rank 15, so it has exactly three independent solutions z0, z1, z2.
  4. Find the new sextic. The Jacobian determinant of (z1/z0, z2/z0) with respect to coordinates of the old plane vanishes on the ramification curve of the new map. The derivatives are taken on the surface, where w is a function of the plane coordinates and ∂w = ∂SE/(2w). Multiplied by w and by z03 it gives a function T that changes sign under the new covering involution, and T2 = c·S′(z0, z1, z2) for a sextic form S′ and a constant c. We find its 28 coefficients by exact linear algebra from a few hundred rational points of the surface and confirm the identity on fresh points.
  5. Put it in minimal coordinates. The height of a line depends on the coordinates, so the sextic is replaced by a minimal and reduced one, computed with Magma's MinRedTernaryForm [6]: a transform of the sextic that is minimal at every prime, with coefficients then made small by a unimodular change of variables. The effect on a search by height can be large. Elkies' sextic is itself not minimal at the prime 2: its minimal form has largest coefficient 35,356 where the published one has 8,697,600, and the 52 lines on which it is a square have height at most 37 there, against 148. Of the 2380 sextics that the construction produces, 397 are not minimal as they come out. A minimal sextic is not unique, however, and MinRedTernaryForm returns one that depends on the coordinates it is given. For example, the sextic S of the one-node model b5 of our list has largest coefficient 6984, and S(z0 − z2, z0 + z2, z1 − z2)/4 is another integral sextic for the same model, with largest coefficient 5134. The change of coordinates has determinant 2, which the division by 4 compensates exactly (26 = 43), so the second sextic is minimal as well. A line has different heights on the two, so which lines have height at most 400 depends on the minimal sextic used (see the limits of the search).

From Elkies' sextic to 2381 models

The five steps are applied to every pair of conics that gives a model with at most three nodes. The pairs are found, and sorted into models, in the lattice of curve classes, which is computed from the sextic.

  1. Lines. An exhaustive search finds the lines of height at most 150 on which Elkies' sextic is a square: 43 tritangents and 9 node lines. Their 104 halves and the curve that the node is blown up into span a lattice of rank 20 and discriminant −163, which is therefore the whole lattice of curve classes. From here on curves are handled through their classes.
  2. Conics. The conics on which the sextic is a square are found modulo 41, lifted to characteristic 0 and verified exactly. The class of a half of a conic is determined by its intersection numbers with the halves of the lines. An enumeration of all classes E with E2 = −2 and degree at most 2 then shows that both tables are complete: the irreducible classes of degree 1 are the halves of the 52 lines, and those of degree 2 are the halves of the 1241 conics.
  3. Pairs. For each pair of halves of conics with Γa·Γb = 3, the vectors of norm −2 orthogonal to D′ = Γa + Γb are counted. They are among the classes just enumerated: such a vector has degree at most 3 on Elkies' model by the Hodge index theorem, and degree 3 would produce a class of square 0 and degree 1, which a model does not have. The classes D′ with one, two or three such vectors up to sign, pairwise orthogonal, are kept: 380, 3378 and 13,966 classes on Elkies' model.
  4. Models. For each D′ kept, the contracted curves and the halves of the special lines of the new model are found by a second enumeration, and the table of their intersection numbers is brought to a canonical form under reordering. In every case these classes generate the whole lattice. So two of the classes D′ have the same canonical form exactly when an isometry of the lattice takes one to the other, that is, exactly when they give isomorphic models.
  5. The other one-node models. On Elkies' model the pairs give seven models with one node, 180 with two and 1531 with three. One of the seven is Elkies' model itself. For each of the other six a sextic is constructed and its node is moved to (0 : 0 : 1); then steps 2 to 4 are repeated on it. No further model with one node appears: the seven are closed under the construction. Together they give 186 models with two nodes and 2188 with three.
  6. Sextics. For each model one pair of conics is chosen, on Elkies' model whenever the model occurs there, and the sextic is constructed. Several things are checked. The lines and conics of the old model whose halves have degree 1 or 0 in the new one must be mapped to lines on which the new sextic is a square and to singular points of it; the new sextic must have the number of nodes and of lines that the lattice gives; and its point counts modulo the primes between 47 and 100 must agree with the CM newform.

The models constructed are therefore the seven models with one node that are reached from Elkies' model by pairs of conics, and all the models with two or three nodes that are sums of two conics on one of these seven: 7 + 186 + 2188 = 2381 models. The census has the same seven models with one node and the same 186 with two, and it has 2192 with three: the 2188 constructed here and four more (the script census.m compares the lattices). These four are not sums of two conics on a one-node model. (A sum of two conics on a one-node model D0 always has D′·D0 = 4, and for each of the four the lattice shows that no one-node model of the census has D′·D0 ≤ 4.) There is no deeper obstruction. They can be written down by another route, for instance starting from a model with two or three nodes, or as a sum of three curves. We have not done that, so these four are the only three-node models of the census that are not searched. The whole construction takes 7 minutes.

The list of models, and a certificate for it

The search uses the 2381 sextics: 7 with one node, 186 with two and 2188 with three. They are in the file models.txt, each with its singular points and the lines on which it is a square. The Magma script models.m verifies every one of them from that file alone, independently of the construction:

The script then computes the lattice D⊥ of each sextic and shows that no two of them are isometric: the 2381 sextics are 2381 different models. This confirms, by another method, the sorting by canonical forms used in the construction.

Worked example: the model of C738

Take the two conics

q1 = 3400x02 + (−3150x1 + 54x2)x0 − 5400x12 + 551x1x2 − 9x22
q2 = 5760x02 + (−5896x1 − 936x2)x0 − 209x12 − 458x1x2 − 9x22

of Elkies' model, the pair that the construction chooses for this model. They meet in one rational point and in three points conjugate over a cubic field. Their lifts agree at the three conjugate points and differ at the rational one, so Γ1·Γ2 = 3. The lattice says that the class Γ1 + Γ2 is a three-node model with 21 tritangents and 18 node lines.

q1 = 0q2 = 0
Elkies' model (chart x1 = 1)
11z0 + z1 + 2z2 = 0
The new model (chart z0 = 1)
Two conics become one line. Left: Elkies' sextic with the two conics. Each conic touches the sextic wherever it meets it, which is why the sextic is a square on it. Right: the three-node model built from them. The chosen lifts of the two conics to the surface are mapped onto the two halves of the line 11z0 + z1 + 2z2 = 0, one of the 21 tritangents of the new model. The map involves w: it is a map of the surface, not of the plane.

Carrying out the five steps gives the model w′2 = S3(z0, z1, z2) with

S3 = 25z06 + (−1930z1 + 3240z2)z05 + (969z12 − 1734z1z2 + 1996z22)z04 + (5752z13 − 6326z12z2 − 996z1z22 − 864z23)z03 + (2748z14 − 1328z13z2 − 3588z12z22 + 960z1z23 − 176z24)z02 + (−144z15 + 216z14z2 − 548z13z22 + 1376z12z23 + 16z1z24)z0 + 36z16 + 12z15z2 − 311z14z22 − 52z13z23 + 676z12z24

This sextic is irreducible and has exactly three singular points, the nodes (0 : 0 : 1), (0 : 2 : 1) and (2 : −2 : −1). It is a square on exactly 21 + 18 lines, as the lattice predicts (15 of the node lines pass through one node, and the other three are the sides of the triangle of nodes), and its point counts match the CM newform. It is already minimal and reduced. Its largest coefficient is 6326, close to the median of the largest coefficient over the 2381 sextics, which is 6073.

A certificate for the model. These statements, and the identification of the surface, can be checked without the construction. The Magma script below (k3model.m) proves them from S3 and the list of the 39 lines alone, in a few seconds. The three nodes are the only singular points. The 78 halves of the 39 lines and the three contracted curves have an intersection matrix of rank 20 and discriminant −163. So they generate the whole lattice of curve classes, and every class is defined over Q. Over an algebraic closure of Q the surface is therefore X163: a complex K3 surface with 20 independent curve classes is determined by its oriented transcendental lattice [16], and there is only one positive definite even binary form of determinant 163. The same computation identifies the surface over Q as well, which the construction also gives directly, as an explicit map over Q from Elkies' model. The reason is that an isomorphism between two such surfaces, on both of which every class is defined over Q, differs from each of its Galois conjugates by an automorphism that fixes every curve class, and on this surface the only such automorphism is the identity: on the transcendental lattice it could only act by ±1, and −1 does not act trivially on the discriminant group Z/163, as it would have to. So the isomorphism is defined over Q. Enumeration in the lattice then shows that S3 is a square on exactly 39 lines and on exactly 810 conics. Finally, the number of points over Fp agrees with the CM newform at the 32 primes 41 ≤ p < 200 where the model has good reduction.

The Magma script k3model.m
// k3model.m -- a Magma certificate for the double plane w^2 = S_3(z0,z1,z2) that carries the curve C_738.
//   magma -b k3model.m                                  (a few seconds)
//   magma -b conics:=data/S3_conics_810.txt k3model.m   (checks the table of conics as well)
// It proves, from the sextic S_3 and the list of 39 lines below and nothing else:
//  1. S_3 is absolutely irreducible and its only singular points are three rational ordinary nodes, so the minimal
//     resolution X of the double plane is a K3 surface with three contracted curves E_1, E_2, E_3.
//  2. S_3 restricts to a perfect square on each of the 39 lines: 21 through no node, 15 through one, 3 through two.
//     Each such line has two lifts L+ and L- to X, smooth rational curves defined over Q.
//  3. The 78 lifts and the E_n span a lattice of rank 20 and discriminant -163.  The Picard number of a K3 surface in
//     characteristic 0 is at most 20 and 163 is squarefree, so this lattice is all of NS(X), every class is defined
//     over Q, and the transcendental lattice is the unique even positive definite binary form of determinant 163,
//     [2,1;1,82].  A K3 surface of Picard number 20 is determined by its transcendental lattice (Shioda-Inose), so
//     X is X_163.
//  4. From the lattice alone: the classes x with x^2 = -2, x.D = 0 form the root system A_1^3; the irreducible
//     classes with x^2 = -2, x.D = 1 are exactly the 78 lifts, so S_3 is a square on exactly 39 lines (78 forced points);
//     the irreducible classes with x^2 = -2, x.D = 2 number 1620, so there are exactly 810 conics on which S_3 is a
//     square.  (An irreducible class of the last kind is a lift of such a conic, or the pullback of a line through
//     two nodes on which S_3 is not a square; by step 2 all three lines through two nodes are among the 39.)
//     (D is the pullback of a line.  A class x with x^2 = -2 and x.D > 0 is effective, and it is
//     irreducible if and only if x.C >= 0 for every smooth rational curve C with C.D < x.D.)
//  5. #X(F_p) = p^2 + 20p + 1 + a_p for the good primes p < 200, where a_p is the coefficient of the CM newform of
//     weight 3 and level 163: a_p = (u^2 - 163 v^2)/2 if 4p = u^2 + 163 v^2, and a_p = 0 if p is inert in
//     Q(sqrt(-163)).
//  6. The line 66 z0 + 31 z1 - 52 z2 = 0, parametrised by (6x - 2 : 4x + 16 : 10x + 7), cuts out y^2 = f_738(x);
//     the 39 lines meet it at 39 distinct values of x, each giving two rational points of C_738.
SetColumns(0); SetQuitOnError(true);
// coefficients of S_3: for d = 0..6 and j = 0..d, the coefficient of z0^(6-d) z1^(d-j) z2^j
S3coef := [25, -1930, 3240, 969, -1734, 1996, 5752, -6326, -996, -864, 2748, -1328, -3588, 960, -176,
           -144, 216, -548, 1376, 16, 0, 36, 12, -311, -52, 676, 0, 0];
// the 39 lines: [a, b, c] is a z0 + b z1 + c z2 = 0
lines := [[0,2,-3], [1,-2,-2], [1,3,-2], [1,42,-62], [5,-11,-4], [5,0,-2], [5,1,4], [5,4,4],
          [6,-4,-3], [10,-3,-8], [10,0,3], [11,1,2], [12,-2,-5], [17,6,10], [19,12,4], [20,26,5],
          [25,-21,4], [25,-3,10], [35,-60,6], [40,30,31], [59,56,-60], [0,1,-2], [1,-2,0], [1,0,0],
          [1,1,-2], [1,1,0], [1,3,-4], [2,3,0], [3,-2,0], [3,-1,2], [4,3,2], [4,5,-2],
          [5,-7,0], [5,1,-2], [6,-11,0], [7,4,6], [8,-1,2], [13,8,10], [23,-2,4]];
f738coef := [6838225, -46160230, 40133975, 172162750, 60026500, -16144220, 743425];       // constant term first
Q := Rationals(); P<z0,z1,z2> := PolynomialRing(Q, 3); R<x> := PolynomialRing(Q);
mons := []; for d in [0..6] do for j in [0..d] do Append(~mons, z0^(6-d)*z1^(d-j)*z2^j); end for; end for;
S3 := &+[S3coef[i]*mons[i] : i in [1..28]];
n := #lines; assert n eq 39;

// 1. singular points
P2 := ProjectiveSpace(P); B := Curve(P2, S3);
assert IsAbsolutelyIrreducible(B);
Sing := SingularSubscheme(B); assert Dimension(Sing) eq 0 and Degree(Sing) eq 3;
nodes := [Eltseq(p) : p in RationalPoints(Sing)]; assert #nodes eq 3;
printf "ok 1: S_3 is absolutely irreducible; its singular subscheme has degree 3 and consists of the\n";
printf "      rational points %o, %o, %o: three ordinary nodes\n", nodes[1], nodes[2], nodes[3];

// 2. the 39 lines.  For a line l with basis p1, p2 of its points, S_3(s p1 + p2) = g(s)^2; the lift L+ is
//    w = G(a, b) at the point a p1 + b p2, where G is g made homogeneous of degree 3, and L- is w = -G.
LD := [];
for l in lines do
  K := Basis(Kernel(Matrix(Q, 3, 1, l))); p1 := Eltseq(K[1]); p2 := Eltseq(K[2]);
  ok, g := IsSquare(Evaluate(S3, [p1[i]*x + p2[i] : i in [1..3]])); assert ok;
  Append(~LD, <p1, p2, g>);
end for;
function wplus(ld, v)
  c := Solution(Matrix(Q, 2, 3, [ld[1], ld[2]]), Vector(Q, v));
  return &+[Coefficient(ld[3], k)*c[1]^k*c[2]^(3-k) : k in [0..3]];
end function;
onnode := [[&+[lines[i][k]*nd[k] : k in [1..3]] eq 0 : nd in nodes] : i in [1..n]];
nn := [#[b : b in onnode[i] | b] : i in [1..n]];
for i in [1..n] do for t in [1..3] do if onnode[i][t] then      // the line meets the sextic exactly twice at the node
  g := LD[i][3]; c := Solution(Matrix(Q, 2, 3, [LD[i][1], LD[i][2]]), Vector(Q, nodes[t]));
  if c[2] ne 0 then assert Evaluate(g, c[1]/c[2]) eq 0 and Evaluate(Derivative(g), c[1]/c[2]) ne 0;
  else assert Degree(g) eq 2; end if;
end if; end for; end for;
assert [#[i : i in [1..n] | nn[i] eq k] : k in [0..3]] eq [21, 15, 3, 0];
printf "ok 2: S_3 is a square on all 39 lines; 21 pass through no node, 15 through one, 3 through two\n";

// 3. intersection numbers on X.  Index 2i-1 = L_i+, 2i = L_i-, 2n+t = E_t.
//    L.L = E.E = -2;  L.E = 1 if the line passes through the node;  L_i+.L_i- = 3 - (number of nodes on the line),
//    because L_i+ + L_i- + (the E_t on the line) = D and D.L = 1;  two lines through a common node have
//    disjoint lifts;  two lines meeting at a point v off the sextic have L_i^a.L_j^b = 1 exactly when the two
//    lifts take the same value w at v.
N := 2*n + 3; M := ZeroMatrix(Integers(), N, N);
for i in [1..n] do
  M[2*i-1,2*i-1] := -2; M[2*i,2*i] := -2; M[2*i-1,2*i] := 3 - nn[i]; M[2*i,2*i-1] := 3 - nn[i];
  for t in [1..3] do if onnode[i][t] then
    for e in [0,1] do M[2*i-e,2*n+t] := 1; M[2*n+t,2*i-e] := 1; end for;
  end if; end for;
end for;
for t in [1..3] do M[2*n+t,2*n+t] := -2; end for;
for i in [1..n] do for j in [i+1..n] do
  v := Eltseq(Basis(Kernel(Transpose(Matrix(Q, 2, 3, [lines[i], lines[j]]))))[1]);
  if exists{nd : nd in nodes | Rank(Matrix(Q, 2, 3, [nd, v])) eq 1} then continue; end if;
  sv := Evaluate(S3, v); assert sv ne 0;
  wi := wplus(LD[i], v); wj := wplus(LD[j], v); assert wi^2 eq sv and wj^2 eq sv;
  for e in [0,1] do a := 2*i - e; b := wi eq wj select 2*j - e else 2*j - 1 + e; M[a,b] := 1; M[b,a] := 1; end for;
end for; end for;
assert Rank(M) eq 20;
Ker := KernelMatrix(M); S, U, V := SmithForm(Ker); assert Set(Diagonal(S)) eq {1};
// 20 rows completing a basis of the kernel to a basis of Z^81; G is the Gram matrix of the lattice of the 81 curves
T := V^-1; Bas := RowSubmatrix(T, Nrows(Ker) + 1, 20);
G := Bas*M*Transpose(Bas);
assert Determinant(G) eq -163;
// the class of an integer combination c of the 81 curves, in the basis Bas
coord := func< c | Vector(Integers(), [(c*V)[k] : k in [Nrows(Ker)+1..N]]) >;
e := func< k | Vector(Integers(), [i eq k select 1 else 0 : i in [1..N]]) >;
ip := func< a, b | (Matrix(a)*G*Transpose(Matrix(b)))[1,1] >;
pull := func< i | coord(e(2*i-1) + e(2*i) + &+[Parent(e(1)) | e(2*n+t) : t in [1..3] | onnode[i][t]]) >;
Dcl := pull(1); assert forall{i : i in [1..n] | pull(i) eq Dcl};       // every line pulls back to the same class D
assert ip(Dcl, Dcl) eq 2;
printf "ok 3: the 78 lifts and the 3 contracted curves span a lattice of rank 20 and discriminant %o,\n", Determinant(G);
printf "      which is therefore NS(X); D^2 = %o\n", ip(Dcl, Dcl);

// 4. counting curves in the lattice.  Pd(x) = (x.D)^2 - x^2 is a positive definite integral form on NS(X).
dv := Matrix(Dcl)*G; Pd := Transpose(dv)*dv - G; assert IsPositiveDefinite(Pd) and Determinant(Pd) eq 163;
LP := LatticeWithGram(Pd);
sv := [Vector(Integers(), Eltseq(Coordinates(LP, t[1]))) : t in ShortVectors(LP, 6)]; sv := sv cat [-v : v in sv];
deg := func< v | ip(v, Dcl) >;
roots := [v : v in sv | deg(v) eq 0 and ip(v, v) eq -2];
Ecl := [coord(e(2*n+t)) : t in [1..3]];
assert Set(roots) eq Set(Ecl) join {-v : v in Ecl};
Lcl := [coord(e(k)) : k in [1..2*n]]; assert #Set(Lcl) eq 78;
d1 := [v : v in sv | deg(v) eq 1 and ip(v, v) eq -2];
d1irr := [v : v in d1 | forall{E : E in Ecl | ip(v, E) ge 0}];
assert Set(d1irr) eq Set(Lcl);
d2 := [v : v in sv | deg(v) eq 2 and ip(v, v) eq -2];
d2irr := [v : v in d2 | forall{C : C in Ecl cat Lcl | ip(v, C) ge 0}];
assert #d2irr eq 1620;
perp := Matrix(Basis(Kernel(Transpose(dv))));  Lperp := perp*G*Transpose(perp);
assert Determinant(Lperp) eq -326 and IsPositiveDefinite(-Lperp);
printf "ok 4: D-perp is negative definite of rank 19 and determinant 326, with roots +-E_1, +-E_2, +-E_3 (A_1^3);\n";
printf "      of the %o classes with x^2 = -2, x.D = 1 the irreducible ones are exactly the 78 lifts:\n", #d1;
printf "      exactly 39 lines, 78 forced points;\n";
printf "      of the %o classes with x^2 = -2, x.D = 2, %o are irreducible: exactly 810 conics\n", #d2, #d2irr;

// 5. point counts.  Each node is replaced by a conic, so #X(F_p) = sum over the plane of (1 + chi(S_3)) + 3p.
function ap(p)
  if KroneckerSymbol(-163, p) ne 1 then return 0; end if;
  ok, u, v := NormEquation(163, 4*p); assert ok and u^2 + 163*v^2 eq 4*p;
  return (u^2 - 163*v^2) div 2;
end function;
good := []; 
for p in PrimesInInterval(5, 200) do
  F := GF(p); Pp<a0,a1,a2> := PolynomialRing(F, 3); Sp := Pp ! S3;
  // skip the primes at which the sextic has more than its three nodes (bad reduction of the model)
  if Degree(SingularSubscheme(Curve(ProjectiveSpace(Pp), Sp))) ne 3 then continue; end if;
  cnt := 0;
  for pt in [[F|1,a,b] : a, b in F] cat [[F|0,1,b] : b in F] cat [[F|0,0,1]] do
    s := Evaluate(Sp, pt); cnt +:= s eq 0 select 1 else (IsSquare(s) select 2 else 0);
  end for;
  assert cnt + 3*p eq p^2 + 20*p + 1 + ap(p);
  Append(~good, p);
end for;
printf "ok 5: #X(F_p) = p^2 + 20p + 1 + a_p for the %o primes %o\n", #good, good;

// 6. the line of C_738
f738 := R ! f738coef; par := [6*x - 2, 4*x + 16, 10*x + 7];
assert 66*par[1] + 31*par[2] - 52*par[3] eq 0 and Evaluate(S3, par) eq 64*f738;
onl := [l[1]*par[1] + l[2]*par[2] + l[3]*par[3] : l in lines]; assert forall{h : h in onl | Degree(h) eq 1};
xs := [Roots(h)[1][1] : h in onl];
assert #Set(xs) eq 39 and forall{t : t in xs | Evaluate(f738, t) ne 0 and IsSquare(Evaluate(f738, t))};
printf "ok 6: S_3(6x - 2, 4x + 16, 10x + 7) = 64 f_738(x); the 39 lines meet this line at 39 distinct x,\n";
printf "      each with f_738(x) a nonzero square\n";

// 7. (optional) the table of conics: one conic per row, coefficients of z0^2, z0 z1, z0 z2, z1^2, z1 z2, z2^2
if assigned conics then
  rows := [[StringToInteger(t) : t in Split(r, " ")] : r in Split(Read(conics), "\n") | #r gt 0 and r[1] ne "#"];
  qm := [z0^2, z0*z1, z0*z2, z1^2, z1*z2, z2^2];
  qs := {}; U2<uu,vv> := PolynomialRing(Q, 2);
  for r in rows do
    q := &+[r[k]*qm[k] : k in [1..6]]; C := Conic(P2, q); assert IsNonsingular(C);
    phi := Parametrization(C); pe := DefiningPolynomials(phi);
    h := Evaluate(Evaluate(S3, [Evaluate(pp, [x, 1]) : pp in pe]), x);
    assert IsSquare(R ! h) and (Degree(R ! h) mod 2 eq 0);
    Include(~qs, q / LeadingCoefficient(q));
  end for;
  assert #rows eq 810 and #qs eq 810;
  printf "ok 7: S_3 is a square on each of the 810 distinct conics of %o, which by step 4 are all of them\n", conics;
end if;
print "ALL CHECKS PASSED";
quit;
Elkies' sextic and the two lifts
SE = 3686400x06 + (−8697600x12 + 2193664x1x2 + 11520x22)x04 + (5130225x14 − 1860100x13x2 + 138414x12x22 − 4180x1x23 + 9x24)x02 − 45144x15x2 + 19960x14x22 − 1672x13x23 + 40x12x24

It has its node at (0 : 0 : 1). The lifts of the two conics are given by the cubic forms

G1 = 120x03 + (−40909x1 − 158x2)x02 + (−2305x12 + 39x1x2)x0 − 2280x13 + 20569x12x2
G2 = −1920x03 + (4480x1 + 624x2)x02 + (−208879x12 + 666x1x2 + 9x22)x0 − 14639x13 − 389x12x2 + 9x1x22

in the sense that SE ≡ Gi2 modulo qi, and Γi = {qi = 0, w = Gi}.

The map from Elkies' model to the new one, in full

The three coordinate functions are zi = Ai + Biw (the common denominator q1q2 is dropped), with

B0 = −800x0 + 1980x1 − 260x2
B1 = 8800x0 − 15130x1 + 350x2
B2 = −3325x1 + 1255x2
A0 = 1536000x04 + (−3801600x1 − 1198080x2)x03 + (−1812000x12 + 1139328x1x2 + 2208x22)x02 + (4484700x13 + 2145348x12x2 − 204876x1x22 + 3276x23)x0 + 704484x13x2 − 42516x12x22 − 2004x1x23 + 36x24
A1 = 22272000x04 + (−47331200x1 + 504960x2)x03 + (−6552400x12 + 3901184x1x2 − 73776x22)x02 + (30724050x13 − 4682002x12x2 + 298950x1x22 − 3798x23)x0 + 2257200x14 − 393572x13x2 − 53660x12x22 + 3156x1x23 − 36x24
A2 = 19584000x04 + (−31806400x1 + 594240x2)x03 + (−13242200x12 + 276256x1x2 − 241224x22)x02 + (24965625x13 − 2984177x12x2 + 59235x1x22 − 243x23)x0 + 1128600x14 + 1038206x13x2 − 143086x12x22 + 7770x1x23 − 18x24

The statement is exact. In the chart x0 = 1, with T the function of step 4, S3(z0, z1, z2) = T2/16 as an identity of polynomials modulo w2 − SE; the script verify.py that accompanies this note checks it, together with the congruences Aj ≡ BjGi modulo qi that express the vanishing conditions. In particular, for every rational point (x0 : x1 : x2; w) of Elkies' model, S3(A0 + B0w, A1 + B1w, A2 + B2w) is a rational square. For example, the line x2 = 0 is a tritangent of Elkies' sextic: there SE = x02(1920x02 − 2265x12)2. Its lift w = x0(1920x02 − 2265x12) is mapped to the single point (0 : 2 : 1), so it is one of the three contracted curves. On the other lift, the point (1 : 1 : 0) with w = 345 maps to (118 : −118 : −75), where S3 = 75708802.

The record curve is one line of this model. The line λ = (66, 31, −52) is parametrized by (z0 : z1 : z2) = (6x − 2 : 4x + 16 : 10x + 7), and

S3(6x − 2, 4x + 16, 10x + 7) = 64 · (743425x6 − 16144220x5 + 60026500x4 + 172162750x3 + 40133975x2 − 46160230x + 6838225),

which is 64 times the sextic of C738. In Elkies' plane the image of the same curve has degree 4. It is not a line there. A curve isomorphic to it could still be a line section of another model, so we checked that directly. On none of the other 2380 sextics of the list is C738 isomorphic, even over an algebraic closure of Q, to the section by any line, rational or not, and on S3 the line λ is the only one. The proof is a computation: two genus 2 curves are isomorphic over an algebraic closure exactly when their Igusa–Clebsch invariants agree, which gives three polynomial equations in the coefficients of the line, and the script linesection.m shows, for all lines in three charts, that they have no solution on the other sextics (by a Gröbner basis, and a second time by resultants) and exactly one on S3.

A model has 2.1 × 108 lines of height at most 400, and the 2381 sextics together have 5.1 × 1011. Reducing the sextic of a line and counting its rational points, even to a small height bound, costs a millisecond or two, far too much to do for every line. So the lines are ranked first, by a count that costs about a tenth of a microsecond per line.

The score of a line. By a surface point of height at most H we mean a point (z0 : z1 : z2) of the plane with coprime integer coordinates of absolute value at most H at which the sextic is the square of a rational number, zero included. It is counted once, although two points of the surface lie over it when the sextic is not zero there. A surface point on a line is a rational x-value of the curve of that line, so a line with many surface points of small height is a line whose curve has many points of small height. The score of a line is the number of surface points of height at most 4000 on it, not counting those on the special lines: every line meets each special line, so they add the same forced points to every score.

The scores of all lines of height at most 400 are computed together, without looking at any line by itself. First the surface points are listed: for each pair (z1, z2) the values of the sextic for |z0| ≤ 4000 are sieved for squares modulo products of small prime powers, 512 values at a time, and the survivors are tested exactly. A model has between half a million and a little over two million such points (on average 1.8, 1.5 and 1.3 million for one, two and three nodes). Then, for each point, the lines of height at most 400 through it are enumerated (they are the short vectors of a lattice of rank 2), and a counter for each of these lines is incremented in a table with one entry for every line of height at most 400. The two steps together take 0.11 microseconds of processor time per line.

The stages. The settings are the same for every model.

  1. Keep the lines with a high score. Lines through a node are discarded, because their curves have genus at most 1. The others are kept if their score is at least 40: on average 7,400 lines on a model with one node, 4,300 with two nodes and 2,600 with three. (Models with fewer nodes have more lines with many points.)
  2. Count points. For each line kept, the sextic is restricted to the line and reduced, and the rational points of height at most 2000 are counted. For the curves with at least 250 of them the points of height at most 104 are counted, and for those with at least 330 of these, the points of height at most 105. The counts use a sieve over the pairs (a, c) with x = a/c that treats 512 numerators at a time.
  3. Close. For each curve with at least 440 points of height at most 105, the set of these points is closed under the four-point relation of the next section. The number of points reported for a curve is the number after closure.

On two threads a three-node model takes about 20 seconds, a two-node model about 35 and a one-node model about 120 (medians).

all lines of height ≤ 4002.1 × 10⁸score at least 40, through no node2,460at least 250 points of height ≤ 200027at least 330 points of height ≤ 10⁴4at least 400 points of height ≤ 10⁵1, with 738 points after closure
The search on the model of C738, stage by stage (bar lengths are logarithmic). It takes 22 seconds on two threads, closure included.

On the model of the previous section there are 1,328,965 surface points of height at most 4000 off the 39 special lines. Of the lines of height at most 400 that pass through no node, 249,542 contain at least 15 of them and 2460 contain at least 40; these are the lines kept. The line (66, 31, −52) has the highest score of all, 112, and it is first at every later stage.

scorepoints of height ≤ 2000≤ 104≤ 105after closure
the line (66, 31, −52)112424536662738
best other line of the model95310346376

(These counts are for the reduced sextic used in the search, which differs from the equation at the top by x ↦ −2x. In the coordinates of the equation at the top the last two are 554 and 666.)

11010²10³10⁴10⁵20406080100score: surface points of height at most 4000 on the linelineskept: score at least 40the line (66, 31, −52): 112next: 95
One line stands apart. The scores of the lines of height at most 400 on the model of C738: the number of lines, on a logarithmic scale, with a given number of surface points of height at most 4000 (lines through a node and lines with fewer than 15 points are not shown). The 2460 lines with a score of at least 40 are kept. The model is ordinary, and the record line is a single outlier that is unmistakable from the score on.
The numbers behind this chart
scorelines
15–19168,927
20–2449,962
25–2917,855
30–347,283
35–393,055
40–441,332
45–49607
50–54289
55–59127
60–6457
65–6919
70–7416
75–797
80–843
90–941
95–991
110–1141

The selection loses curves

A line of height at most 400 whose curve has many points is missed if its score is below 40, or if one of its counts falls below a threshold. The record line is far from these limits. The margin is small for weaker curves, however. On the model above, the four curves with at least 330 points of height at most 104 have ranks 1, 2, 34 and 2105 by score among the 2460 lines kept: the last has score 41, just above the bound, and 262 points of height at most 2000, against a threshold of 250. Its curve has 372 points of height at most 105; lines of that quality with a score of 39 are not seen. The section on what the search does not cover measures what is lost in the whole search.

Finding the large points

The point counts of the search stop at height 105, and going much further is expensive. On curves like these there is a far cheaper way to find the larger points, because the large points are tied to the small ones by simple relations in the Jacobian. The idea goes back to Stoll's search, which tested whether five known points lie on a cubic [19].

Take four known points P1, …, P4 on the curve y2 = f(x), with four different values of x, and let p(x) be the cubic polynomial whose graph passes through them. The graph meets the curve in six points: the four we started with and two more. The x-coordinates of those two are the roots of a quadratic r(x), the factor left over in

p(x)2 − f(x) = (p32 − f6) (x − x1)(x − x2)(x − x3)(x − x4) r(x),

where p3 and f6 are the leading coefficients (we assume p32 ≠ f6; otherwise, which requires f6 to be a square, one or two of the six intersections are at a point at infinity, and the factor left over has degree one or zero). If the discriminant of r is a square, the two new points are rational. In the Jacobian this is the relation P5 + P6 ~ 3D∞ − (P1 + P2 + P3 + P4), where D∞ is the divisor at infinity.

the curve C738the cubic through P1, …, P4four known pointsthe two further intersections
−15,000−10,000−5,00005,00010,000−2−1.5−1−0.500.51xyP₁P₂P₃P₄x = −102/109x = 1474359/1473364
Four small points produce a large one. The cubic through four points of height at most 5 meets the curve again at x = −102/109, which was known, and at x = 1474359/1473364, whose height is far beyond the point counts of the search.

In this example the four points are

P1 = (0, 2615),   P2 = (−32, 360258),   P3 = (13, 4647527),   P4 = (−53, 24881527)

the cubic through them is

p(x) = −254773x3 − 961529x2 + 54713x + 2615

and

p(x)2 − f(x) = 281 x (2x + 3) (3x − 1) (3x + 5) (109x + 102) (1473364x − 1474359).

The last two factors give the new points. The second has y = −47232542363775629256305/14733643.

Replacing some of the four points by their mirror images (x, −y) changes the cubic, so each set of four x-values gives eight tests. The closure runs over all sets of four known points and all sign patterns, adds the points it finds, and repeats until nothing new appears. With n known x-values that is about n4/3 tests, each asking whether a single rational number is a square. Almost all of them are decided without computing that number. Take a prime p of good reduction at which f has a root, and let W be the Weierstrass point of the curve over Fp above that root. The six points in which a cubic meets the curve have classes [Pi − W] that add up to zero in the Jacobian over Fp. So four points can have two rational partners only if minus the sum of their four classes is a sum of two such classes. The classes of the known points are computed once for the primes p ≤ 257 of this kind, the test for a set of four is then a table lookup for each prime, and the sets that pass for all of them (about one in five hundred on C738) are tested exactly. Two variants run as well: relations in which the cubic is tangent to the curve at a known point, and, when f6 is a square, relations through the points at infinity.

Unlike the five-point test, this finds two new points from four known ones and needs no fifth point on the cubic. There is a second difference. Write [P] for the class in the Jacobian of P minus its mirror image. The relation gives [P5] + [P6] = −([P1] + [P2] + [P3] + [P4]). So the sum of the two new classes lies in the subgroup generated by the known points, but the two new points separately need not. With five known points on a cubic the class of the sixth is minus the sum of theirs, so that test stays inside the subgroup. The closure can enlarge it, and here it does: the five starting values of the certificate below give classes that generate a subgroup of rank at most 5, and their closure is the whole list, which generates rank 21.

The points of C738

height of x at most10102103104105106107108109
points36188390554666712732736738

The count to height 105 finds 333 x-values. From those, the first pass of the closure runs through 5.0 × 108 sets of four in a few seconds and finds the other 36 x-values, of heights between 1.06 × 105 and 1.86 × 108. The second pass finds nothing. The same 369 x-values are reached from the points of height at most 103 or 104 alone, and a sieve to 107 finds no point the closure had not already found. The largest point has x = 185880347/11934737.

No point count is even needed. The closure of the five x-values of height at most 3 is already the whole list, as the next section shows.

A certificate you can run

The claim of 738 rational points can be checked from ten of them. The curve has the points

(0, ±2615),   (−12, ±381158),   (12, ±344758),   (−32, ±360258),   (13, ±4647527)

with x = 0, −1/2, 1/2, −3/2 and 1/3, and their four-point closure consists of the 369 x-values listed under Data. So these five x-values, together with the closure rule, are a certificate for the whole list: anyone can regenerate every point from them and check each one on the curve.

The program below does this in your browser. It verifies the starting points, runs the closure, shows the points as it finds them, and checks every one by testing that c6f(a/c) is a perfect square, in exact integer arithmetic. It works for any curve y2 = f(x) with f of degree 5 or 6 and any rational starting points. Points with y = 0 are used like the others. If the curve has rational points at infinity (the leading coefficient is a square, or the degree is 5), they are known from the start, they are shown as ∞, and the relations through them are found as well. To keep the running time down it only tries sets of four that contain at least two of the 16 known points of smallest height, which is enough here; with 0 in that box it tries every set of four, which finds the same 369 x-values on this curve but takes far longer. Every point remembers the four points and the cubic that first produced it. When the run has finished you can pick any point, see that relation with its exact check, and follow the four points back, step by step, to the five starting values; on this curve no point is more than 7 steps away.

Not run yet. On this curve it takes roughly ten to thirty seconds.

What is known about the curve

What the search finds

The search ran on the 2381 sextics with the settings above. In all 5.1 × 1011 lines were scored and 6.6 × 106 kept. Of their curves, 119,276 have at least 250 points of height at most 2000, 17,338 have at least 330 of height at most 104, 6,118 have at least 400 of height at most 105, and 1,061 have at least 440 and were closed. Up to isomorphism these are 1,035 different curves: 26 of them occur twice on their sextic (seven of these on Elkies' sextic, which is unchanged by x0 ↦ −x0). No curve occurs on two different models.

nodesmodelsforced pointscurves with ≥ 400 points of height ≤ 105, per modeldifferent curves with ≥ 500 points≥ 550≥ 600most points
one7104–1087256112642
two18686–9612202285638
three2,18872–861.6271479738

A model with fewer nodes is richer: it has more forced points, and many more lines with many points. But the three-node models are so much more numerous that nine of the sixteen curves with at least 600 points, and the two best, are found on them.

on a model with three nodestwo nodesone node
020406080100120140160180450500550600650700750rational points on the curve, after closurecurvesElkies–Stoll 2008: 642654738
The 1,035 different curves closed by the search, by number of rational points and by the number of nodes of the model. Curves with fewer than 440 points of height at most 105 were not closed, so the left part of the picture is incomplete. Point at a bar for the numbers.
The numbers behind this chart
pointsthree nodestwo nodesone node
450–459311
460–46925244
470–47979478
480–489815410
490–4991015612
500–509765814
510–519634410
520–52933336
530–539302310
540–54922165
550–5591983
560–569742
570–579552
580–589220
590–599542
600–609211
610–619200
620–629220
630–639120
640–649001
650–659100
730–739100

A second curve, with 654 points

Three curves have at least 642 points: C738, the curve C642 of Elkies and Stoll, which the search finds on the line (95, −9, −82) of a one-node model with 106 forced points, and

C654 :   y2 = 963396x6 − 25653180x5 − 257315387x4 + 5817403620x3 + 43135861216x2 − 76150919040x + 48962544400

It is the section by the line (83, 6, 125) of a three-node model with 84 forced points (26 tritangents and 16 node lines). It has 570 points of height at most 105, and the closure adds 42 x-values, for 654 points. Its Jacobian is absolutely simple of rank at least 22, its automorphism group has order 2, and its sextic has Galois group S6. On its model no other line has as many as 440 points of height at most 105, so this curve too is an outlier line on an ordinary model.

Sixteen curves have at least 600 points.

pointsof height ≤ 105nodesforced pointsmodellinescorerank by score
738662378t0453(66, 31, −52)1121
654570384t1272(83, 6, 125)821
6425541106b4(95, −9, −82)8032
638*604292p144(145, 2, 50)901
636562292p008(154, −369, 102)51629
632546382t2139(22, −295, −100)5849
626528290p098(225, 219, 112)471,067
620558378t1782(387, 99, −16)46830
620544294p101(165, 157, −106)7628
620538378t0497(291, 130, −215)7033
614524380t1614(227, 224, 186)59110
610574378t0555(30, 114, 313)777
604524376t1068(35, 76, 224)844
604522378t2005(13, −230, 1)6445
602552292p053(173, 109, −56)804
6025121108b5(106, −113, −28)65242

The model is the name of its row in models.txt, and the line is given in the coordinates of that row. The last two columns are the score of the line and its position among the lines of its model ordered by score. The Jacobian of the curve marked * is isogenous to a product of two elliptic curves; the other fifteen are absolutely simple.

The curves are generic. For 1,060 of the 1,061 curves the sextic has Galois group S6, so by Zarhin's theorem [23] the Jacobian has endomorphism ring Z over an algebraic closure: it is absolutely simple, and the curve has no automorphism besides the hyperelliptic involution. The exception is the curve with 638 points in the table. It has an extra involution and its Jacobian splits, so it belongs with the 588-point and 648-point curves of the introduction, and it has fewer points than the latter. Curves with split Jacobians are a different subject with their own constructions; they are not considered further here.

Forced points do not predict the best curves. Between the three kinds of model the number of forced points matters, as the table above shows. Among models of one kind it does not help. Among the three-node models the mean number of curves with at least 400 points of height at most 105 falls from 2.4 per model for the models with 72 forced points to 1.3 for those with 84 and 0.8 for the five with 86 (rank correlation −0.13 over the 2188 models). The models of the two best curves have 78 and 84 forced points, of a possible 72 to 86, and on each of them no other line has as many as 440 points of height at most 105. The records are isolated lines on models that are otherwise unremarkable.

012372747678808284861046331687684359665forced points of the modelcurves per model
More forced points, no better. Three-node models: the mean number of curves with at least 400 points of height at most 105 on a model, by its number of forced points. The number above each bar is the number of models. Point at a bar for the numbers.
The numbers behind this chart
forced pointsmodelscurves with ≥ 400 points of height ≤ 105, per modelmodels with a curve of at least 500 points
72102.401
74461.546
763311.9141
786871.6478
806841.5889
823591.2534
84661.267
8650.800

What the search does not cover

Two remarks. The first is that the number of forced points, which is what one sees first in a model, is a poor guide to its best curve. The best curves are isolated lines, and they turned up among the models with the fewest forced points of the three kinds, because there are so many of those. The second is that the space of models is very large and almost entirely unexplored. The models with four or more singular points number tens of thousands at least, none of which is searched here, and on the models that are searched only the lines of height at most 400 in one system of coordinates have been looked at. There is no reason to think that 738 is the largest number of points on a line section of X163.

Reproducing the results

These files accompany the note. They contain what is needed to check each statement, and to repeat the computation from Elkies' sextic. The scripts need Magma, PARI/GP or Python, as indicated; each one says in its first lines what it proves and how to run it.

The computation

The input is the file elkies163.txt: the 28 coefficients of Elkies' sextic. The script run_all.sh does everything described under Building the models and Finding the line: build_models.sh constructs the 2381 sextics (the programs it uses are described in models/README.md), search_all.sh searches them, and the results are collected in tables and compared with the stored ones. The scripts call C programs, whose sources are included (k3pts.c lists the surface points, k3inc.c computes the scores, linesearch.c restricts the sextic to a line, g2rp.c counts points, cubic.c is the closure; they need a processor with AVX-512), PARI/GP, and Magma. Nothing in the computation is random: the output is determined by the input and the settings. (The sextic of a model depends on what MinRedTernaryForm returns, so another version of Magma could give other coordinates for some models.) A model can be searched by itself:

search/search_model.sh t0453

prints the numbers of the section Finding the line and ends with the line (66, 31, −52) and 738 points, in about twenty seconds, and search/line.sh b5 327 -285 -512 treats a single line. The results of the whole run are in models.txt (the sextics), search_models.txt (one row per sextic: the numbers at each stage and the times), search_curves.txt (the 1,061 curves closed, with model, line, reduced sextic, numbers of points, score, Galois group and isomorphism class), S3_search.txt (the scores and counts of the lines kept on the model of C738), search_sample.txt (the sample of the section on what the search does not cover) and search_summary.txt (machine, settings and times). README.md lists every file.

Certificates

statementfiles
at least 738 rational pointsthe program on this page (closure.js), which regenerates them from five; or witnesses_738.json, one relation for each of the other 364 x-values, verified without any search by check_witnesses.py (Python); the list itself is points_738.txt
Elkies' sextic is a model, over Q, of the surface given by Inose's equationverify.m and verify_inverse.m (Magma, about a minute), with the maps in map_data.m and inverse_data.m; inose/README.md describes them
each of the 2381 sextics is a model of X163 with the nodes and lines listed; they are 2381 different modelsmodels.m (Magma, about ten minutes), with models.txt
the 2381 models are models of the census: all those with one or two nodes, and 2188 of the 2192 with threecensus.m (Magma), with census.txt, the lattices of the census with at most three nodes
the model S3: exactly 39 special lines and 810 conics, point countsk3model.m (Magma), with S3_coefficients.txt, S3_lines.txt, S3_conics_810.txt
the displayed formulas, and the map between the two modelsverify.py (Python and PARI/GP), with model_map.txt
no torsion, rank at least 21, the basis of Λ, saturation at 2, …, 13, regulatormw.m (Magma), with mw_basis.txt and mw_coords.txt
rank at most 21, assuming GRHselmer.m (Magma, about four hours)
C738 is a line section of exactly one of the 2381 sextics, by exactly one linelinesection.m (Magma, a few hours in all; it can be run in pieces)
the curves of the search: each is the stated line section; Galois groups; isomorphism classescurves.m (Magma), with search_curves.txt
the masses of the two genera; the census counts; the figures quoted for the searchmass.m (Magma); check_census.py (Python), which checks the census counts against the census file of the project and the search figures against the data files
this note as a papercurve738.pdf

Two things are not in the package: the random walks of the census (the computation does not use the census, and the sextics are certified one by one), and the enumeration up to height 350 quoted above, which is evidence and not part of any proof.

Computations

The programs that list surface points, score lines, count points and close them are C programs written for this project. They use AVX-512 instructions: the two sieves treat 512 values at a time. The lattice computations and the construction of the models use PARI/GP [13] (2.18.1) and Magma [10] (V2.29-10), and the certificates use Magma. Everything ran under Linux on one machine with a 16-core AMD Ryzen AI Max+ 395 processor (Zen 5, up to 5.2 GHz, 32 threads).

The whole computation, from Elkies' sextic to the curves, took 62 minutes on that machine.

partminutesprocessor hours
constructing the models7.33.5
listing the points of the surfaces27.513.4
scoring the lines5.12.3
the curves of the lines kept21.99.6
   restriction and reduction3.21.4
   points of height at most 20007.63.4
   points of height at most 1041.60.6
   points of height at most 1057.73.4
   closure1.70.7
total61.828.8

Sixteen models are searched at a time, on two threads each, so the parts of the search overlap; its time is divided among them in proportion to the time that the sixteen jobs spent in each.

The only statement of this note that depends on an unproved hypothesis is the upper bound for the rank, where Magma's class group computations assume GRH. Canonical heights are computed numerically; they enter only the regulator and the enumeration up to height 350, which is evidence and not a proof.

AI use

The search, the computations, and the text and figures of this note were produced by Claude Fable 5.2 (Anthropic), working under the direction of the author, who posed the problem and supervised the work. “We” refers to both.

Acknowledgments

We thank GPT-6 Astra (OpenAI) for helpful reviews of two earlier drafts of this manuscript.

Data

The 369 x-values of C738, with y

One line per x-value, in order of height: (a : b : c) stands for the two points with x = a/c and y = ±b/c3.

(0 : 2615 : 1)
(-1 : 38115 : 2)
(1 : 34475 : 2)
(-3 : 36025 : 2)
(1 : 46475 : 3)
(-3 : 242515 : 4)
(1 : 54135 : 4)
(4 : 633545 : 3)
(-5 : 248815 : 3)
(-2 : 591745 : 5)
(-1 : 506205 : 5)
(5 : 491985 : 2)
(-6 : 450065 : 1)
(-7 : 920315 : 3)
(-7 : 107055 : 5)
(4 : 1913805 : 7)
(1 : 732305 : 8)
(8 : 45745 : 1)
(-11 : 3101735 : 6)
(9 : 14050435 : 11)
(12 : 11937445 : 7)
(12 : 7240625 : 5)
(13 : 1254295 : 2)
(-16 : 4192965 : 11)
(3 : 3662395 : 16)
(16 : 35839195 : 11)
(16 : 439065 : 1)
(-17 : 6711705 : 1)
(-17 : 12726845 : 5)
(2 : 7376985 : 17)
(4 : 3850115 : 17)
(5 : 6072885 : 17)
(-5 : 25721465 : 18)
(19 : 13590395 : 5)
(11 : 41519205 : 20)
(-22 : 25260795 : 5)
(-23 : 22768165 : 3)
(23 : 2743895 : 3)
(-24 : 36429715 : 11)
(-25 : 17903655 : 17)
(25 : 10171665 : 4)
(-27 : 11154155 : 28)
(-29 : 65781485 : 12)
(-31 : 61190145 : 5)
(-31 : 74066895 : 8)
(-30 : 4606525 : 31)
(31 : 5888025 : 4)
(9 : 36015385 : 32)
(1 : 84134995 : 33)
(31 : 497697795 : 34)
(-29 : 178214085 : 37)
(-17 : 242041635 : 37)
(37 : 86976365 : 9)
(37 : 15542425 : 5)
(-38 : 60277455 : 1)
(-38 : 26351815 : 27)
(-41 : 79293375 : 43)
(13 : 105767595 : 43)
(43 : 29737015 : 2)
(-4 : 304163545 : 45)
(8 : 87633275 : 45)
(-47 : 128253125 : 2)
(-50 : 130440915 : 1)
(50 : 1498776895 : 43)
(50 : 47624005 : 7)
(50 : 83472935 : 1)
(-52 : 324176265 : 11)
(-53 : 312389595 : 31)
(23 : 478263555 : 53)
(53 : 843459925 : 27)
(57 : 593821295 : 20)
(58 : 1186350165 : 31)
(-46 : 729843345 : 59)
(50 : 2300851215 : 59)
(61 : 159801525 : 1)
(-41 : 1158200505 : 64)
(-61 : 599331425 : 67)
(-71 : 594371745 : 7)
(11 : 489133575 : 74)
(-75 : 1140742405 : 28)
(-67 : 1532079045 : 80)
(-25 : 2570693895 : 83)
(19 : 440888305 : 83)
(83 : 6407919315 : 68)
(85 : 1376197405 : 24)
(86 : 718275915 : 17)
(-65 : 2638328175 : 88)
(-76 : 1975780485 : 89)
(-57 : 3115203455 : 89)
(26 : 855879255 : 89)
(34 : 1699133345 : 89)
(-1 : 1977167885 : 90)
(95 : 72734725 : 6)
(-97 : 1544060915 : 10)
(-101 : 2674596925 : 48)
(-91 : 2216258055 : 101)
(-103 : 1727791205 : 9)
(-86 : 3690429545 : 105)
(-107 : 3021871765 : 27)
(107 : 834257655 : 2)
(-102 : 1926936035 : 109)
(6 : 2731302145 : 109)
(-66 : 6659172205 : 113)
(-65 : 7925109995 : 119)
(119 : 883980625 : 4)
(-122 : 1541678985 : 85)
(-125 : 4641534495 : 67)
(127 : 252409185 : 16)
(-136 : 6365233955 : 37)
(-141 : 3297553595 : 5)
(-50 : 15221202815 : 149)
(117 : 34809586385 : 157)
(-29 : 18469555445 : 168)
(-169 : 3943310525 : 118)
(-19 : 17208080505 : 170)
(179 : 13934334645 : 53)
(179 : 779941575 : 11)
(199 : 2563600725 : 10)
(-209 : 17631092115 : 127)
(212 : 16025777495 : 51)
(-185 : 27933010105 : 216)
(118 : 53426272095 : 223)
(223 : 4972487135 : 9)
(-225 : 30827322085 : 86)
(-140 : 54403489305 : 229)
(61 : 11549123145 : 229)
(236 : 13677759095 : 45)
(-226 : 20173686925 : 241)
(40 : 14767750305 : 241)
(149 : 90368033525 : 241)
(-251 : 39546602865 : 127)
(62 : 13120171155 : 251)
(-193 : 73370786335 : 265)
(265 : 215139516725 : 222)
(-236 : 51536875995 : 271)
(109 : 54180074795 : 272)
(276 : 3077731865 : 35)
(29 : 36443002995 : 281)
(283 : 149613228765 : 160)
(284 : 6632892035 : 15)
(307 : 19455521785 : 6)
(131 : 90317942935 : 310)
(-233 : 114162130565 : 311)
(290 : 395381992005 : 311)
(-303 : 1677044005 : 313)
(229 : 267152016285 : 313)
(316 : 159368200145 : 149)
(-29 : 106937356925 : 317)
(317 : 23672982375 : 53)
(337 : 43773419925 : 67)
(166 : 161004648035 : 339)
(325 : 544669490745 : 343)
(-358 : 120499983835 : 111)
(367 : 128211698615 : 113)
(135 : 109871643875 : 373)
(-376 : 18847067085 : 389)
(-383 : 161100244305 : 424)
(443 : 1290319905875 : 445)
(451 : 54013120695 : 70)
(172 : 225541721025 : 469)
(471 : 10396374065 : 59)
(-437 : 184440514615 : 473)
(32 : 210415706865 : 473)
(-477 : 237959606885 : 272)
(477 : 42079094915 : 67)
(-153 : 515266350655 : 484)
(-206 : 549807038235 : 487)
(494 : 67489965865 : 15)
(499 : 36380639295 : 67)
(-379 : 535870545595 : 516)
(-509 : 57144179365 : 527)
(223 : 449738884315 : 534)
(544 : 118957143825 : 7)
(-577 : 224685318375 : 20)
(-142 : 837433279485 : 581)
(-596 : 221279042205 : 409)
(107 : 216034056385 : 606)
(-607 : 594006067795 : 198)
(615 : 164056785805 : 101)
(-641 : 100212068385 : 457)
(668 : 4050414054225 : 629)
(-617 : 541833196275 : 670)
(-656 : 270651330675 : 685)
(191 : 345779618805 : 686)
(-559 : 1051092848855 : 687)
(695 : 2576225707825 : 436)
(279 : 918583932695 : 707)
(-710 : 583766758575 : 67)
(-38 : 1202033509875 : 733)
(802 : 143365460175 : 43)
(867 : 2499174561205 : 341)
(-865 : 929531997405 : 914)
(231 : 659992994335 : 916)
(945 : 70776197635 : 118)
(-985 : 2032047791375 : 172)
(-387 : 4933090879915 : 1016)
(1019 : 312870114585 : 137)
(1025 : 170685529245 : 62)
(-481 : 5421892873815 : 1043)
(-1079 : 810943889525 : 761)
(31 : 3378816867195 : 1126)
(-901 : 5842451534465 : 1171)
(-1147 : 1211258431295 : 1194)
(-699 : 8897466284525 : 1240)
(1283 : 6600639783465 : 443)
(-1367 : 2914353229645 : 930)
(1371 : 405500721535 : 83)
(1372 : 3150866570995 : 281)
(374 : 2624983062975 : 1373)
(-1376 : 2507479356935 : 15)
(378 : 2767831028105 : 1433)
(-1389 : 2868681708695 : 1454)
(1473 : 42314804851165 : 1361)
(293 : 3482135893775 : 1570)
(1572 : 61696557065 : 101)
(1659 : 1522570612175 : 227)
(-220 : 18297263123065 : 1717)
(-1739 : 10699925879475 : 1033)
(843 : 21567155951575 : 1762)
(283 : 7179636116835 : 1837)
(-1842 : 15979664297905 : 901)
(-1856 : 17278897760135 : 687)
(-298 : 38763662343795 : 2195)
(2358 : 2695338114835 : 137)
(2411 : 13648104995525 : 447)
(-2127 : 36874184278165 : 2435)
(-2459 : 38049693529365 : 1201)
(763 : 21501418743585 : 2566)
(2591 : 5952016366425 : 356)
(2752 : 7506471010675 : 131)
(-2759 : 45254421029675 : 1590)
(-2305 : 77739323386725 : 2861)
(-2204 : 89277519488495 : 2891)
(71 : 62810168021055 : 2969)
(1081 : 56537921817265 : 2999)
(196 : 55434990298545 : 3019)
(-3048 : 50353401671675 : 1907)
(-3050 : 27365538504645 : 2107)
(712 : 27665271652095 : 3295)
(-3310 : 92105081744445 : 911)
(3421 : 14546200536985 : 162)
(1075 : 59702794749455 : 3537)
(-556 : 169022068457165 : 3543)
(3569 : 209253053849075 : 1582)
(886 : 38258882300485 : 3585)
(933 : 42899208591085 : 3656)
(-3701 : 102115936307385 : 571)
(-3505 : 117965694423915 : 3863)
(-4316 : 162282173431375 : 669)
(4507 : 23178561715465 : 249)
(4512 : 3551736348745 : 289)
(-847 : 405914697144905 : 4687)
(-5189 : 378167336690715 : 1996)
(-5451 : 437285812293785 : 1990)
(-5538 : 189549547102745 : 157)
(5751 : 141132361932265 : 961)
(5795 : 947309736178575 : 2666)
(6016 : 874890359007665 : 2439)
(-5459 : 662404427602785 : 6301)
(2907 : 963375346353875 : 6680)
(-6925 : 143445159058465 : 4929)
(2663 : 841875308805685 : 7335)
(-7459 : 1118782858805755 : 3099)
(-5368 : 1726910076384685 : 7531)
(7636 : 164455838167005 : 1063)
(1725 : 359008702198055 : 7762)
(-7771 : 1065320586165825 : 1583)
(1649 : 420231903686595 : 8054)
(-8141 : 1274537407541555 : 4388)
(818 : 1022250627968985 : 8435)
(1865 : 486526106241945 : 8576)
(1498 : 638330385372795 : 8629)
(8791 : 265878424629405 : 1237)
(8896 : 673263145105915 : 1637)
(9190 : 39703281215675 : 587)
(69 : 2286420042989005 : 9641)
(9773 : 751508620922565 : 59)
(10249 : 7418137746550905 : 5968)
(10541 : 1692664527021125 : 2344)
(10579 : 2720427321531645 : 3031)
(-4727 : 6039644355870125 : 10818)
(11141 : 1687423866012225 : 2282)
(-3793 : 7108578644071905 : 11579)
(11837 : 657960480920655 : 1670)
(-12185 : 4260491349935445 : 6586)
(-12270 : 2961554393933135 : 7897)
(2081 : 2288598477390225 : 12821)
(391 : 7190964285833195 : 14474)
(10775 : 28654273970648235 : 15074)
(1241 : 7218001990726065 : 15668)
(-16136 : 8611867128632005 : 17373)
(-163 : 14579873131612635 : 17552)
(17565 : 4565386109940425 : 37)
(4377 : 4787032037123105 : 18119)
(-18811 : 6460531090074215 : 222)
(20259 : 7572984333232045 : 3655)
(-3975 : 37749442614919805 : 21163)
(21203 : 10004751871146945 : 4058)
(-3493 : 38836017395275265 : 21649)
(-22423 : 11753940783801195 : 446)
(-23241 : 16058826080126245 : 1114)
(2489 : 21406556256384915 : 23633)
(3765 : 15584859520126445 : 23971)
(-25162 : 39060300786837155 : 6243)
(25625 : 40171132744253665 : 7509)
(-26062 : 43900259420498355 : 6677)
(4135 : 22669782648887605 : 26919)
(25829 : 275679297902020005 : 27428)
(-27963 : 40048984831609985 : 17327)
(20114 : 186195053878768935 : 28115)
(-28202 : 55938081507962225 : 7343)
(10777 : 55238036419490265 : 29167)
(29456 : 283312780661578545 : 23945)
(-30153 : 31287505288341575 : 944)
(4213 : 49178085963400485 : 32779)
(-35929 : 92298545415097175 : 5392)
(29559 : 499610576958907645 : 36230)
(-36244 : 124888365833605605 : 16817)
(-3939 : 169570994792930885 : 36544)
(4325 : 74819204735810535 : 36791)
(-7033 : 203393342172153455 : 37068)
(-46953 : 197734776153048365 : 28654)
(-51253 : 293046201388858645 : 29359)
(49051 : 1858184926364861185 : 51473)
(55304 : 902678305985250795 : 27059)
(8489 : 204023856451466205 : 55787)
(55829 : 149839974472181105 : 9853)
(-58957 : 513545976445095405 : 15572)
(10738 : 251568352223063515 : 62823)
(14163 : 206494244731150805 : 64511)
(39731 : 1915421578146731655 : 70859)
(-12149 : 1518085837438614675 : 73303)
(-78874 : 990646352321673985 : 47023)
(105599 : 2903561998642024405 : 31543)
(29492 : 1285081691671477175 : 107955)
(116429 : 591380001758190155 : 5409)
(16483 : 2195170614190728405 : 119032)
(-128255 : 5505710012294584015 : 60248)
(-126883 : 774321928808236195 : 131299)
(-144133 : 3355076889239355 : 103349)
(-147674 : 7073253324897034275 : 84115)
(-158875 : 10784140141202943515 : 67204)
(142935 : 54001535446875089855 : 169259)
(-69343 : 29249235495581860395 : 183986)
(186068 : 46180542049365695895 : 111347)
(-9323 : 19809451945272711015 : 186853)
(-140473 : 29220704506192434235 : 194365)
(-212023 : 50086060664262124645 : 254470)
(-161081 : 235024313065359870325 : 366554)
(-368617 : 109592204906049101775 : 68795)
(124843 : 93688744530339843885 : 413152)
(468718 : 112428370443197874065 : 91269)
(-474212 : 45575075619486852075 : 490951)
(-714674 : 791009551319834160545 : 772439)
(-782807 : 166063834999375648335 : 558652)
(-883007 : 771518401465257183465 : 601876)
(-1021593 : 2869660629719151131185 : 361772)
(1253101 : 4891592881753275668355 : 376132)
(1474359 : 47232542363775629256305 : 1473364)
(1569292 : 53439226063774431440315 : 1496615)
(2849584 : 8786941423420006701855 : 131239)
(-3293069 : 53924053696892642961635 : 255735)
(-3450501 : 42304341977344557116545 : 3608630)
(1014709 : 77078856269975182433415 : 4642999)
(-6033322 : 44041868071125499038915 : 4319267)
(-6163180 : 4402591487801175816374535 : 9929603)
(-15897946 : 6545059057243516112665495 : 1495381)
(-20767006 : 18993516384121797154110705 : 12077099)
(185880347 : 146509009612568651183470335 : 11934737)
The other curves
C654 :   y2 = 963396x6 − 25653180x5 − 257315387x4 + 5817403620x3 + 43135861216x2 − 76150919040x + 48962544400
C642 :   y2 = 82342800x6 − 470135160x5 + 52485681x4 + 2396040466x3 + 567207969x2 − 985905640x + 247747600

The 648-point curve of GPT-6 Astra is given in the introduction.

References

  1. N. Bruin, M. Stoll, The Mordell–Weil sieve: proving non-existence of rational points on curves, LMS J. Comput. Math. 13 (2010), 272–306. doi:10.1112/S1461157009000187.
  2. L. Caporaso, J. Harris, B. Mazur, Uniformity of rational points, J. Amer. Math. Soc. 10 (1997), 1–35. doi:10.1090/S0894-0347-97-00195-1.
  3. A. Degtyarev, Smooth models of singular K3-surfaces, Rev. Mat. Iberoam. 35 (2019), 125–172. doi:10.4171/rmi/1051.
  4. N. D. Elkies, About the cover: Rational curves on a K3 surface, in: Arithmetic Geometry, Clay Math. Proc. 8, Amer. Math. Soc., 2009, 1–4.
  5. N. D. Elkies, Mordell–Weil generators for singular Shioda–Inose surfaces over Q, web page, people.math.harvard.edu/~elkies/K3_20SI.html (accessed 17 November 2026).
  6. A.-S. Elsenhans, M. Stoll, Minimization of hypersurfaces, Math. Comp. 93 (2024), 2513–2555. doi:10.1090/mcom/3924. arXiv:2110.04625.
  7. Epoch AI, Many rational points on a genus-2 curve, FrontierMath: Open Problems, epoch.ai/frontiermath/open-problems/genus-2-rational-points (accessed 14 November 2026).
  8. H. Inose, Defining equations of singular K3 surfaces and a notion of isogeny, in: Proceedings of the International Symposium on Algebraic Geometry (Kyoto, 1977), Kinokuniya, Tokyo, 1978, 495–502.
  9. W. Keller, L. Kulesz, Courbes algébriques de genre 2 et 3 possédant de nombreux points rationnels, C. R. Acad. Sci. Paris Sér. I Math. 321 (1995), 1469–1472.
  10. W. Bosma, J. Cannon, C. Playoust, The Magma algebra system I: The user language, J. Symbolic Comput. 24 (1997), 235–265. doi:10.1006/jsco.1996.0125.
  11. J. S. Müller, M. Stoll, Canonical heights on genus-2 Jacobians, Algebra Number Theory 10 (2016), 2153–2234. doi:10.2140/ant.2016.10.2153.
  12. V. V. Nikulin, Integral symmetric bilinear forms and some of their applications, Math. USSR Izv. 14 (1980), 103–167. doi:10.1070/IM1980v014n01ABEH001060.
  13. The PARI Group, PARI/GP version 2.18.1, Univ. Bordeaux, pari.math.u-bordeaux.fr.
  14. B. Saint-Donat, Projective models of K-3 surfaces, Amer. J. Math. 96 (1974), 602–639. doi:10.2307/2373709.
  15. T. Shioda, Kummer sandwich theorem of certain elliptic K3 surfaces, Proc. Japan Acad. Ser. A Math. Sci. 82 (2006), 137–140. doi:10.3792/pjaa.82.137.
  16. T. Shioda, H. Inose, On singular K3 surfaces, in: Complex Analysis and Algebraic Geometry, Iwanami Shoten, Tokyo, 1977, 119–136. doi:10.1017/CBO9780511569197.010.
  17. M. Stoll, Implementing 2-descent for Jacobians of hyperelliptic curves, Acta Arith. 98 (2001), 245–277. doi:10.4064/aa98-3-4.
  18. M. Stoll, Documentation for the ratpoints program, arXiv:0803.3165.
  19. M. Stoll, On the average number of rational points on curves of genus 2, arXiv:0902.4165.
  20. M. Stoll, Rational points on curves of genus 2: experiments and speculations, Dagstuhl Seminar Proceedings 09221 (2009). doi:10.4230/DagSemProc.09221.5.
  21. M. Stoll, A genus 2 curve with at least 642 rational points, web page, mathe2.uni-bayreuth.de/stoll/recordcurve.html.
  22. J. Voight, Kneser's method of neighbors, arXiv:2308.11566.
  23. Yu. G. Zarhin, Hyperelliptic Jacobians without complex multiplication, Math. Res. Lett. 7 (2000), 123–132. doi:10.4310/MRL.2000.v7.n1.a11.