Lines on K3 quartic surfaces in characteristic 3

We investigate the number of straight lines contained in a K3 quartic surface X defined over an algebraically closed field of characteristic 3. We prove that if X contains 112 lines, then X is projectively equivalent to the Fermat quartic surface; otherwise, X contains at most 67 lines. We improve this bound to 58 if X contains a star (ie four distinct lines intersecting at a smooth point of X). Explicit equations of three 1-dimensional families of smooth quartic surfaces with 58 lines, and of a quartic surface with 8 singular points and 48 lines are provided.


Introduction
A quartic surface in P 3 whose singularities are isolated rational double points is called a K3 quartic surface.Together with [17] and [18], this paper forms a trilogy dedicated to the study of the number of straight lines lying on a K3 quartic surface X, denoted by |Fn X|.
The field K over which X is defined is always supposed to be algebraically closed.The aim of this paper is to investigate the case char K = 3.
1.1.Background.Unlike smooth cubic surfaces, which always contain 27 lines, a dimension count shows that a general smooth quartic surface does not contain any line at all.One is therefore led to study the maximal number of lines that a quartic surface can contain.
Historically, the main focus has been on smooth complex quartic surfaces, which are an example of algebraic K3 surfaces.In 1882 F. Schur [12] discovered a surface with 64 lines that now carries his name, given by the following equation: The fact that |Fn X| ≤ 64 for a smooth complex quartic surface X was proven by B. Segre in 1943 [14].Around 70 years later, though, Rams and Schütt [9] discovered a flaw in Segre's argument (further investigated in a spin-off paper [7]) and fixed his proof, extending Segre's theorem to the case char K = 2, 3.
Approximately at the same time, Degtyarev, Itenberg and Sertöz, spurred by a remark of Barth [1], gave another solution with an algebraic approach.Their work resulted in a complete classification of large configurations of lines on smooth complex quartic surfaces up to projective equivalence [5].There are 8 possible configurations with more than 52 lines, corresponding to 10 distinct surfaces.A list of explicit equations for these surfaces, initiated by Schur [12], Rams and Schütt [9], Degtyarev, Itenberg and Sertöz [5], and Shimada and Shioda [15], was completed by the author [19].
If char K = 2, 3, the inequality |Fn X| ≤ 64 holds more generally for a (possibly non-smooth) K3 quartic surface X [18].The bound of 64 lines is not known to be sharp for non-smooth K3 quartic surfaces.To the author's knowledge, the record for the number of lines contained in a non-smooth K3 quartic surface in the case char K = 0 (respectively char K > 3) is attained by a surface with 52 lines (respectively by its reduction modulo 5 with 56 lines), whose equation is published in this paper for the first time (see Example 5.5).
If char K = 2, a sharp bound is known for both smooth and non-smooth K3 surfaces.More precisely, if X is smooth, then |Fn X| ≤ 60.The bound was found by Degtyarev [4] and a surface attaining the bound was found by Rams and Schütt [10].More generally, if X is a K3 quartic surface, |Fn X| ≤ 68 and if equality holds, then X is not(!) smooth and projectively equivalent to a member of a 1-dimensional family discovered by Rams and Schütt [17].
1.2.Principal results.The Fermat quartic surface, which is defined by (1) x 4 + y 4 + w 4 + z 4 = 0, contains 48 lines over C. Its reduction modulo 3 is a smooth, supersingular (ie of Picard rank 22) surface and contains 112 lines.It is the smooth quartic surface with the highest number of lines in characteristic 3, a result proven independently by Rams and Schütt [8] and Degtyarev [4].The main result of this paper is the following generalization.
We refine Theorem 1.1 in the case that X contains four distinct lines meeting in a smooth point, a configuration which we call a star.This configuration is particularly relevant in characteristic 3.For instance, each line in the Fermat surface is contained in 10 stars, and if two lines on the Fermat surface intersect, then they belong to the same star.
Addendum 1.2 (see Proposition 4.9).Assume char K = 3.If a K3 quartic surface X is not projectively equivalent to the Fermat surface and contains a star, then |Fn X| ≤ 58.
For smooth surfaces more precise results were found by Degtyarev [4], distinguishing between supersingular and non-supersingular surfaces.
A smooth supersingular quartic surface can contain 112, 58 or at most 52 lines ([4, Theorem 1.2]).There exist three different possible configurations of 58 lines.In Example 5.1, Example 5.2 and Example 5.3, explicit equations of the three 1-dimensional families of smooth quartic surfaces containing these configurations are provided.All of them contain a star.The first two families were discovered independently by Degtyarev ([4, Proposition 8.13 and Proposition 8.14]).To the author's knowledge, the third one is new.
Addendum 1.2 and the known results about smooth surfaces provide evidence for the following conjecture.
Neither the bound of 67 lines in Theorem 1.1 nor the bound of 58 lines in Addendum 1.2 are known to be sharp for non-smooth K3 quartic surfaces.In Example 5.4 we show the existence of a non-smooth K3 quartic surface with 48 lines in characteristic 3, the highest number known to the author.
Finally, in Example 5.5 we exhibit a non-smooth K3 quartic surface with 52 lines in characteristic 0 and 56 lines in characteristic 5.

1.3.
Contents of the paper.The main technique used in this paper is the study of the genus 1 fibration induced by a line on a K3 quartic surface.A genus 1 fibration can be elliptic or a quasi-elliptic, the latter appearing only in characteristic 2 and 3 by a theorem of Tate [16].Heuristically, quasi-elliptic fibrations can carry a higher number of fiber components than elliptic fibrations, thus allowing for a higher number of lines on X.This technique is explained in §2, where the notation is also fixed.
The technical core of the paper is contained in §3, where several bounds on the possible valencies of a line (see Definition 2.3) are given.
The proofs of the main results are then carried out in §4.Finally, various examples of K3 quartic surfaces with many lines are discussed in §5.
Acknowledgments.I am particularly indebted to Sławomir Rams and Matthias Schütt for suggesting the problem of studying the number of lines on K3 quartic surfaces.I also wish to thank Fabio Bernasconi, Alex Degtyarev and Víctor González-Alonso for many stimulating discussions.

Preliminaries
In this section we fix our notation and recall some general facts about K3 quartic surfaces (see [17, §2] for more details).
2.1.Lines and singular points.Starting from here and throughout the paper, X denotes a K3 quartic surface defined over a field K of characteristic p ≥ 0, Sing(X) is the set of singular points on X (which can be of type A n , D n or E n ), ρ : Z → X is the minimal resolution of X (Z is a K3 surface), H is a hyperplane divisor in the linear system defined by ρ * (O X (1)), ℓ is a line on X, L is the strict transform of ℓ on Z and |Fn X| is the number of lines lying on X.
The pencil of planes {Π t } t∈P 1 containing ℓ induces a genus 1 fibration π : Z → P 1 .A line ℓ is elliptic (respectively quasi-elliptic) if it induces an elliptic (respectively quasielliptic) fibration.For t ∈ P 1 , the residual cubic c t is the union of the irreducible components of X ∩ Π t different from ℓ.A fiber F t of π is the pullback through ρ of the residual cubic c t contained in X ∩ Π t .A general fiber of π is denoted by F .
The restriction of π to L is again denoted by π.The degree of ℓ is the degree of the morphism π : L → P 1 .(If π : L → P 1 is constant, we say that ℓ has degree 0.) The line ℓ is separable (respectively inseparable) if π : L → P 1 is separable (respectively inseparable).The singularity of ℓ is the cardinality of Sing(X) ∩ ℓ.
A point P on a separable line ℓ is a point of ramification n l if the corresponding point on L has ramification index n and length(Ω L/P 1 ) = l.If char K does not divide n, then l = n − 1 and can be omitted, whereas if char K divides n, then l ≥ n.We say that ℓ has ramification (n l ) r (n ′ l ′ ) r ′ . . .if it has r points of ramification n l and so on.Lemma 2.1 ([18, Proposition 1.5]).If ℓ ⊂ X has degree d and singularity s, then d ≤ 3 − s, and d = 3 if and only if s = 0. Lemma 2.2 ([18, Lemma 2.7]).If P ∈ Sing(X), then there are at most 8 lines on X passing through P .

2.2.
Valency.Let Π ⊂ P 3 be a plane such that X ∩ Π splits into four lines ℓ 1 , . . ., ℓ 4 (not necessarily distinct).If a line ℓ ′ ⊂ X not lying on Π meets two or more distinct lines ℓ i , then their point of intersection must be a singular point of the surface.It follows that |Fn X| is bounded by #{lines not in Π meeting ℓ i in a smooth point}.
It holds A line ℓ is said to be of type (p, q) if there exist exactly p points t ∈ P 1 such that c t splits into three (not necessarily distinct) lines and exactly q points t ∈ P 1 such that c t splits into a line and an irreducible conic.It follows from (3) that if ℓ has type (p, q) and degree d, then (4) v(ℓ) ≤ dp + q.
2.3.Lines of the first and second kind.Let x 0 , x 1 , x 2 , x 3 be the coordinates of P 3 .We denote by K[x 0 , . . ., x 3 ] (d) the space of forms (ie homogeneous polynomials) of degree d.Up to projective equivalence, we can suppose that the line ℓ is given by x 0 = x 1 = 0, so that the quartic X is defined by (5) X : where i 0 , . . ., i 4 are non-negative integers.
The planes containing ℓ are parametrized by Π t : x 0 = tx 1 , where t = ∞ denotes the plane x 1 = 0.The residual cubic c t is defined by the equation of Π t and the equation g ∈ K[t][x 1 , x 2 , x 3 ] (3) obtained by substituting x 0 with tx 1 in (5) and factoring out x 1 .An explicit computation shows that there exist two forms α, β ∈ K[x 2 , x 3 ] (deg(ℓ)) such that ℓ ∩ c t consists of the points [0 : 0 : An inflection point of a (possibly reducible) cubic curve c ⊂ P 2 is a point of c which is a zero of the hessian of the cubic.If c contains a line as a component, all the points of the line are inflection points of c.The following lemma can be checked by an explicit computation, too.
Lemma 2.4.If c ⊂ P 2 is a reducible cubic that is the union of an irreducible conic and a line m, then the locus of inflection points of c is exactly m.
If the surface X is defined as in (5), then the hessian of the equation g defining the residual cubic c t restricted to the line ℓ is given by ( 7) which is a polynomial of degree 5 in t, with forms of degree 3 in (x 2 , x 3 ) as coefficients (if char K = 2, this definition has to be slightly modified, see [8]).
The resultant R(ℓ) with respect to the variable t of the polynomials ( 6) and ( 7) is called the resultant of the line ℓ.We say that a line ℓ of positive degree is a line of the second kind if its resultant is identically equal to zero.Otherwise, we say that ℓ is a line of the first kind.
A root [x 2 : x 3 ] of R(ℓ) corresponds to a point P = [0 : 0 : x 2 : x 3 ] on ℓ.If P is a smooth surface point, then it is an inflection point for the residual cubic passing through it.

2.4.
Triangle-free surfaces.We follow here the nomenclature of [18, §5].A Dynkin diagram (resp.extended Dynkin diagram) is also called an elliptic graph (resp.parabolic graph).The line graph Γ(X) of X is the dual graph of the strict transforms of its lines on Z.A triangle on X is the union of three lines intersecting pairwise in smooth points of X.We say that X is triangle-free if it contains no triangles.

Lemma 2.7 ([17, Lemma 2.23]).
A triangle on X is contained in a plane Π ⊂ P 3 such that the intersection of Π and X is one of the configurations pictured in Figure 1.
From now on, we suppose that char K = 3.We adopt Kodaira's notation for the possible types of a fiber F appearing in a genus 1 fibration.We denote by e(F ) and δ(F ) respectively the Euler-Poincaré characteristic and the wild ramification index in characteristic 3 of F .The possible values of e(F ) and δ(F ) according to the type of F are displayed in Table 1.
The following formula holds for an elliptic fibration on a K3 surface. ( Possible configurations of lines on a plane with a triangle.Singular points are marked white.In configurations D 0 and E 0 the singular points might coincide. 2.6.Quasi-elliptic fibrations and cuspidal lines.A fiber F of a quasi-elliptic fibration can only have one of the following types: II, IV, IV * , II * .A K3 surface with a quasi-elliptic fibration is necessarily supersingular and the following formula holds (9) If iv, iv * and ii * denote the numbers of reducible fibers of the respective type, then (9) can be also written (10) iv + 3 iv * + 4 ii * = 10.
The closure of the locus of cusps is a smooth curve K, called cuspidal curve, such that K • F = 3.The restriction of the fibration to K is an inseparable morphism of degree 3. The cuspidal curve meets a reducible fiber in one of the ways picture in Figure 2. In Table 1.Euler-Poincaré characteristic and possible wild ramification indices of a fiber F in a genus 1 fibration.
Cuspidal curve (white vertex) intersecting a reducible fiber of a quasi-elliptic fibration.Multiple white dots represent different possibilities.
particular, the way K intersects F is uniquely determined unless F is of type II * .Note that K intersects a fiber of type IV at the intersection point of the three components.
A line ℓ ⊂ X is said to be cuspidal if it is quasi-elliptic and the cuspidal curve K of the induced fibration coincides with the strict transform L of ℓ.A cuspidal line is necessarily inseparable.

Bounds on the valency
From now on we assume that X is a K3 quartic surface defined over a field K of characteristic 3.This section is devoted to the study of the possible valencies of a line ℓ ⊂ X (Definition 2.3).

Separable elliptic lines.
The results which we will prove presently are summarized in Table 2.
Lemma 3.1.Let ℓ ⊂ X be a separable line of the second kind and P ∈ ℓ a smooth point of ramification 2.Then, either the corresponding fiber is of type II with a cusp in P , or the corresponding residual cubic splits into a double line and a simple line.
Table 2. Bounds for the valency of a separable elliptic line.kind degree singularity valency first kind Proof.Note that only lines of degree 3 and 2 can have a point P of ramification 2. We choose coordinates so that P is given by [0 : 0 : 0 : 1].This means that a 0103 = 0 and a 0112 = 0.
Since P is of ramification index 2 and it is non-singular, we can normalize Since ℓ is of the second kind, the following relations must be satisfied: 0211 and a 0310 = a 0211 a 0220 .This means that the residual cubic in x 0 = 0 corresponding to P is given by where f 3 is a form of degree 3. Looking at the explicit formula for f 3 (which we do not write here), we see that either this cubic is irreducible and gives rise to a fiber of type II, or the polynomial g = a 0211 x 1 − x 2 divides f 3 ; in the latter case it is immediate to compute that also g 2 divides f 3 .Lemma 3.2.Let ℓ ⊂ X be a separable line of the second kind and P ∈ ℓ a point of ramification 3 4 .Then, either the corresponding fiber is of type II with a cusp in P , or the corresponding residual cubic splits into three lines (not necessarily distinct) intersecting in P .
A calculation with local parameters shows that length(Ω L/P 1 ) = 4 if and only if a 1012 = 0.Moreover, the following three coefficients must be different from 0: a 0130 , a 1003 and a 1021 ; the first two because otherwise there would be singular points on ℓ (implying that the degree of ℓ is less than 3), the third because otherwise ℓ would be inseparable.We can normalize them to 1, rescaling coordinates.Two necessary conditions for the line ℓ to be of the second kind are a 0202 = 0 and a 0211 = 0.
Hence, the residual cubic in Π 0 is given by . Either the fiber is irreducible and has a cusp in P (a 0301 = 0), or it splits into three concurrent lines (a 0301 = 0).Remark 3.3.Suppose that ℓ is an elliptic line of degree 3. From Table 1 we observe since a reducible fiber has e(F t ) ≥ 2 and a fiber with at least three components has e(F t ) ≥ 3. From equations (3) and ( 8) we infer that (11) v(ℓ) = Table 3. Bounds for the valency of a separable elliptic line of the second kind of degree 3.
ramification valency for some N ≥ 0, then all fibers F s must be of type I 3 and, in particular, N must be divisible by 3.
An application of the Riemann-Hurwitz formula yields the following lemma.
Lemma 3.4.If ℓ ⊂ X is a separable line of degree 3, then ℓ has ramification 2 4 1 , 2 1 3 3 or 3 4 .Proposition 3.5.If ℓ ⊂ X is a separable elliptic line of the second kind of degree 3, then the valency of ℓ is bounded according to Table 3.
Proof.Suppose first that ℓ has a point P of ramification index 2.According to Lemma 3.1, the corresponding fiber F P is either of type II, so that e(F P ) + δ P ≥ 2 + 1 = 3 (Table 1) and v(F P ) = 0, or it contains a double component, so that e(F P ) ≥ 6 and v(F P ) = 2; in any case, the difference e(F P ) + δ P − v(F P ) is always at least 3. Therefore, if there are 4 points of ramification 2, then by formula (11) v(ℓ) is not greater than 24 − 4 • 3 = 12, while if there is just one, v(ℓ) is not greater than 24 − 3 = 21.
Suppose now that ℓ has no point of ramification index 2, ie ℓ has ramification 3 4 .If the ramified fiber F 0 is of type II, then there can be at most 24 − 3 = 21 lines meeting ℓ.Otherwise, F 0 splits into three concurrent lines by Lemma 3.2.Then, e(F 0 ) + δ 0 ≥ 5 (type IV has wild ramification, too), which means that the contribution to v(ℓ) of the other fibers is not greater than 24−5 = 19.Nonetheless, by Remark 3.3 this contribution cannot be exactly 19, since 19 is not divisible by 3. Hence, again, we can have at most 3 + 18 = 21 lines intersecting ℓ.
If the residual cubic corresponding F Q splits into a double line and a simple line, then it contributes 1 to the valency and at least 6 to the Euler-Poincaré characteristic.Applying formula (4) again yields v(ℓ) ≤ 13.
Example 3.9.The following surface contains an elliptic line ℓ : x 0 = x 1 = 0 of degree 2 with valency 14, thus attaining the bound in Proposition 3.8.The surface contains one point P = [0 : 0 : 0 : 1] of type A 1 .The line ℓ has 7 fibers of type I 3 and one ramified fiber of type II.The other ramified fiber corresponds to the plane x 0 = 0 and is smooth.Proof.The proof can be carried over word by word from the characteristic 0 case (see [18, Propositions 2.13 and 2.14]).

3.2.
Quasi-elliptic lines of degree 3. Table 4 summarizes the known bounds for the valency of a quasi-elliptic line, which we are about to prove.Lemma 3.11.If ℓ ⊂ X is an inseparable line with v(ℓ) > 12, then ℓ is cuspidal (in particular, quasi-elliptic).
Proof.Up to coordinate change, we can suppose that the residual cubic contained in x 0 = tx 1 intersects the line ℓ : x 0 = x 1 = 0 in [0 : 0 : 0 : 1] for t = 0 and in [0 : 0 : 1 : 0] for t = ∞.This means that the following coefficients vanish: a 0103 , a 0112 , a 0121 ; a 1012 , a 1021 , a 1030 .Moreover, a 1003 and a 0130 must be different from 0, and can be normalized to 1 and −1, respectively, by rescaling coordinates.Up to a Frobenius change of parameter t = s 3 , we can explicitly write the intersection point P s of the residual cubic with ℓ, which is given by P s = [0 : 0 : s : 1].
If a residual cubic c s is reducible, then all components must pass through P s ; in particular, P s must be a singular point of c s .One can see explicitly that P s is a singular point of c s if and only if s is a root of the following degree 8 polynomial: Furthermore, it can be checked by a local computation that if c s splits into three (not necessarily distinct) lines, then s is a double root of g.This implies that the valency of ℓ is not greater than 3 • 8/2 = 12, unless the polynomial g vanishes identically, but g ≡ 0 implies that all points P s are singular for c s , ie the line ℓ is cuspidal.Corollary 3.12.If ℓ ⊂ X is a cuspidal line, then X is projectively equivalent to a member of the family C defined by C : x 0 x 3 3 − x 1 x 3 2 + x 2 q 3 (x 0 , x 1 ) + x 3 q ′ 3 (x 0 , x 1 ) + q 4 (x 0 , x 1 ), where q 3 , q ′ 3 and q 4 are forms of degree 3, 3 and 4, respectively.Proof.The family can be found imposing that the polynomial g in the proof of Lemma 3.11 vanishes identically.Corollary 3.13.If ℓ ⊂ X is a cuspidal line, then a residual cubic corresponding to a reducible fiber of ℓ is either the union of three distinct concurrent lines, or a triple line.
Proof.The intersection of a residual cubic with ℓ is always one single point.
A residual cubic of ℓ cannot be the union of a line and an irreducible conic, because the line and the conic would result in a fiber of type I n (because the conic has to be tangent to ℓ), which is impossible in a quasi-elliptic fibration.
Therefore, a residual cubic relative to a reducible fiber must split into three (not necessarily distinct) lines.If at least two of them coincide, the plane on which they lie contains at least a singular point P of the surface (which is not on ℓ, since ℓ has degree 3).An explicit inspection of this configuration in the family C (for instance, supposing up to change of coordinates that P is given by [0 : 1 : 0 : 0]) shows that the residual cubic degenerates to a triple line.Lemma 3.14.If ℓ ⊂ X is a quasi-elliptic line of degree 3, then v(ℓ) ≤ 30.
Proof.The fibration induced by the line ℓ has at most 10 reducible fibers, each of which can contribute at most 3 to its valency.
Remark 3.15.The bound of Lemma 3.14 is sharp.As soon as a K3 quartic surface X is smooth, the valency of a quasi-elliptic line of degree 3 on X is automatically 30, because the fibration induced by ℓ can only have 10 reducible fibers of type IV, whose residual cubics are the union of three concurrent lines.Notably, this happens for all 112 lines of the Fermat surface.Lemma 3.16.Let ℓ ⊂ X be any line of degree 3. A fiber of type IV of the fibration induced by ℓ corresponds to one of the residual cubics pictured in Figure 3, with the only restriction that ℓ cannot pass through a singular point.
Proof.A fiber of type IV contains three simple components; hence, the corresponding residual cubic can also have only simple components.Since it cannot contain cycles, it must be one of the following, as in Figure 3: a cusp, a conic and a line meeting tangentially in one point, or three distinct lines meeting in one point.The remaining components must come from the resolution of the singular points on the surface.The types of the singular points can be immediately deduced from the respective Dynkin diagrams.
Lemma 3.17.Let ℓ ⊂ X be any line of degree 3. A fiber of type IV * of the fibration induced by ℓ corresponds to one of the residual cubics pictured in Figure 4, with the only restriction that ℓ cannot pass through a singular point.
Proof.Besides the residual cubics with only simple components described in the previous lemma, we can also have multiple components, namely a double line and a simple line, or a triple line.
In the former case, the strict transforms of the lines can intersect (if their intersection point is smooth) or not (if their intersection point is singular), giving rise to two different configurations, which we distinguish by the letters a and b.In the latter case, there is no ambiguity, since a fiber of type IV * contains only one triple component.
We denote by iv 0 , iv 1 ,. . . the number of fibers of type IV 0 , IV 1 , and so on.Note that the subscript indicates the local valency of the fiber.Lemma 3.18.If ℓ ⊂ X is a separable quasi-elliptic line of degree 3, then the degree of its cuspidal curve is at least 3 and at most 7.
The cuspidal curve K and the line L are distinct because ℓ is separable.The curve K can meet L only in points of ramification; moreover, a local computation shows that if K is tangent to L, then ramification 3 4 occurs, and that higher order tangency cannot happen.We thus obtain the following bounds according to the ramification type of ℓ: K • L ≤ 4 for type 2 4  1 , In particular, if ii * > 0, then v(ℓ) ≤ 20, so we can suppose that ℓ has no II * -fibers.Similarly, we can suppose that ℓ has at most one fiber of type IV * .
If ℓ has no IV * -fiber, then it must have 10 fibers of type IV.Using the classification of Lemma 3.16, we list the possible configurations with v(ℓ) > 21 (16 cases) in Table 5.
For each case, we consider the lattice generated by L, a general fiber F , the fiber components of the reducible fibers and the cuspidal curve K (which must be different from L, since ℓ is separable).All intersection numbers are univocally determined (L • F = 3 because ℓ has degree 3), except for but k can only take up the values 3, . . ., 7 on account of Lemma 3.18.We check that this lattice has rank bigger than 22 in all cases, except for case 6 with k = 3 (ie K • L = 0).
On the other hand, this case does not exist.In fact, suppose that ℓ is as in case 6 with K • L = 0; in particular, ℓ has no ramified fibers with multiple components and, since v(ℓ) = 24, ℓ is of the second kind.It follows that if ℓ has a point of ramification 2, then by Lemma 3.1, the ramified fiber must be a cusp, ie K intersects L so K • L > 0; if ℓ has ramification 3 4 , then by Lemma 3.2 the ramified fiber must be either a cusp or the union of three distinct lines; in both cases, K • L > 0. Example 3.20.The following surface contains a separable quasi-elliptic line ℓ : x 0 = x 1 = 0 of degree 3 with valency 21, thus attaining the bound of Proposition 3.19: . It contains only one singular point [1 : 0 : 0 : 0] of type E 6 .
Proof.By Proposition 2.5, we can assume that ℓ is of the second kind.Let P be the singular point on ℓ and suppose t ∈ P 1 is such that v t (ℓ) > 0. The cubic c = c t is reducible, because it contains at least a line.Suppose that c splits into a line m and an irreducible conic q (which must be tangent to each other because fibers of type I n are not admitted in a quasi-elliptic fibration).Since v t (ℓ) > 0, the line m meets ℓ in a smooth point; hence, the three configurations in Figure 5  The third configuration gives rise to a fiber of type III, but this type does not appear in a quasi-elliptic fibration.(In the other configurations, the point of intersection of m with the residual conic is singular.)In the first configuration, if q ∩ ℓ = {P, Q}, then Q is not an inflection point of the residual cubic by Lemma 2.4, contradicting the fact that ℓ is of the second kind.The second configuration can be ruled out by an explicit parametrization (in a line of the second kind, either the point P is a ramification point, or the cubic passing twice through P is singular at P ).
Thus, all three configurations are impossible, so c must split into three (not necessarily distinct) lines and at least one of them should pass through P .Since there can be at most 8 lines through a singular point (Lemma 2.2), there can be at most 7 such reducible fibers, each of them contributing at most 2 to the valency of ℓ, whence v(ℓ) ≤ 14.
Example 3.22.The bound given by Proposition 3.21 is sharp.In fact, the following quartic surface contains a quasi-elliptic line ℓ : x 0 = x 1 = 0 of degree 2 and valency 14: The quartic contains two singular points, P = [0 : 0 : 0 : 1] of type A 1 and Q = [−1 : 1 : 1 : 0] of type E 6 .The fibration induced by the line ℓ has 7 fibers of type IV and one fiber of type IV * corresponding to the plane containing Q.
Proof.The fibration induced by the line ℓ has at most 10 reducible fibers, each of which contributes at most 1 to its valency.

Proof of Theorem 1.1
In order to prove the main theorem, we divide our arguments according to whether the K3 quartic surface contains a star ( §4.1) or more generally a triangle ( §4.2), or contains no triangles at all ( §4.3).Throughout the section, X denotes a K3 quartic surface defined over a field of characteristic 3.

4.1.
Star case.We will start by analyzing quartic surfaces containing a star, which, we recall, is the union of four distinct lines meeting in a smooth point.
Since the four lines in a star are necessarily coplanar, a star is the same as a configuration C 0 (see Figure 1).The lines have necessarily degree 3 because there are no singular points on the plane containing them.
We will first need a series of lemmas.In all of them, we will parametrize the surface X as in (5) in such a way that the star is contained in the plane x 0 = 0 and the lines meet at [0 : 0 : 0 : 1], ie setting the following coefficients equal to 0: a 0301 , a 0211 , a 0121 , a 0202 , a 0112 , a 0103 .
If necessary, we will parametrize a second line in the star ℓ ′ as x 0 = x 2 = 0, by further assuming a 0400 = 0. Proof.It can be checked by an explicit computation that the resultant of ℓ has a root of order 6 at the center of the star; this implies that there are at most 18 − 6 = 12 lines meeting ℓ not contained in the star.Lemma 4.2.If ℓ ⊂ X is a separable line of ramification 2 1 3 3 contained in a star, then it is of the first kind.
Proof.By a change of coordinates, we can assume that the point of ramification index 2 is [0 : 0 : 1 : 0], and that ramification occurs at x 1 = 0. Imposing that ℓ is of the second kind leads to a contradiction (ℓ cannot be separable).Proof.Beside ℓ : x 0 = x 1 = 0 and ℓ ′ : x 0 = x 2 = 0, we can suppose without loss of generality that a third line is given by ℓ ′′ : x 0 = x 1 + x 2 = 0, setting a 0220 = a 0130 + a 0310 .The conditions for ℓ, ℓ ′ or ℓ ′′ to be of ramification 3 4 are a 1012 = 0, a 1102 = 0 and a 1012 = a 1102 , respectively.Two of them imply the third one.
Lemma 4.4.If three lines in a star contained in X are separable, then at most two of them can be of the second kind.
Proof.We parametrize ℓ, ℓ ′ and ℓ ′′ as in the previous Lemma.Imposing that all three of them are of the second kind leads to a contradiction (at least one of them must be inseparable).
Lemma 4.5.Let ℓ, ℓ ′ ⊂ X be two lines in a star.If ℓ is a separable line of the second kind, and ℓ ′ is a line of the first kind of ramification 3 4 , then v(ℓ ′ ) ≤ 12.
Proof.This can be checked again by an explicit computation of the resultant of ℓ ′ , which has now a root of order 9 at the center of the star.Lemma 4.6.Let ℓ, ℓ ′ ⊂ X be two lines in a star.If ℓ is a cuspidal line, and ℓ ′ is not cuspidal, then v(ℓ ′ ) ≤ 12.
Proof.We parametrize ℓ : x 0 = x 1 = 0 and ℓ ′ : x 0 = x 2 = 0 as in Corollary 3.12, ie we suppose that X is given by the family C where the following coefficients are set to zero: a 0400 , a 0301 ; a 1201 , a 1300 ; a 2200 , a 2101 , a 1210 .
By a further rescaling we put a 0310 = 1 and we consider ℓ ′′ : x 0 = x 1 − x 2 = 0.The line ℓ ′′ is inseparable and we can compute its polynomial g as in formula (12) in the proof of Lemma 3.11 (by parametrizing the pencil with x 0 = s 3 (x 1 − x 2 )), which turns out to be g(s) = a 2110 s 8 .
This means that ℓ ′′ has only one singular fiber in s = 0 (namely a fiber of type IV with the maximum possible index of wild ramification), unless a 2110 = 0 and g ≡ 0, in which case ℓ ′′ is cuspidal.
Lemma 4.8.If ℓ ⊂ X is a cuspidal line which is not contained in at least two stars, then v(ℓ) ≤ 6.
Proof.On account of Corollary 3.13, the number of stars in which ℓ is contained is exactly equal to the number of fibers of type IV in its fibration; moreover, v t (ℓ) = 1 if t ∈ P 1 corresponds to a fiber of type IV * or II * , yielding Recalling formula (10), we deduce that if iv < 2 then v(ℓ) ≤ 6.
Proposition 4.9.If X contains a star and is not projectively equivalent to the Fermat surface, then |Fn X| ≤ 58.
Proof.If ℓ 1 , ℓ 2 , ℓ 3 , ℓ 4 are the lines contained in the star, it holds from ( 2) Suppose first that all lines ℓ i are not cuspidal.

Triangle case.
In this section we study the case in which X admits a triangle.The three lines forming the triangle need to be coplanar, and we denote by Π the plane on which they lie.The plane Π intersects X also in a fourth line, which might coincide with one of the first three.Proof.Let ℓ 0 be the double line in the plane Π containing one of the two configurations, and let ℓ 1 and ℓ 2 be the two simple lines.Lines meeting ℓ 0 different from ℓ 1 and ℓ 2 must pass through the singular points; hence, by Lemma 2.2 there can be at most 3•(8−1) = 21 of them.
Note that ℓ 1 and ℓ 2 cannot be cuspidal because of Corollary 3.13.
Suppose that ℓ 1 is quasi-elliptic.The plane Π corresponds to a fiber of type IV * or II * ; hence, there can be only one singular point on ℓ 0 : in fact, by inspection of the Dynkin diagrams, a component of multiplicity 2 in these fiber types meets at most 2 other components (and one of them is the strict transform of ℓ 2 ).The lines ℓ 1 and ℓ 2 not being cuspidal, we know that they have valency at most 21.It follows from ( 2 Proof.Let Π be the plane containing ℓ and ℓ ′ .Both lines are separable, since otherwise the respective residual cubics would intersect them in one point.We suppose that also v(ℓ ′ ) > 18 and look for a contradiction.
Two necessary condition for ℓ and ℓ ′ to be lines of the second kind are a 0310 = a 2 0211 and a 0130 = a 2 0121 .This means that the residual conic in Π : x 0 = 0 is given explicitly by (13) a 2 0211 x 2 1 + a 2 0121 x 2 2 + a 0220 x 1 x 2 + a 0211 x 1 x 3 + a 0121 x 2 x 3 + x 2 3 = 0. Proof.The subgraph D induces a genus 1 fibration ([6, §3, Theorem 1]), which can be elliptic or quasi-elliptic.The vertices in D ∪ (Γ span D) are fiber components of this fibration.If the fibration is elliptic, there cannot be more than 24 components, on account of the Euler-Poincaré characteristic.If the fibration is quasi-elliptic, we obtain from formula (10) that ( 14) iv * + ii * ≤ 3.
A fiber of type IV can contain at most 2 lines, since there are no triangles.Hence, from (10) and ( 14 Proof.Thanks to Lemma 2.6, we can suppose that ℓ is quasi-elliptic.We claim that for all t ∈ P it holds (15) Indeed, this is true if the residual cubic c t is irreducible (v t (ℓ) = 0 and e(F t ) ≥ 2), or if c t is the union of a line and a reducible conic (v t (ℓ) ≤ 1 and e(F t ) ≥ 4 because F t is reducible).If c t splits into three lines (not necessarily distinct), then these lines meet in one singular point because X is triangle-free and fibers of type I n do not appear in a quasi-elliptic fibration.Hence, F t is of type IV * or II * and (15) holds because v t (ℓ) ≤ 3 and e(F t ) ≥ 8.
From (3), ( 9) and ( 15) it follows that Proof.We can adapt the proof in [18, Proposition 5.9].Indeed, the proof of [18, Proposition 5.8] is also valid in characteristic 3, since we merely employ the analogous result of Proposition 4.13 and the fact that the Picard lattice of X has rank at most 22.Therefore, if X contains no quadrangle, then |Fn X| ≤ 54.
If X contains a quadrangle, then we can apply Lemma 4.14 to the Ã3 -subgraph corresponding to the quadrangle, finding |Fn X| ≤ 4 • (12 − 2) + 24 = 64.4.4.Proof of Theorem 1.1.The case in which X contains a star is treated in Proposition 4.9, the case in which X contains no stars, but contains a triangle is treated in Proposition 4.12, and the case in which X contains no triangles is treated in Proposition 4.15, so the proof is now complete.

3 Figure 3 .
Figure 3. Possible residual cubics corresponding to a fiber of type IV.

Figure 4 .
Figure 4. Possible residual cubics corresponding to a fiber of type IV * .

Figure 5 .
Figure 5. Possible residual cubics appearing in the proof of Proposition 3.21.

Lemma 4 . 1 .
If ℓ ⊂ X is a line of the first kind in a star, then v(ℓ) ≤ 15.

Lemma 4 . 3 .
If three lines in a star contained in X are separable and at least two of them have ramification 3 4 , then the third one also has ramification 3 4 .

Table 4 .
Bounds for the valency of a quasi-elliptic line.Proof.By Proposition 2.5, we can assume that ℓ is of the second kind.Since ℓ has degree 2, it must have singularity 1 (Lemma 2.1).Let P be the singular point on ℓ.The morphism π : L → P 1 , being of degree 2, is separable and has two points of ramification index 2.At least one of the point of ramification must be different from P .Let us call it Q.By Lemma 3.1 either the fiber corresponding to Q is of type II or the residual cubic splits into a double line and a simple line.