betaFeedback

WorksMathematics

Terao's freeness conjecture for arrangements of 15 and 16 lines

Abstract

We prove that Terao's freeness conjecture holds for arrangements of 15 and of 16 lines in the complex projective plane. Together with the theorem of Barakat and Kühne for at most 14 hyperplanes, this shows that freeness of a central arrangement of at most 16 planes in C3\mathbb{C}^3 is determined by its intersection lattice. The proof has a theoretical and a computational part. Combining published combinatorial criteria for freeness (due to Abe, Dimca, Dimca-Kühne-Pokora and Yoshinaga) with Bojanowski's Hirzebruch-type inequality and an elementary counting identity, we show that a counterexample would have exponents (6,8)(6,8) or (7,7)(7,7) for 15 lines and (7,8)(7,8) for 16 lines, and an intersection lattice subject to strong numerical constraints. We then enumerate all 22,544,280 simple rank 3 matroids satisfying these constraints, up to isomorphism, and determine the complex realization space of each of them exactly over Q\mathbb{Q}. Exactly eleven of them are realizable over C\mathbb{C}, and on each of these the freeness of the realizations is constant. As a by-product we classify the free arrangements of 15 and 16 lines that are not divisionally free, have maximal multiplicity smaller than the smaller nontrivial exponent, and have at least four intersection points on every line.

Introduction

Let A\mathcal{A} be a central arrangement of hyperplanes in V=CℓV = \mathbb{C}^{\ell}, let S=C[x1,…,xℓ]S = \mathbb{C}[x_{1}, \ldots, x_{\ell}], and let

Q=∏H∈AαHQ = \prod_{H \in\mathcal{A}} \alpha_{H}

be a defining polynomial. The module of logarithmic derivations

D(A)={θ∈Der⁡C(S):θ(Q)∈QS}D(\mathcal{A}) = \{\theta\in\operatorname{Der}_{\mathbb{C}}(S) : \theta(Q) \in QS\}

is a reflexive graded SS-module of rank ℓ\ell, and A\mathcal{A} is called free if D(A)D(\mathcal{A}) is a free SS-module. In that case the degrees d1,…,dℓd_{1}, \ldots, d_{\ell} of a homogeneous basis are the exponents of A\mathcal{A}. Terao’s factorization theorem [23] states that the characteristic polynomial then factors as χ(A,t)=∏i(t−di)\chi(\mathcal{A},t) = \prod_{i}(t-d_{i}); in particular the exponents of a free arrangement are determined by its intersection lattice L(A)L(\mathcal{A}). The converse question is the central open problem of the subject.

Conjecture 1.1 (Terao, see [19 Conj. 4.138]). Let A\mathcal{A} and B\mathcal{B} be arrangements over the same field with isomorphic intersection lattices. If A\mathcal{A} is free, then B\mathcal{B} is free.

The conjecture is open already for ℓ=3\ell= 3, i.e. for line arrangements in the projective plane, which is the case considered here. Over C\mathbb{C} it has been verified for arrangements of at most 12 lines by Faenzi and Vallès [FV14], of 13 lines by Dimca, Ibadula and Măcinic [10], and of 14 lines, over fields of arbitrary characteristic, by Barakat, Behrends, Jefferson, Kühne and Leuner [4]. Barakat and Kühne [5] proved it for all rank 3 arrangements with at most 14 hyperplanes, again over fields of arbitrary characteristic. Their proof enumerates the relevant simple rank 3 matroids with at most 14 elements and computes their moduli spaces, in particular their non-free loci. That approach becomes impractical for larger numbers of hyperplanes, since the number of simple rank 3 matroids grows very fast with the number of elements [16].

Our main result is the following.

Main Theorem. Let d∈{15,16}d \in\{15,16\}, and let A\mathcal{A} and B\mathcal{B} be arrangements of dd lines in P2(C)\mathbb{P}^{2}(\mathbb{C}) with isomorphic intersection lattices. If A\mathcal{A} is free, then B\mathcal{B} is free.

Corollary 1.2. Terao’s conjecture holds for all central arrangements of at most 16 planes in C3\mathbb{C}^{3}.

Proof. Arrangements of rank at most two are free. For rank three, combine the Main Theorem with [5 Thm. 1.2]. □\square

Strategy

Instead of examining all matroids with 15 or 16 elements, we first use theory to confine a potential counterexample to a very small region. Section 33 combines the following ingredients:

  • results of Dimca [8] and Dimca–Kühne–Pokora [11], which force m≤d1−1m \le d_{1}-1, where mm is the maximal multiplicity of an intersection point and (d1,d2)(d_{1},d_{2}), d1≤d2d_{1} \le d_{2}, are the nontrivial exponents of the free arrangement;

  • Abe’s restriction on the number of points on a line of a free arrangement [1] together with his division theorem [2], which force every line to contain at most d1d_{1} intersection points;

  • Yoshinaga’s criterion [25], together with the 3-transitivity of PGL2\mathrm{PGL}_{2} on P1\mathbb{P}^{1} (compare Wakamiko’s explicit formula [24]), which forces every line to contain at least four intersection points;

  • an elementary identity, which we call the budget identity (Lemma 3.23.2), and Bojanowski’s Hirzebruch-type inequality [15, 21].

The outcome (Proposition 3.83.8) is that a counterexample with d∈{15,16}d \in\{15,16\} lines is of type (d;d1,d2)∈{(15;6,8),(15;7,7),(16;7,8)}(d;d_{1},d_{2}) \in\{(15;6,8),(15;7,7),(16;7,8)\} and an intersection lattice satisfying five explicit numerical conditions (R1)–(R5).

In Section 44 we enumerate all simple rank 3 matroids satisfying (R1)–(R5), up to isomorphism: there are 3, 6 091 145 and 16 453 132 of them, respectively. In Section 55 we compute the complex realization space of each of them exactly over Q\mathbb{Q}, and we decide freeness on each Galois orbit of realizations by certificates that are verified in exact arithmetic. The results are in Section 66. Only 11 of the 22 544 280 matroids are realizable over C\mathbb{C}: one of type (15;6,8)(15;6,8), ten of type (15;7,7)(15;7,7), and none of type (16;7,8)(16;7,8). Each of them is either free for all realizations or non-free for all realizations. This proves the Main Theorem. As a by-product we obtain:

Theorem 1.3. Let A\mathcal{A} be a free arrangement of d∈{15,16}d \in\{15,16\} lines in P2(C)\mathbb{P}^{2}(\mathbb{C}) with exponents (1,d1,d2)(1,d_{1},d_{2}), d1≤d2d_{1} \le d_{2}, and maximal multiplicity mm. Suppose that m≤d1−1m \le d_{1}-1, that every line of A\mathcal{A} contains at least four intersection points, and that A\mathcal{A} is not divisionally free.

(a) If d=16d=16, no such arrangement exists.

(b) If d=15d=15, then L(A)L(\mathcal{A}) is isomorphic to one of the six lattices L0,F1,…,F5L_{0},F_{1},\ldots,F_{5} of Table 22. They are the lattices of the reflection arrangement A(5,5,3)A(5,5,3) (L0L_{0}), of Cuntz’s one-parameter family [3 §7] (F1F_{1}), of two sub-arrangements of the reflection arrangement of the complex reflection group G26G_{26} (F3,F5F_{3},F_{5}), and of rigid arrangements defined over Q\mathbb{Q} (F2F_{2}) and over Q(5)\mathbb{Q}(\sqrt{5}) (F4F_{4}).

Conversely, every complex arrangement whose lattice is one of L0,F1,…,F5L_{0},F_{1},\ldots,F_{5} is free and satisfies the hypotheses above.

Here A\mathcal{A} is divisionally free if χ(AH,t)\chi(\mathcal{A}^{H},t) divides χ(A,t)\chi(\mathcal{A},t) for some line HH (for ℓ=3\ell=3 this agrees with Abe’s definition [2]). Divisional freeness, the condition m≥d1m \ge d_1, and the existence of a line with at most three points are three standard situations in which freeness is known to be combinatorial (§ 3). Theorem 1.3 describes the free arrangements of 15 and 16 lines not covered by them. Abe [1] gives further combinatorial criteria. For instance, freeness is combinatorial when d1≤5d_1 \le5 [1 Cor. 5.5] or d2≥2d1−3d_2 \ge2d_1-3 [1 Cor. 5.4]. These already exclude the five types that we remove in Table 1 by a different argument. The six lattices of Theorem 1.3 are not covered by these criteria, nor by [1 Thm. 2.8(2)] (see Remark 3.6).

On the computer-assisted part

The code, the results and a script that reruns the enumeration and the screening are available as ancillary files (§ 7). The programs implement the specification justified in § 3–§ 5.

Preliminaries

Line arrangements and their lattices

Throughout, A\mathcal{A} is an arrangement of dd distinct lines in P2=P2(C)\mathbb{P}^{2}=\mathbb{P}^{2}(\mathbb{C}), or equivalently a central arrangement of dd planes in C3\mathbb{C}^{3}. A pencil (all lines through one point) is always free, so we may and do assume that the lines are not all concurrent. For an intersection point PP let mPm_P be the number of lines through PP. Let trt_r be the number of points of multiplicity rr, and let m=max⁡PmPm=\max_P m_P. For a line HH let rH=∣AH∣r_H=\lvert\mathcal{A}^{H}\rvert be the number of intersection points on HH, and put

e(H)=∑P∈H(mP−2).e(H)=\sum_{P\in H}(m_P-2).

Since every other line meets HH in exactly one point, ∑P∈H(mP−1)=d−1\sum_{P\in H}(m_P-1)=d-1, and hence

e(H)=d−1−rH.(1)e(H)=d-1-r_H. \tag*{(1)}

The intersection lattice L(A)L(\mathcal{A}) is determined by the simple rank 3 matroid on the set of lines whose rank 2 flats are the sets of lines through the intersection points. We encode it by the family of its blocks, the sets of lines through the points with mP≥3m_P\ge3; any two blocks share at most one line. We identify L(A)L(\mathcal{A}) with this matroid and use the words lattice and matroid interchangeably; in the language of incidence geometry, such a matroid is a linear space. All quantities above depend only on the lattice. The characteristic polynomial is χ(A,t)=(t−1)χ0(A,t)\chi(\mathcal{A},t)=(t-1)\chi_0(\mathcal{A},t) with

χ0(A,t)=t2−(d−1)t+b,b=∑P(mP−1)−(d−1).(2)\chi_0(\mathcal{A},t)=t^2-(d-1)t+b,\qquad b=\sum_P(m_P-1)-(d-1). \tag*{(2)}

Freeness and exponents

If A\mathcal{A} is free, its exponents are (1,d1,d2)(1,d_1,d_2) with d1+d2=d−1d_1+d_2=d-1 and d1d2=bd_1d_2=b [23]. We always order them so that d1≤d2d_1\le d_2, and we say that L(A)L(\mathcal{A}) is of type (d;d1,d2)(d;d_1,d_2) whenever χ0(A,t)=(t−d1)(t−d2)\chi_0(\mathcal{A},t)=(t-d_1)(t-d_2) with integers d1≤d2d_1\le d_2.

Let f=Q(A)f=Q(\mathcal{A}), and let mdr⁡(f)\operatorname{mdr}(f) be the minimal degree of a nonzero syzygy afx+bfy+cfz=0af_x+bf_y+cf_z=0. These Jacobian syzygies form the module D0(A)D_0(\mathcal{A}), and D(A)=SθE⊕D0(A)D(\mathcal{A})=S\theta_E\oplus D_0(\mathcal{A}), where θE\theta_E is the Euler derivation. Hence a free arrangement has mdr⁡(f)=d1\operatorname{mdr}(f)=d_1. Let τ(A)=∑P(mP−1)2\tau(\mathcal{A})=\sum_P(m_P-1)^2 be the global Tjurina number of ff, and put τmax⁡(d,r)=(d−1)2−r(d−1−r)\tau_{\max}(d,r)=(d-1)^2-r(d-1-r).

Theorem 2.1 ([9 Cor. 1.2], based on du Plessis–Wall [12]). If r=mdr⁡(f)<d/2r=\operatorname{mdr}(f)<d/2, then A\mathcal{A} is free if and only if τ(A)=τmax⁡(d,r)\tau(\mathcal{A})=\tau_{\max}(d,r).

Ziegler's multirestriction and Yoshinaga's criterion

For a line H∈AH \in\mathcal{A}, the Ziegler multirestriction [Zie89] is the multiarrangement (AH,μH)(\mathcal{A}^{H},\mu^{H}) on H≅P1H \cong\mathbb{P}^{1} consisting of the rHr_{H} points of HH, where PP has multiplicity μH(P)=mP−1\mu^{H}(P)=m_{P}-1. It is free (being of rank 22), with exponents (e1H,e2H)(e_{1}^{H},e_{2}^{H}), e1H≤e2He_{1}^{H}\leq e_{2}^{H}, e1H+e2H=∑P∈H(mP−1)=d−1e_{1}^{H}+e_{2}^{H}=\sum_{P\in H}(m_{P}-1)=d-1. Moreover e1He_{1}^{H} is the minimal degree of a nonzero element of D(AH,μH)D(\mathcal{A}^{H},\mu^{H}).

Theorem 2.2 (Yoshinaga [25 Thm. 3.2]; see also [1 Thm. 2.3]). For every line HH of an arrangement A\mathcal{A} of lines in P2(C)\mathbb{P}^{2}(\mathbb{C}), b≥e1He2Hb\geq e_{1}^{H}e_{2}^{H}, and A\mathcal{A} is free if and only if b=e1He2Hb=e_{1}^{H}e_{2}^{H}.

Lemma 2.3. Let L(A)L(\mathcal{A}) be of type (d;d1,d2)(d;d_{1},d_{2}) and let H∈AH\in\mathcal{A}. Then A\mathcal{A} is free if and only if e1H≥d1e_{1}^{H}\geq d_{1}, i.e. if and only if D(AH,μH)D(\mathcal{A}^{H},\mu^{H}) has no nonzero element of degree d1−1d_{1}-1.

Proof. If A\mathcal{A} is free, then e1He2H=d1d2e_{1}^{H}e_{2}^{H}=d_{1}d_{2} and e1H+e2H=d1+d2e_{1}^{H}+e_{2}^{H}=d_{1}+d_{2}, so e1H=d1e_{1}^{H}=d_{1}. Conversely, the function x↦x(d−1−x)x\mapsto x(d-1-x) is increasing for x≤(d−1)/2x\leq(d-1)/2, so d1≤e1H≤(d−1)/2d_{1}\leq e_{1}^{H}\leq(d-1)/2 implies e1He2H≥d1d2=be_{1}^{H}e_{2}^{H}\geq d_{1}d_{2}=b. Theorem 2.2 then gives equality, hence freeness. □\square

Concretely, let H≅P1H\cong\mathbb{P}^{1} have homogeneous coordinates (X:Y)(X:Y) and write P=(sP:tP)P=(s_{P}:t_{P}). A derivation θ=g1∂X+g2∂Y\theta=g_{1}\partial_{X}+g_{2}\partial_{Y}, with g1,g2g_{1},g_{2} homogeneous of degree kk, lies in D(AH,μH)D(\mathcal{A}^{H},\mu^{H}) if and only if (tPX−sPY)μH(P)(t_{P}X-s_{P}Y)^{\mu^{H}(P)} divides hP:=tPg1−sPg2h_{P}:=t_{P}g_{1}-s_{P}g_{2} for every P∈HP\in H. If tP≠0t_{P}\neq0, this divisibility is equivalent to the vanishing at (sP,tP)(s_{P},t_{P}) of the Hasse derivatives ∂X(j)hP\partial_{X}^{(j)}h_{P}, 0≤j<μH(P)0\leq j<\mu^{H}(P); if tP=0t_{P}=0 one uses ∂Y(j)\partial_{Y}^{(j)} instead. This is valid in every characteristic. We obtain a system of d−1d-1 linear equations in the 2(k+1)2(k+1) coefficients of g1,g2g_{1},g_{2}, and we denote its matrix by MH,kM_{H,k}. If XX-derivatives are used at every point, the resulting system is implied by the exact one, so its solution space contains D(AH,μH)kD(\mathcal{A}^{H},\mu^{H})_{k}, with equality when all tP≠0t_{P}\neq0.

Reduction to three lattice types

Definition 3.1. A lattice LL is a Terao counterexample if there are complex line arrangements A,B\mathcal{A},\mathcal{B} with L(A)≅L(B)≅LL(\mathcal{A})\cong L(\mathcal{B})\cong L, A\mathcal{A} free and B\mathcal{B} not free.

In this section LL is a Terao counterexample with dd lines, realized by A\mathcal{A} (free) and B\mathcal{B} (not free), and (d1,d2)(d_{1},d_{2}), d1≤d2d_{1}\leq d_{2}, are the exponents of A\mathcal{A}. We put

X=∑P(mP−12),W=∑P(mP−22)=t4+3t5+6t6+⋯ .X=\sum_{P}\binom{m_{P}-1}{2},\qquad W=\sum_{P}\binom{m_{P}-2}{2}=t_{4}+3t_{5}+6t_{6}+\cdots.

Lemma 3.2 (counting). For every lattice of type (d;d1,d2)(d;d_{1},d_{2}):

(a) ∑P(mP2)=(d2)\sum_{P}\binom{m_{P}}{2}=\binom{d}{2} and ∑P(mP−1)=d1d2+d−1\sum_{P}(m_{P}-1)=d_{1}d_{2}+d-1;

(b) X=(d−12)−d1d2X=\binom{d-1}{2}-d_{1}d_{2} and τ=(d−1)2−d1d2=τmax⁡(d,d1)\tau=(d-1)^{2}-d_{1}d_{2}=\tau_{\max}(d,d_{1});

(c) ∑He(H)=3X−W\sum_{H}e(H)=3X-W; in particular, with β:=3X−dd2\beta:=3X-dd_{2},

W+∑H(e(H)−d2)=β(3)W+\sum_{H}\bigl(e(H)-d_{2}\bigr)=\beta \tag*{(3)}

(budget identity).

Conversely, for every simple rank 33 matroid on dd elements, X=(d2)−∑P(mP−1)=(d2)−(b+d−1)X=\binom{d}{2}-\sum_{P}(m_{P}-1)=\binom{d}{2}-(b+d-1) by (2). So XX determines χ0\chi_{0}, and every lattice with X=(d−12)−d1d2X=\binom{d-1}{2}-d_{1}d_{2} is of type (d;d1,d2)(d;d_{1},d_{2}).

Proof. (a) The first identity counts pairs of lines; the second is (2) with b=d1d2b=d_{1}d_{2}. (b) X=∑P(mP2)−∑P(mP−1)X=\sum_{P}\binom{m_{P}}{2}-\sum_{P}(m_{P}-1). For τ\tau, note that (m−1)2=2(m−12)+(m−1)(m-1)^{2}=2\binom{m-1}{2}+(m-1), so τ=2X+d1d2+d−1=(d−1)2−d1d2\tau=2X+d_{1}d_{2}+d-1=(d-1)^{2}-d_{1}d_{2}. (c) ∑He(H)=∑PmP(mP−2)\sum_{H}e(H)=\sum_{P}m_{P}(m_{P}-2), and m(m−2)=3(m−12)−(m−22)m(m-2)=3\binom{m-1}{2}-\binom{m-2}{2} for all mm. □\square

Lemma 3.3. m≤d1−1m \le d_1 - 1.

Proof. This is the first assertion of [8 Cor. 1.4]: Terao’s conjecture holds for a free arrangement A\mathcal{A} with m≥d1m \ge d_1. For completeness we include a short argument; in the case d1∈{m−1,m}d_1 \in\{m - 1,m\} it uses [11 Props. 2.1–2.2] instead of the original argument of [8]. We have mdr⁡(fA)=d1\operatorname{mdr}(f_{\mathcal{A}}) = d_1. By Dimca’s theorem [8 Thm. 1.2], applied to a point of maximal multiplicity mm, either d1=d−md_1 = d - m or m−1≤d1≤d−m−1m - 1 \le d_1 \le d - m - 1.

Case d1=d−md_1 = d - m. This case is covered by [8 Cor. 1.4(1)]; we give the argument. Then m=d−d1≥d2+1>d1m = d - d_1 \ge d_2 + 1 > d_1. Apply [8 Thm. 1.2] to B\mathcal{B}, which has the same mm. Either mdr⁡(fB)=d−m=d1\operatorname{mdr}(f_{\mathcal{B}}) = d - m = d_1, or m−1≤mdr⁡(fB)≤d−m−1=d1−1m - 1 \le\operatorname{mdr}(f_{\mathcal{B}}) \le d - m - 1 = d_1 - 1. The second alternative would give m≤d1m \le d_1, which is impossible. Hence mdr⁡(fB)=d1≤(d−1)/2<d/2\operatorname{mdr}(f_{\mathcal{B}}) = d_1 \le(d - 1)/2 < d/2, and τ(B)=τ(A)=τmax⁡(d,d1)\tau(\mathcal{B}) = \tau(\mathcal{A}) = \tau_{\max}(d,d_1) by Lemma 3.2(b). By Theorem 2.1, B\mathcal{B} is free, a contradiction.

Case d1∈{m−1,m}d_1 \in\{m - 1,m\}. By [11 Prop. 2.1], a line arrangement in P2(C)\mathbb{P}^{2}(\mathbb{C}) with dd lines, maximal multiplicity mm and τ=τmax⁡(d,m−1)\tau= \tau_{\max}(d,m - 1) is free, with exponents (m−1,d−m)(m - 1,d - m). By [11 Prop. 2.2], if instead τ=τmax⁡(d,m)\tau= \tau_{\max}(d,m), it is free with exponents (m,d−m−1)(m,d - m - 1). Now m(B)=mm(\mathcal{B}) = m, and τ(B)=τ(A)=τmax⁡(d,d1)\tau(\mathcal{B}) = \tau(\mathcal{A}) = \tau_{\max}(d,d_1) by Lemma 3.2(b), which is τmax⁡(d,m−1)\tau_{\max}(d,m - 1) or τmax⁡(d,m)\tau_{\max}(d,m). Hence B\mathcal{B} is free, a contradiction. Hence d1≥m+1d_1 \ge m + 1. □

Lemma 3.4. rH≤d1r_H \le d_1, i.e. e(H)≥d2e(H) \ge d_2, for every line HH.

Proof. Fix HH, and choose H0∈AH_0 \in\mathcal{A} with H0≠HH_0 \ne H. The deconing of A\mathcal{A} with respect to H0H_0 is the affine arrangement d⁡A\operatorname{d}\mathcal{A} of the d−1d - 1 lines H′∖H0H' \setminus H_0, H′≠H0H' \ne H_0, in P2∖H0≅C2\mathbb{P}^{2} \setminus H_0 \cong\mathbb{C}^{2}. Its cone is A\mathcal{A}, so it is free in the sense of [1], and χ(d⁡A,t)=(t−d1)(t−d2)\chi(\operatorname{d}\mathcal{A},t) = (t - d_1)(t - d_2). The affine line H∖H0H \setminus H_0 contains exactly the intersection points of A\mathcal{A} on HH other than H∩H0H \cap H_0, i.e. rH−1r_H - 1 of them. By [1 Cor. 1.2(1)], rH−1≤d1r_H - 1 \le d_1 or rH−1=d2r_H - 1 = d_2.

Suppose that rH∈{d1+1,d2+1}r_H \in\{d_1 + 1,d_2 + 1\}. Then χ(BH,t)=(t−1)(t−rH+1)\chi(\mathcal{B}^{H},t) = (t - 1)(t - r_H + 1) divides χ(B,t)=(t−1)(t−d1)(t−d2)\chi(\mathcal{B},t) = (t - 1)(t - d_1)(t - d_2), and BH\mathcal{B}^{H} is free since it has rank 2. By Abe’s division theorem [2 Thm. 1.1], B\mathcal{B} is free, a contradiction. Hence rH≤d1r_H \le d_1. The reformulation in terms of e(H)e(H) is (1). □

Lemma 3.5. rH≥4r_H \ge4, i.e. e(H)≤d−5e(H) \le d - 5, for every line HH.

Proof. Suppose that rH≤3r_H \le3. The Ziegler multirestrictions of A\mathcal{A} and B\mathcal{B} on HH consist of at most three points with the same multiplicities. Since PGL⁡2(C)\operatorname{PGL}_2(\mathbb{C}) acts 3-transitively on P1\mathbb{P}^{1}, they are isomorphic multiarrangements, so e1He_1^H is the same for A\mathcal{A} and B\mathcal{B}. (For three points the exponents are given explicitly by Wakamiko’s formula [24].) Since A\mathcal{A} is free, e1H=d1e_1^H = d_1 by Lemma 2.3, and then B\mathcal{B} is free by the same lemma, a contradiction. □

Remark 3.6. All arrangements considered here are balanced: every multiplicity μH(P)=mP−1\mu^H(P) = m_P - 1 is at most half of d−1d - 1, because m≤d1−1≤(d−3)/2m \le d_1 - 1 \le(d - 3)/2. For balanced arrangements, [1 Thm. 2.8] gives e2H−e1H≤rH−2e_2^H - e_1^H \le r_H - 2, and freeness is combinatorial as soon as d2−d1∈{rH−2,rH−3}d_2 - d_1 \in\{r_H - 2,r_H - 3\} for some line HH. This would happen Lemma 3.5 to rH≥d2−d1+4r_H \ge d_2 - d_1 + 4, which we do not need. For the lattices of Theorem 1.3 we have d2−d1∈{0,2}d_2 - d_1 \in\{0,2\} and rH≥6r_H \ge6, so this criterion does not apply to them.

Lemma 3.7 (Bojanowski’s inequality [6]; see [21 Thm. 19], [20 Rem. 2.5]). Let A\mathcal{A} be an arrangement of d≥6d \ge6 lines in P2(C)\mathbb{P}^{2}(\mathbb{C}) with tr=0t_r = 0 for r>2d/3r > 2d/3. Then

t2+34t3≥d+∑r≥5(r24−r)tr.t_2 + \frac{3}{4}t_3 \ge d + \sum_{r \ge5}\left(\frac{r^2}{4} - r\right)t_r.

If L(A)L(\mathcal{A}) is of type (d;d1,d2)(d; d_1, d_2), this inequality is equivalent to

3W≥5X−4d1d2+4.(4)3W \ge5X - 4d_1d_2 + 4. \tag*{(4)}

Bojanowski’s inequality follows from an inequality of Langer for line arrangements, [15 Prop. 11.3.1], which Langer derives from his orbifold Miyaoka–Yau inequality.

Proof of the equivalence with (4). For every integer kk,

54(k−12)−(k−1)−(k24−k)=34(k−22),\frac{5}{4}\binom{k-1}{2}-(k-1)-\left(\frac{k^2}{4}-k\right)=\frac{3}{4}\binom{k-2}{2},

as both sides equal 38k2−158k+94\frac{3}{8}k^2-\frac{15}{8}k+\frac{9}{4}. Since k2/4−kk^2/4-k equals −1-1, −34-\frac{3}{4} and 00 for k=2,3,4k=2,3,4, summing over all points and using Lemma 3.2(a) gives

34W=54X−(d1d2+d−1)+(t2+34t3−∑r≥5(r24−r)tr).\frac{3}{4}W=\frac{5}{4}X-(d_1d_2+d-1)+\left(t_2+\frac{3}{4}t_3-\sum_{r\geq5}\left(\frac{r^2}{4}-r\right)t_r\right).

Hence Bojanowski’s inequality is equivalent to 34W≥54X−d1d2+1\frac{3}{4}W\geq\frac{5}{4}X-d_1d_2+1.

Proposition 3.8. Let LL be a Terao counterexample with d∈{15,16}d\in\{15,16\} lines, of type (d;d1,d2)(d;d_1,d_2). Then

(d;d1,d2)∈{(15;6,8), (15;7,7), (16;7,8)},(d;d_1,d_2)\in\{(15;6,8),\ (15;7,7),\ (16;7,8)\},

and LL satisfies:

(R1) 3≤mP≤d1−13\leq m_P\leq d_1-1 for every point with mP≥3m_P\geq3;

(R2) d2≤e(H)≤d−5d_2\leq e(H)\leq d-5 for every line HH;

(R3) X=(d−12)−d1d2X=\binom{d-1}{2}-d_1d_2;

(R4) W≥Wmin⁡:=⌈(5X−4d1d2+4)/3⌉W\geq W_{\min}:=\left\lceil(5X-4d_1d_2+4)/3\right\rceil;

(R5) W+∑H(e(H)−d2)=β=3X−dd2W+\sum_H(e(H)-d_2)=\beta=3X-dd_2.

Proof. (R1)–(R5) follow from Lemmas 3.3, 3.4, 3.5, 3.2 and 3.7. Lemma 3.7 applies because d≥6d\geq6 and m≤d1−1≤(d−1)/2−1<2d/3m\leq d_1-1\leq(d-1)/2-1<2d/3. Since X=(d−12)−d1d2≥(d−12)−(d−1)24>0X=\binom{d-1}{2}-d_1d_2\geq\binom{d-1}{2}-\frac{(d-1)^2}{4}>0, some point has mP≥3m_P\geq3, so m≥3m\geq3 and Lemma 3.3 gives d1≥4d_1\geq4. By Lemma 3.2(c), (R5) holds for every lattice satisfying (R3); we list it because it is used for pruning. By (R2) every summand in (R5) is nonnegative, so Wmin⁡≤W≤βW_{\min}\leq W\leq\beta. Table 1 lists all types with 4≤d1≤(d−1)/24\leq d_1\leq(d-1)/2; all but the three stated ones have Wmin⁡>βW_{\min}>\beta.

Remark 3.9. Each line has two degrees of freedom, a projective frame fixes four lines, and a point of multiplicity mPm_P imposes mP−2m_P-2 conditions. The expected dimension of the space of realizations modulo PGL3\mathrm{PGL}_3 is therefore

2(d−4)−∑P(mP−2)=2d−8−(X−W).2(d-4)-\sum_P(m_P-2)=2d-8-(X-W).

using m−2=(m−12)−(m−22)m-2=\binom{m-1}{2}-\binom{m-2}{2}. In the three remaining types it equals −12-12 (type (15;6,8)(15;6,8), where (R4) and (R5) force W=9W=9), W−20≤1W-20\le1, and W−25≤−6W-25\le-6, respectively. This is only a parameter count, not a dimension bound: the incidence conditions need not be independent. It suggests that realizable lattices may be rare, which is what we find.

Remark 3.10. In type (15;6,8)(15;6,8) we have Wmin⁡=β=9W_{\min}=\beta=9. By (R5) this forces W=9W=9 and e(H)=8e(H)=8, i.e. rH=6r_H=6, for every line.

Enumeration of the residual lattices

For a type (d;d1,d2)(d;d_1,d_2) from Proposition 3.8, let L(d;d1,d2)\mathcal{L}(d;d_1,d_2) be the set of isomorphism classes of simple rank 3 matroids on dd elements satisfying (R1)–(R5). By Proposition 3.8 and Lemma 3.2, every Terao counterexample with 15 or 16 lines lies in one of the three sets L(15;6,8)\mathcal{L}(15;6,8), L(15;7,7)\mathcal{L}(15;7,7), L(16;7,8)\mathcal{L}(16;7,8). In this section we enumerate these sets.

Encoding

A matroid is given by its family of blocks (§ 2.1). Two matroids are isomorphic if and only if their bipartite point–block incidence graphs are isomorphic by an isomorphism preserving the bipartition. We use the canonical labelling of nauty [18] (program labelg, with the points forming the first colour class) and store one canonical graph6 string per class.

Stage 1: heavy configurations

For a matroid LL, let H(L)\mathcal{H}(L) be its set of blocks of size at least 4, the heavy configuration of LL. Triple points contribute 0 to WW, so W(L)=W(H(L))W(L)=W(\mathcal{H}(L)). For an element vv of the matroid (a line of the arrangement; in this section we call the elements points, as in the theory of linear spaces) put

h(v)=∑B∈Hv∈B(∣B∣−2),XH=∑B∈H(∣B∣−12).h(v)=\sum_{\substack{B\in\mathcal{H}\\v\in B}}(|B|-2),\qquad X_{\mathcal{H}}=\sum_{B\in\mathcal{H}}\binom{|B|-1}{2}.

Then e(v)=h(v)+τ3(v)e(v)=h(v)+\tau_3(v), where τ3(v)\tau_3(v) is the number of triple points through vv, and t3=X−XHt_3=X-X_{\mathcal{H}}. Put hmax⁡=min⁡(d−5,d2+β−Wmin⁡)h_{\max}=\min(d-5,d_2+\beta-W_{\min}).

Lemma 4.1. If LL satisfies (R1)–(R5), then H=H(L)\mathcal{H}=\mathcal{H}(L) satisfies:

(H1) its blocks have sizes in [4,d1−1][4,d_1-1] and pairwise share at most one element;

(H2) h(v)≤hmax⁡h(v)\le h_{\max} for all vv, and XH≤XX_{\mathcal{H}}\le X;

(H3) W(H)+∑v(h(v)−d2)+≤βW(\mathcal{H})+\sum_v(h(v)-d_2)^+\le\beta, and ∑v(h(v)−d2)+≤β−Wmin⁡\sum_v(h(v)-d_2)^+\le\beta-W_{\min};

(H4) W(H)≥Wmin⁡W(\mathcal{H})\ge W_{\min};

(H5) for each vv, let u(v)u(v) be the number of w≠vw\ne v not in a common block of H\mathcal{H} with vv, and put c(v)=min⁡(⌊u(v)/2⌋,hmax⁡−h(v))c(v)=\min(\lfloor u(v)/2\rfloor,h_{\max}-h(v)). Then (d2−h(v))+≤c(v)(d_2-h(v))^+\le c(v) for every vv, and

∑v(d2−h(v))+≤3(X−XH)≤∑vc(v).\sum_v(d_2-h(v))^+\le3(X-X_{\mathcal{H}})\le\sum_v c(v).

Conditions (H1)–(H3) are inherited by every subset of H\mathcal{H}.

Proof. (H1) is (R1). For (H2): h(v)≤e(v)≤d−5h(v) \le e(v) \le d-5, by (R5) e(v)−d2≤β−W≤β−Wmin⁡e(v)-d_2 \le\beta-W \le\beta-W_{\min}, and X−XH=t3≥0X-X_{\mathcal{H}}=t_3 \ge0. For (H3): e(v)≥max⁡(h(v),d2)e(v) \ge\max(h(v),d_2), so e(v)−d2≥(h(v)−d2)+e(v)-d_2 \ge(h(v)-d_2)^{+}; now use (R5) and (R4). (H4) is (R4). For (H5): the triple points through vv use disjoint pairs among the u(v)u(v) elements not yet collinear with vv. Moreover τ3(v)=e(v)−h(v)\tau_3(v)=e(v)-h(v) lies between (d2−h(v))+(d_2-h(v))^{+} and hmax⁡−h(v)h_{\max}-h(v), and ∑vτ3(v)=3t3\sum_v \tau_3(v)=3t_3. Heredity holds because hh, XHX_{\mathcal{H}} and WW can only decrease when blocks are removed.

We generate all heavy configurations satisfying (H1)–(H3), one per isomorphism class, by McKay’s canonical augmentation [17], adding one block at a time, and output those satisfying (H4) and (H5).

  • The candidate blocks are the sets of size 4,…,d1−14,\ldots,d_1-1 containing no pair already covered, and containing no vv with h(v)+2>hmax⁡h(v)+2>h_{\max}.

  • The canonical deletion of a configuration is the block whose vertex comes last in the canonical labelling of the incidence graph. Blocks are coloured by size, after the points, so the last vertex is always a block.

  • A child is accepted if and only if the new block lies in the orbit of the canonical deletion under the automorphism group of the child. Isomorphic accepted siblings are identified by comparing canonical graphs.

Every child is tested again against (H1)–(H3), since a candidate block of size at least 5 may overshoot hmax⁡h_{\max}. McKay’s scheme [17] applies because:

  • (H1)–(H3) are hereditary, so deleting the canonical block of a valid configuration gives a valid configuration;

  • the candidate filter admits the canonical block of every valid child, since ∣B∣−2≥2|B|-2 \ge2;

  • the acceptance test is isomorphism invariant, because the ordered colour partition is fixed.

Hence this produces every isomorphism class of configurations satisfying (H1)–(H3), each at least once; the final deduplication with labelg removes any remaining repetition.

Stage 2: completion by triple points

Let H\mathcal{H} be a representative heavy configuration. A completion is a set TT of 3-subsets such that H∪T\mathcal{H}\cup T is a linear space with heavy configuration H\mathcal{H} satisfying (R1)–(R5). We enumerate the completions by backtracking over the points v=0,1,…,d−1v=0,1,\ldots,d-1, ordered by decreasing h(v)h(v). At step vv we choose all triples whose smallest element is vv, so that e(v)e(v) takes its final value, and we prune with (R2), (R5) and capacity bounds as in (H5).

We break symmetry as follows. At step vv, the points w>vw>v that lie in exactly the same existing blocks form classes, and new triples through vv use only initial segments of the currently unused members of each class. Permuting the remaining unused members of such a class fixes the pre-existing blocks, all processed points, and the triples already chosen at this step. Induction over the triple choices therefore shows that every completion is equivalent to one respecting this rule.

Every LL satisfying (R1)–(R5) is isomorphic to a completion of the representative of H(L)\mathcal{H}(L). Therefore the completions of all representatives, after deduplication by canonical form, are exactly the elements of L(d;d1,d2)\mathcal{L}(d;d_1,d_2).

Results of the enumeration

typeheavy configurationscompletions foundclasses ∣L∣|\mathcal{L}|
(15; 6, 8)1 9521 3463
(15; 7, 7)96 72510 016 8676 091 145
(16; 7, 8)8 324 81117 590 57316 453 132

Table 1 (PDF p. 9).

For (16; 7, 8), the canonical augmentation visited 8 885 007 nodes (about 9 minutes on one core), and the completion visited 2.97×10102.97 \times10^{10} nodes (about 25 minutes).

Exact computation of realization spaces

Let LL be a simple rank 3 matroid on {1,…,d}\{1,\ldots,d\}. A realization of LL over a field KK (of any characteristic) is an arrangement (ℓ1,…,ℓd)(\ell_1,\ldots,\ell_d) of lines in P2(K)\mathbb{P}^2(K), labelled by the elements of LL, whose labelled lattice, defined exactly as over C\mathbb{C}, is LL.

Normal form and construction sequence

Suppose that LL admits a frame: four elements, no three of them in a common block. Every residual matroid used below is explicitly checked to admit such a frame. Since PGL⁡3\operatorname{PGL}_3 acts simply transitively on ordered 4-tuples of lines in general position, every realization is projectively equivalent to a unique one in which the frame lines are x=0x=0, y=0y=0, z=0z=0, x+y+z=0x+y+z=0. Order the remaining elements. When an element ℓ\ell is placed, call a block determined if it contains ℓ\ell and at least two placed elements; the corresponding point is then already constructed. There are three kinds of steps:

  • det: ℓ\ell lies in at least two determined blocks, so it is the line through two constructed points;

  • pencil: ℓ\ell lies in exactly one determined block, through the point P=ℓa∩ℓbP=\ell_a\cap\ell_b, so ℓ=ℓa+tℓb\ell=\ell_a+t\ell_b for a parameter tt;

  • free: ℓ\ell lies in no determined block, so it depends on two parameters.

In a pencil step, the only line through PP not of the form ℓa+tℓb\ell_a+t\ell_b is ℓb\ell_b, which is already one of the lines of LL. So no realization is lost. The cost of the sequence is the number of parameters. We choose frames and orders greedily so as to minimize the cost; correctness does not depend on this choice, and the program checks explicitly that the chosen frame is in general position. With our greedy choice, all 22 544 280 matroids except one received a sequence of cost at most 1.

One parameter

For cost ≤1\leq1, every line is represented by a vector ℓi(t)∈Z[t]3\ell_i(t)\in\mathbb{Z}[t]^3, built by cross products and pencils. Each vector is made primitive, and every common factor removed is recorded as a forbidden polynomial; at its roots two distinct points or lines of the construction would coincide. Let G∈Z[t]G\in\mathbb{Z}[t] be the gcd of the required determinants det⁡(ℓb0,ℓb1,ℓx)\det(\ell_{b_0},\ell_{b_1},\ell_x), for all blocks B∋b0,b1,xB\ni b_0,b_1,x. Let NN consist of the forbidden polynomials and of the determinants det⁡(ℓi,ℓj,ℓk)\det(\ell_i,\ell_j,\ell_k) of all triples not contained in a block.

Proposition 5.1. Suppose the sequence has cost 1. Every complex realization of LL is projectively equivalent to exactly one arrangement (ℓi(t0))i(\ell_i(t_0))_i with t0∈Ct_0\in\mathbb{C}. This arrangement is a realization if and only if G(t0)=0G(t_0)=0 and n(t0)≠0n(t_0)\ne0 for all n∈Nn\in N. Consequently:

(a) if some ℓi\ell_i is identically 0, or GG is a nonzero constant, or every irreducible factor of GG divides an element of NN, or G=0G=0 and some element of NN is identically 0, then LL is not realizable over C\mathbb{C};

(b) if G≠0G\ne0, the realizations correspond to the roots of the irreducible factors qq of GG that divide no element of NN, and the realizations coming from one factor qq form a single orbit under Aut⁡(C/Q)\operatorname{Aut}(\mathbb{C}/\mathbb{Q});

(c) if G=0G=0, the realizations form the Zariski open subset {t0:n(t0)≠0 ∀n∈N}\{t_0:n(t_0)\ne0\ \forall n\in N\} of A1\mathbb{A}^1.

Proof. The construction reproduces any normalized realization step by step: det steps are forced, and pencil steps were discussed above. If a constructed vector vanishes at t0t_0, then two elements that are distinct in LL give the same point or line. The conditions G(t0)=0G(t_0)=0 and n(t0)≠0n(t_0)\ne0 say exactly that the blocks of LL are collinear and that no other triple is. For an irreducible g∈Q[t]g\in\mathbb{Q}[t] and n∈Q[t]n\in\mathbb{Q}[t], nn vanishes at one root of gg if and only if g∣ng\mid n, if and only if nn vanishes at all roots of gg; and the roots of gg are permuted transitively by Aut⁡(C/Q)\operatorname{Aut}(\mathbb{C}/\mathbb{Q}). □

By the same argument, if the sequence has cost 0, the vectors ℓi\ell_i are integral, GG is an integer, and LL has at most one realization up to projective equivalence. It exists, and is then defined over Q\mathbb{Q}, if and only if G=0G=0 and no element of NN vanishes.

Several parameters

For the single matroid of cost 2, the construction has two pencil steps and no free step. The same cross-product construction therefore gives vectors in Z[t1,t2]\mathbb{Z}[t_1,t_2] and covers every normalized realization. Required determinants must vanish and nonblock determinants must be nonzero, exactly as in the one-parameter case.

This class has the following elementary obstruction, verified directly in the ancillary module exact_multi.py and in the script identifications/certificates.py. It is the class with index 13095 (counted from 0) in the list of type (15;7,7)(15;7,7); its elements are numbered from 0. With the frame (3,4,7,8)(3,4,7,8) and parameters t1,t2t_1,t_2, the construction gives

q=det⁡(ℓ1,ℓ2,ℓ14)=−4t12+4t1−1,f=det⁡(ℓ0,ℓ1,ℓ7)=2t1−1.q=\det(\ell_1,\ell_2,\ell_{14})=-4t_1^2+4t_1-1,\qquad f=\det(\ell_0,\ell_1,\ell_7)=2t_1-1.

The triple {1,2,14}\{1,2,14\} lies in a block and the triple {0,1,7}\{0,1,7\} does not, so a realization would satisfy q=0q=0 and f≠0f\ne0. This is impossible, since q+f2=0q+f^2=0 identically.

Deciding freeness

Freeness is invariant under Aut⁡(C/Q)\operatorname{Aut}(\mathbb{C}/\mathbb{Q}) acting on the coefficients, so by Proposition 5.1(b) it is constant on each orbit of realizations. By Lemma 2.3, deciding freeness of a realization of type (d;d1,d2)(d;d_1,d_2) amounts to deciding whether the matrix MH,d1−1M_{H,d_1-1} has trivial kernel, for one line HH. Coordinates on HH are given by two fixed integral linear forms whose common zero does not lie on HH.

Proposition 5.2 (certificates). Let g∈Z[t]g\in\mathbb{Z}[t] be a primitive irreducible factor as in Proposition 5.1(b), let K=Q[t]/(g)K=\mathbb{Q}[t]/(g), and let AKA_K be the corresponding realization.

(a) (free) Suppose pp is a prime not dividing the leading coefficient of gg, r∈Fpr\in\mathbb{F}_p is a root of g mod pg\bmod p, and the reduced arrangement Ap=(ℓi(r))iA_p=(\ell_i(r))_i over Fp\mathbb{F}_p has the same lattice. If D(ApH,μH)D(A_p^H,\mu^H) has no nonzero element of degree d1−1d_1-1 (computed with the exact system of § 2.3 over Fp\mathbb{F}_p), then AKA_K is free.

(b) (non-free) Suppose all points PP of HH have tP≠0t_P\ne0 in KK. If MH,d1−1M_{H,d_1-1} has a nonzero kernel over KK, computed by exact Gaussian elimination in KK, then AKA_K is not free.

Proof. We apply Lemma 2.3 to AKA_K through an embedding K↪CK\hookrightarrow\mathbb{C}. (a) Let Λ\Lambda be the image of Z[1/lc⁡(g)][t]\mathbb{Z}[1/\operatorname{lc}(g)][t] in KK; it maps to Fp\mathbb{F}_p by t↦rt\mapsto r. Choose coordinate forms on HH with integral coefficients whose common zero avoids HH also modulo pp. The coordinates (sP,tP)(s_P,t_P) of the points of HH then lie in Λ\Lambda and reduce to the coordinates of the points of ApA_p on HH, which are distinct because the lattices agree. At each point use the XX-direction if tP≠0t_P\ne0 and the YY-direction otherwise. The resulting matrix over Λ\Lambda is exact both over KK and, after reduction, over Fp\mathbb{F}_p. Its rank over Fp\mathbb{F}_p is at most its rank over KK, since a nonzero maximal minor modulo pp is the image of a nonzero minor. Hence a trivial kernel modulo pp gives a trivial kernel over KK, so e1H≥d1e_1^H\ge d_1 and AKA_K is free. (b) Under the hypothesis the system is exact, so e1H≤d1−1e_1^H\le d_1-1 and AKA_K is not free. □

For a family (G=0G = 0), let DH∈Z[t]D_H \in\mathbb{Z}[t] be the determinant of MH,d1−1(t)M_{H,d_1-1}(t) when the matrix is square. Otherwise let DHD_H be the gcd of the determinants of a few random integral square combinations R⋅MH,d1−1(t)R \cdot M_{H,d_1-1}(t). If a member At0\mathcal{A}_{t_0} is not free, then the rank of MH,d1−1(t0)M_{H,d_1-1}(t_0) drops and DH(t0)=0D_H(t_0) = 0. This holds even where the XX-direction encoding is not exact, because that encoding only enlarges the kernel (§ 2.3). We take three lines HH such that the common zero of the two coordinate forms lies on H(t0)H(t_0) for no t0t_0 in the realization space, so that the forms restrict to coordinates on every H(t0)H(t_0), and let D=gcd⁡HDHD = \gcd_H D_H. (For F1F_1 and F2F_2, the only lattices to which this criterion is applied, the three lines and the forms do not depend on tt. If a point of HH with μH(P)≥2\mu^H(P) \ge2 had tP=0t_P = 0, then DHD_H would vanish identically and the criterion would simply fail.) If D≠0D \ne0 and every irreducible factor of DD divides an element of NN, then every realization in the family is free. If DD is a nonzero constant, all members are free.

In all certificates of type (b) that occur, we checked that every tPt_P is nonzero in KK.

Results

Applied to all classes of § 4 (we call this the screening), the computation of § 5 gives the following verdicts.

(15;6,8)(15; 6, 8)(15;7,7)(15; 7, 7)(16;7,8)(16; 7, 8)
classes36 091 14516 453 132
not realizable: GG constant, or each factor of GG divides an element of NN26 021 70116 175 061
not realizable: a constructed vector vanishes identically069 433278 071
not realizable: elementary certificate (cost 2)010
realizable: finitely many (one Galois orbit) / one family1 / 09 / 10 / 0

Table 1 (PDF p. 11).

The remaining case of Proposition 5.1(a), G=0G = 0 with an element of NN vanishing identically, does not occur. Each realizable lattice has a single Galois orbit of realizations, or a single family. The orbit of F2F_2 has one element, since its construction has cost 0. (The program treats this cost-0 case like a family with constant lines, so the ancillary file results/verdict_15_77.txt lists 8 classes with status points and 2 with status family; in the table above F2F_2 is counted among the lattices with finitely many realizations.) All eleven are listed in Table 2. Their blocks are listed in Appendix § 9. Explicit equations of all of them are in the ancillary file results/realizable_lattices.txt.

(t2,t3,t4,t5,t6)(t_2,t_3,t_4,t_5,t_6)rHr_Hmmfieldrealizationsfreeness, remarks
L0L_0(0,25,0,3,0)(0,25,0,3,0)6156^ {15}55Q(ζ5)\mathbb{Q}(\zeta_5)44 (one orbit)free (6, 8); A(5,5,3)\mathcal{A}(5,5,3)
F1F_1(27,6,0,6,0)(27,6,0,6,0)637126^3 7^{12}55Q(a)\mathbb{Q}(a)A1\mathbb{A}^1 minus 7 pointsfree (7, 7); Cuntz’s family
F2F_2(21,6,6,3,0)(21,6,6,3,0)66796^6 7^955Q\mathbb{Q}11free (7, 7)
F3F_3(21,6,6,3,0)(21,6,6,3,0)66796^6 7^955Q(−3)\mathbb{Q}(\sqrt{-3})22 (one orbit)free (7, 7); ⊂A(G26)\subset\mathcal{A}(G_{26})
F4F_4(21,6,6,3,0)(21,6,6,3,0)66796^6 7^955Q(5)\mathbb{Q}(\sqrt{5})22 (one orbit)free (7, 7)
F5F_5(19,6,8,2,0)(19,6,8,2,0)67786^7 7^855Q(−3)\mathbb{Q}(\sqrt{-3})22 (one orbit)free (7, 7); ⊂A(G26)\subset\mathcal{A}(G_{26})
N1N_1(27,0,8,3,0)(27,0,8,3,0)5627125 6^2 7^{12}55Q(i)\mathbb{Q}(i)22 (one orbit)not free
N2N_2(22,8,4,2,1)(22,8,4,2,1)4627124 6^2 7^{12}66Q(i)\mathbb{Q}(i)22 (one orbit)not free
N3N_3(24,12,0,0,3)(24,12,0,0,3)637126^3 7^{12}66Q(ζ5)\mathbb{Q}(\zeta_5)44 (one orbit)not free; ⊂A(5,1,3)\subset\mathcal{A}(5,1,3)
N4N_4(17,16,0,1,2)(17,16,0,1,2)66796^6 7^966Q(ζ5)\mathbb{Q}(\zeta_5)44 (one orbit)not free; ⊂A(5,1,3)\subset\mathcal{A}(5,1,3)
N5N_5(10,20,0,2,1)(10,20,0,2,1)69766^9 7^666Q(ζ5)\mathbb{Q}(\zeta_5)44 (one orbit)not free; ⊂A(5,1,3)\subset\mathcal{A}(5,1,3)

Table 2. The realizable lattices in L(15;6,8)∪L(15;7,7)\mathcal{L}(15;6,8)\cup\mathcal{L}(15;7,7); L(16;7,8)\mathcal{L}(16;7,8) contains none. The column rHr_H gives the distribution of the number of points per line (637126^3 7^{12}: three lines with 6 points and twelve with 7). “Realizations” counts labelled realizations modulo PGL⁡3(C)\operatorname{PGL}_3(\mathbb{C}).

Proof of the Main Theorem. Let LL be a Terao counterexample with d∈{15,16}d \in\{15,16\} lines. By Proposition 3.8, L∈L(15;6,8)∪L(15;7,7)∪L(16;7,8)L \in\mathcal{L}(15; 6, 8) \cup\mathcal{L}(15; 7, 7) \cup\mathcal{L}(16; 7, 8). By § 4, these sets are completely known. By § 5 and the table at the beginning of this section, exactly eleven of their elements are realizable over C\mathbb{C}. For each of them, freeness is constant on the set of all complex realizations:

  • for the nine lattices L0L_0, F3F_3, F4F_4, F5F_5, N1,…,N5N_1,\ldots,N_5, the realizations form one Galois orbit, whose freeness is certified by Proposition 5.2;

  • for F2F_2, whose construction has cost 0 and whose unique realization is defined over Q\mathbb{Q}, by Lemma 2.3: the matrix MH,6M_{H,6} has trivial kernel over Q\mathbb{Q} (in the family criterion of § 5, DD is a nonzero constant);

  • for F1F_1, by the family criterion (D≠0D \ne0, and every irreducible factor of DD divides an element of NN); see also the certificate in § 6.1.

Hence LL cannot be a Terao counterexample. ∞ӘА

Proof of Theorem 1.3. Let A\mathcal{A} be as in the theorem. Abe’s restriction (proof of Lemma 3.4) and the absence of divisional freeness give rH≤d1r_H\le d_1 for all HH. Together with m≤d1−1m\le d_1-1, rH≥4r_H\ge4 and Lemmas 3.2 and 3.7, L(A)L(\mathcal{A}) satisfies (R1)–(R5) for its type. Since X>0X>0 we have m≥3m\ge3, hence d1≥4d_1\ge4, so Table 1 covers all possible types. Table 1 leaves only the three types above, and L(A)L(\mathcal{A}) is one of the realizable lattices of Table 2 that has a free realization. Conversely, these six lattices have m=5≤d1−1m=5\le d_1-1, rH∈{6,7}r_H\in\{6,7\}, and rH∉{d1+1,d2+1}r_H\notin\{d_1+1,d_2+1\} for all HH; so all their complex realizations, which are free by Section 6, satisfy the hypotheses and are not divisionally free. □

dd(d1,d2)(d_1,d_2)XXβ=3X−dd2\beta=3X-dd_25X−4d1d2+45X-4d_1d_2+4Wmin⁡W_{\min}m≤m\leqtype
15(4, 10)51399333excluded
15(5, 9)46354184excluded
15(6, 8)4392795remains
15(7, 7)42211866remains
16(4, 11)617133453excluded
16(5, 10)55579274excluded
16(6, 9)51943155excluded
16(7, 8)49192596remains

Table 1. The numerical reduction (Proposition 3.8).

Identifications

  • L0L_0. The reflection arrangement A(5,5,3)\mathcal{A}(5,5,3) is given by ∏k=04(x−ζky)(y−ζkz)(x−ζkz)\prod_{k=0}^{4}(x-\zeta^k y)(y-\zeta^k z)(x-\zeta^k z) with ζ=e2πi/5\zeta=e^{2\pi i/5}. Its 15 lines form the 3-net of the cyclic group of order 5, and the five lines of each class are concurrent. There are exactly two main classes of Latin squares of order 5. Correspondingly, L(15;6,8)\mathcal{L}(15;6,8) contains a second matroid with (t2,…,t6)=(0,25,0,3,0)(t_2,\ldots,t_6)=(0,25,0,3,0): a 3-net of the non-cyclic class whose three classes are pencils. It is not realizable, which is also seen directly. Each centre lies on the five lines of its class, and m=5m=5, so no line of the arrangement passes through two centres.

    • If the three centres are not collinear, take them to be the coordinate points. The lines are then x=ayx=ay, y=bzy=bz, x=czx=cz with a∈Aa\in A, b∈Bb\in B, c∈Cc\in C, where A,B,C⊂C∗A,B,C\subset\mathbb{C}^* have five elements, and concurrency means c=abc=ab. So ab=Cab=C for all a∈Aa\in A, and AA−1AA^{-1} lies in the stabilizer of BB, a subgroup of order at most 5. Hence AA and BB are cosets of μ5\mu_5, and the Latin square is cyclic.

    • If the centres are collinear, a projective change of coordinates turns the classes into the parallel classes x=ax=a, y=by=b, x+y=cx+y=c. Then A+B⊆CA+B\subseteq C, which contradicts ∣A+B∣≥∣A∣+∣B∣−1=9|A+B|\ge|A|+|B|-1=9 in (C,+)(\mathbb{C},+).

The 3-net alone, without concurrent classes, can be realized for a non-group Latin square of order 5; compare [22]. Barakat and Kühne [5 Sect. 8 of the arXiv version] determined the non-free loci of the matroids underlying the reflection arrangements of the groups G(n,n,3)G(n,n,3) for n<10n<10; for prime nn they found that a representation is free if and only if the characteristic of the field is not nn. For n=5n=5 this already shows that every complex arrangement with lattice L0L_0 is free.

  • N3N_3, N4N_4, N5N_5. The lattices of sub-arrangements below were computed exactly, with coefficients in the cyclotomic integers (ancillary script identifications/ident.py). The 816 sub-arrangements of 15 lines of A(5,1,3)=A(5,5,3)∪{x,y,z}\mathcal{A}(5,1,3)=\mathcal{A}(5,5,3)\cup\{x,y,z\} have 11 distinct lattices. Four of them satisfy (R1)–(R5): L0L_0 (remove x,y,zx,y,z), and N3N_3, N4N_4, N5N_5, of type (15;7,7)(15;7,7).

  • F3F_3, F5F_5. These are sub-arrangements of the reflection arrangement of the Shephard–Todd group G26G_{26}. With ω=e2πi/3\omega=e^{2\pi i/3}, A(G26)\mathcal{A}(G_{26}) consists of the 21 lines x,y,zx,y,z, x−ωjyx-\omega^j y, y−ωjzy-\omega^j z, x−ωjzx-\omega^j z and x+ωjy+ωkzx+\omega^j y+\omega^k z (0≤j,k≤20\le j,k\le2); it has (t2,t3,t4,t5)=(36,0,9,12)(t_2,t_3,t_4,t_5)=(36,0,9,12) and exponents (1,7,13)(1,7,13). Its 15-line sub-arrangements have 206 distinct lattices, and exactly two of them satisfy (R1)–(R5).

  • A(6,1,3)\mathcal{A}(6,1,3). None of its 231 lattices of 15-line sub-arrangements satisfies (R1)–(R5).

  • F1F_1. This is the lattice of Cuntz’s arrangements AtA_t [[7 §4.2]; see Appendix § 9. In Cuntz’s parametrization, an exact computation (ancillary script identifications/cuntz.py) shows that AtA_t has lattice F1F_1 exactly for

t∉{0,1,12}∪{t:t2+t−1=0}∪{t:t2−3t+1=0}.t\notin\left\{0,1,\frac{1}{2}\right\}\cup\{t:t^2+t-1=0\}\cup\{t:t^2-3t+1=0\}.
  • For t∈{0,1,12}t\in\left\{0,1,\frac{1}{2}\right\}, some of the 15 lines coincide.

  • At the four roots of t2+t−1t^2+t-1 and t2−3t+1t^2-3t+1, the lattice becomes that of the Coxeter arrangement A(H3)\mathcal{A}(H_3), with (t2,t3,t4,t5)=(15,10,0,6)(t_2,t_3,t_4,t_5)=(15,10,0,6) and exponents (1,5,9)(1,5,9).

In the published version [[3 §7]] the family is given in a different, symmetric parametrization, and [[3 Prop. 7.2]] describes its exceptional parameters correctly; its generic member also has lattice F1F_1 (checked by the same script). The description in the preprint [[7 §4.2]] lists −1-1 and 32±2\frac{3}{2}\pm\sqrt{2} instead of 3±52\frac{3\pm\sqrt{5}}{2} among the exceptional parameters. The arrangement AtA_t is free with exponents (1,7,7)(1,7,7) [[3 Prop. 7.2]], and Theorem 1.3 confirms that every complex arrangement with lattice F1F_1 is free.

The script identifications/certificates.py verifies the following short certificate. We use the normalized parameter aa of the ancillary files (not Cuntz’s tt); the arrangements A(a)\mathcal{A}(a) are given explicitly in results/realizable_lattices.txt. The irreducible factors of the polynomials in NN are exactly the factors of

B(a)=a(a−1)(a+1)(a2−a−1)(a2+a−1),B(a)=a(a-1)(a+1)(a^2-a-1)(a^2+a-1),

so by Proposition 5.1(c) the realizations of F1F_1 are the arrangements A(a)\mathcal{A}(a) with B(a)≠0B(a)\ne0. The first line of A(a)\mathcal{A}(a) is z=0z=0. In the coordinates (x,y)(x,y) on it, the Ziegler multirestriction is

x2(x+y)4y4(x−ay)((a−1)x−y)((1−a)x−ay)(ax−y).x^2(x+y)^4y^4(x-ay)((a-1)x-y)((1-a)x-ay)(ax-y).

Represent its points by (sP,tP)=(0,1),(1,−1),(1,0),(a,1),(1,a−1),(a,1−a),(1,a)(s_P,t_P)=(0,1),(1,-1),(1,0),(a,1),(1,a-1),(a,1-a),(1,a), in this order, and use the exact system of § 2.3, with Hasse derivatives in XX at the points with tP≠0t_P\ne0 and in YY at (1,0)(1,0). With the rows ordered by points and then by increasing order of the derivative, and the unknowns ordered as the coefficients of g1g_1 and then of g2g_2, by decreasing powers of XX, the 14×1414\times14 matrix of conditions for a derivation of degree 6 satisfies

det⁡M(a)=2a6(a−1)4(a+1)4(a2−a−1)2(a2+a−1)2.\det M(a)=2a^6(a-1)^4(a+1)^4(a^2-a-1)^2(a^2+a-1)^2.

So det⁡M(a)≠0\det M(a)\ne0 whenever B(a)≠0B(a)\ne0, and by Lemma 2.3 every realization of F1F_1 is free.

We have not tried to identify F2F_2, F4F_4, N1N_1 and N2N_2 in the literature. F2F_2 is a real arrangement defined over Q\mathbb{Q}, with defining polynomial

xyz(x+y)(x+z)(y+2z)(x+y+z)(x+y+2z)(x−y−z)⋅(x−y+2z)(x+2y+2z)(2x+y+z)(2x+y+2z)(2x+y+4z)(2x+y−2z).xyz(x+y)(x+z)(y+2z)(x+y+z)(x+y+2z)(x-y-z) \cdot(x-y+2z)(x+2y+2z)(2x+y+z)(2x+y+2z)(2x+y+4z)(2x+y-2z).

Reproducibility

The ancillary archive contains:

  • the C programs heavyaug (Stage 1) and lsgen3 (Stage 2), which require nauty;

  • the Python modules for the exact analysis of § 5 and § 6, which use python-flint [14] and SymPy; the class of cost 2 is excluded by the elementary certificate of § 5.3;

  • the scripts that compute the identifications of § 6.1 exactly and verify the short certificates of § 5.3 and § 6.1, in the folder identifications;

  • the results, including explicit equations of all realizable lattices;

  • a script run.sh that reruns the enumeration and screening of each type, checks the expected counts and canonical-list checksums, and compares all verdicts and detailed records with the supplied results. Its –certificates mode recomputes the eleven realizable cases, the cost-2 obstruction and the stated identifications, without rerunning the exhaustive enumeration.

Every entry of the lists is checked to encode a linear space satisfying (R1)–(R5) when it is screened, and the final verdict rejects any missing, duplicate, out-of-range or inconsistent record. The original computation used gcc 13.3, nauty 2.8.8, Singular 4.3.2, Python 3.13 and python-flint 0.9.0. Exact screening took about 0.9–1.2 ms per class: 1.5 CPU-hours for (15; 7, 7) and 5.5 for (16; 7, 8). The lists L(d;d1,d2)\mathcal{L}(d;d_1,d_2), as sorted files of canonical graph6 strings, have SHA-256 checksums

L(15;6,8)\mathcal{L}(15;6,8) 8de6cde5fc270e15cae212de4ee41d5202197e8aeab78cf6e65614650746dcd9 L(15;7,7)\mathcal{L}(15;7,7) f1be0ac91ef3d5e230d2812b08c3e67e2090445398d9bfd5f7905ab0677a9040 L(16;7,8)\mathcal{L}(16;7,8) 12a4cb695442e91e9d7fc55e9d5abc608c8cf99c14d96f308a4876e668fc9f4d

These refer to the canonical forms of nauty 2.8.8; the reproduction script requires these checksums and stops on a mismatch. Generated lists and intermediate output are not distributed, since the supplied programs regenerate them. The supplied reproduction scripts do not require Singular.

Statements and declarations

Competing interests. The authors have no financial interests to disclose.

Data availability. The code and the results of the computations are available as ancillary files of the arXiv version of this paper (§ 7).

Use of generative AI. Claude (Anthropic) was used for literature search, programming, and drafting the text; ChatGPT (OpenAI) for auditing, and editorial assistance. After using these tools, the authors reviewed and edited the content and takes full responsibility for the content of the publication.

Appendix A. Blocks of the realizable lattices

Lines are numbered 1, . . . , 9, a, . . . , f (so a = 10, . . . , f = 15). Each block, a point of multiplicity at least 3, is written as the word of its lines. The numbering is the canonical one of the ancillary file.

L0L_0, type (15; 6, 8), A(5,5,3)\mathcal{A}(5, 5, 3). 12345 6cdef 789ab 167 18c 19d 1af 1be 269 27c 28e 2ad 2bf 368 37d 39f 3ae 3bc 46b 47f 48d 49e 4ac 56a 57e 58f 59c 5bd

F1F_1, type (15; 7, 7), Cuntz’s family. 13bcf 149ae 2367e 2458f 57acd 689bd 12d 34d 5be 6af 79f 8ce

F2F_2, type (15; 7, 7). 124ad 136be 235cf 189c 279b 378a 457e 469f 568d 7df 8ef 9de abf ace bcd

F3F_3, type (15; 7, 7), sub-arrangement of A(G26)\mathcal{A}(G_{26}). 127ad 138be 239cf 145c 246b 356a 489d 579e 678f 4ef 5df 6de 7bc 8ac 9ab

F4F_4, type (15; 7, 7). 145ab 278ac 369bc 168e 179d 249e 256f 347f 358d 46d 57e 89f adf bef cde

F5F_5, type (15; 7, 7), sub-arrangement of A(G26)\mathcal{A}(G_{26}). 1389b 247ab 126c 145f 235e 346d 578c 59ad 679f 68ae 7de 8df 9ce acf bcd bef

N1N_1, type (15; 7, 7). 12345 169cd 178ab 2ade 2bcf 37ce 38df 46af 489e 56be 579f

N2N_2, type (15; 7, 7). 123456 18def 19abc 279d 37be 47cf 578a 2ae 39f 48b 5cd 689 6af 6bd 6ce

N3N_3, type (15; 7, 7), sub-arrangement of A(5,1,3)\mathcal{A}(5, 1, 3). 124567 138ace 239bdf 4ab 4cf 4de 589 5ad 5bc 68b 69a 6ef 78f 79e 7cd

N4N_4, type (15; 7, 7), sub-arrangement of A(5,1,3)\mathcal{A}(5, 1, 3). 5bcdef 6789af 12345 16b 17e 18d 1ac 27c 28e 29d 2ab 36c 37d 39b 3ae 46e 48b 49c 4ad

N5N_5, type (15; 7, 7), sub-arrangement of A(5,1,3)\mathcal{A}(5, 1, 3). 123456 1789a bcdef 27b 28d 29f 2ae 37e 38c 39d 3af 47f 48b 49c 4ad 57c 58f 59e 5ab 67d 68e 69b 6ac

Cuntz’s family. Following [7 §4.2], At\mathcal{A}_t is defined by the normal vectors

(1,0,0), (1,1,0), (1,0,1), (1,1,1), (1,t,1), (0,1,0), (2,1,1), (t+1,t,1),(1,0,0),\ (1,1,0),\ (1,0,1),\ (1,1,1),\ (1,t,1),\ (0,1,0),\ (2,1,1),\ (t+1,t,1),
(t+1,1,1), (2t,t,1), (1,1−t,1), (1−3t,t2−3t+1,−t),(t+1,1,1),\ (2t,t,1),\ (1,1-t,1),\ (1-3t,t^2-3t+1,-t),
(3t−1,t,t), (1−3t,−t2,−t), (3t−1,2t−1,t).(3t-1,t,t),\ (1-3t,-t^2,-t),\ (3t-1,2t-1,t).

References

Simion Stoilow Institute of Mathematics of the Romanian Academy, P.O. Box 1-764, RO-014700 Bucharest, Romania Email address: [email protected]

Simion Stoilow Institute of Mathematics of the Romanian Academy, P.O. Box 1-764, RO-014700 Bucharest, Romania Email address: [email protected]

References

  1. [1]T. Abe, Roots of characteristic polynomials and intersection points of line arrangements, J. Singul. 8 (2014), 100–116.arxiv.org/abs/1302.3822
  2. [2]T. Abe, Divisionally free arrangements of hyperplanes, Invent. Math. 204 (2016), 317–346.DOI
  3. [3]T. Abe, M. Cuntz, H. Kawanoue, T. Nozawa, Non-recursive freeness and non-rigidity, Discrete Math. 339 (2016), 1430–1449.DOI
  4. [4]M. Barakat, R. Behrends, C. Jefferson, L. Kühne, M. Leuner, On the generation of rank 3 simple matroids with an application to Terao’s freeness conjecture, SIAM J. Discrete Math. 35 (2021), 1201–1223.DOI
  5. [5]M. Barakat, L. Kühne, Computing the nonfree locus of the moduli space of arrangements and Terao’s freeness conjecture, Math. Comp. 92 (2023), 1431–1452; arXiv:2112.13065.DOI
  6. [6]R. Bojanowski, Zastosowania uogólnionej nierówności Bogomolova–Miyaoka–Yau [Applications of the generalized Bogomolov–Miyaoka–Yau inequality], Master’s thesis, University of Warsaw, 2003 (in Polish).
  7. [7]M. Cuntz, Free, not recursively free and non rigid arrangements, arXiv:1406.6154v1 (2014); incorporated into [ACKN16].arxiv.org/abs/1406.6154
  8. [8]A. Dimca, Curve arrangements, pencils, and Jacobian syzygies, Michigan Math. J. 66 (2017), 347–365.arxiv.org/abs/1601.00607
  9. [9]A. Dimca, Freeness versus maximal global Tjurina number for plane curves, Math. Proc. Cambridge Philos. Soc. 163 (2017), 161–172.DOI
  10. [10]A. Dimca, D. Ibadula, A. Măcinic, Freeness for 13 lines arrangements is combinatorial, Discrete Math. 342 (2019), 2445–2453.DOI
  11. [11]A. Dimca, L. Kühne, P. Pokora, Free line arrangements with low maximal multiplicity, J. Algebraic Combin. 63 (2026), Art. 45; arXiv:2505.01733.DOI
  12. [12]A. A. du Plessis, C. T. C. Wall, Application of the theory of the discriminant to highly singular plane curves, Math. Proc. Cambridge Philos. Soc. 126 (1999), 259–266.DOI
  13. [14]The FLINT team, FLINT: Fast Library for Number Theory, version 3.6.0 (via python-flint 0.9.0), 2026, https://flintlib.org.
  14. [15]A. Langer, Logarithmic orbifold Euler numbers of surfaces with applications, Proc. London Math. Soc. (3) 86 (2003), 358–396.
  15. [16]Y. Matsumoto, S. Moriyama, H. Imai, D. Bremner, Matroid enumeration for incidence geometry, Discrete Comput. Geom. 47 (2012), 17–43.DOI
  16. [17]B. D. McKay, Isomorph-free exhaustive generation, J. Algorithms 26 (1998), 306–324.DOI
  17. [18]B. D. McKay, A. Piperno, Practical graph isomorphism, II, J. Symbolic Comput. 60 (2014), 94–112.
  18. [19]P. Orlik, H. Terao, Arrangements of Hyperplanes, Grundlehren Math. Wiss. 300, Springer, Berlin, 1992.DOI
  19. [20]P. Pokora, The orbifold Langer–Miyaoka–Yau inequality and Hirzebruch-type inequalities, Electron. Res. Announc. Math. Sci. 24 (2017), 21–27.DOI
  20. [21]P. Pokora, Hirzebruch-type inequalities viewed as tools in combinatorics, Electron. J. Combin. 28 (2021), #P1.9.DOI
  21. [22]J. Stipins, Old and new examples of k-nets in ℙ², arXiv:math/0701046 (2007).arxiv.org/abs/math/0701046
  22. [23]H. Terao, Generalized exponents of a free arrangement of hyperplanes and Shephard–Todd–Brieskorn formula, Invent. Math. 63 (1981), 159–179.
  23. [24]A. Wakamiko, On the exponents of 2-multiarrangements, Tokyo J. Math. 30 (2007), 99–116.DOI
  24. [25]M. Yoshinaga, On the freeness of 3-arrangements, Bull. London Math. Soc. 37 (2005), 126–134.DOI

Paper details

vlnc:W_hmaz5kc2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents