Chapter 7

Constructions and Exact Values

Kakeya Sets with Multiplicity

7.1 Pencil complements

Proposition 7.1. Let TAG(2,q)T\subseteq AG(2,q) be ss-valid (at most ss lines of every direction meet TT), 1sq1\le s\le q. Then κqs(q)q2T\kappa_{q-s}(q)\le q^2-|T|. In particular, taking TT to be ss collinear points, κqs(q)q2s\kappa_{q-s}(q)\le q^2-s for 1sq1\le s\le q.

Proof. In each direction choose qsq-s lines that miss TT, which is possible since at most ss lines of that direction meet TT. The chosen lines form a (qs)(q-s)-fold configuration whose union avoids TT. ∎

For s=1s=1 this is the pencil complement of §4.3, the lines avoiding one point, with union of size q21q^2-1, attaining Theorem A. For s=q1s=q-1 it is the tautology κ1(q)q2T\kappa_1(q)\le q^2-|T| for TT the complement of a Kakeya set. Theorem 5.4 says that for sD(q)2s\le D(q)-2 the collinear choice is optimal.

7.2 Dual Denniston arcs and their sub-configurations

Proposition 7.2. Let q=2hq=2^h and nqn\mid q, n2n\ge2. The dual of a Denniston arc of degree nn (Theorem 2.4) with respect to its origin is an (n1)(n-1)-fold configuration attaining Theorem A, with all multiplicities nn. For every nnn'\mid n it contains an (n1)(n'-1)-fold sub-configuration attaining Theorem A. Hence κn1(q)=n1nq(q+1)\kappa_{n-1}(q)=\frac{n-1}nq(q+1) for all nqn\mid q.

Proof. Theorem 4.2 and Remark 2.5. ∎

The construction was carried out explicitly (Appendix A.2). For q=8q=8 and n=4n=4 one may take β=1\beta=1 (since Tr(1)=1\operatorname{Tr}(1)=1 in F8\mathbb F_8) and A={0,1,x,x+1}A=\{0,1,x,x+1\} in F8=F2[x]/(x3+x+1)\mathbb F_8=\mathbb F_2[x]/(x^3+x+1). The arc has 2828 points; its 6363 affine secants each carry 44 points and its other 99 affine lines none. The dual configuration has 2727 lines, 33 per direction, union of 5454 points, every one of multiplicity 44. For q=16q=16, n=4n=4, with β\beta of trace 11 and AA the span of 11 and xx: 5252 points, dual union 204204, all multiplicities 44.

Choosing r<n1r<n-1 lines per direction inside the dual arc gives an rr-fold configuration whose union is a subset of the arc's union, and one may minimise over the (n1r)q+1\binom{n-1}r^{q+1} choices. For q=8q=8, n=4n=4, r=2r=2 the minimum over the 393^9 choices is 5151, with spectrum {2:18, 3:24, 4:9}\{2{:}18,\ 3{:}24,\ 4{:}9\}; the exhaustive search over all 22-fold configurations of AG(2,8)AG(2,8) (Appendix A.1) also gives 5151, with a different spectrum {1:2, 2:12, 3:30, 4:7}\{1{:}2,\ 2{:}12,\ 3{:}30,\ 4{:}7\}. So κ2(8)=51\kappa_2(8)=51, and the dual Denniston arc contains an optimal 22-fold sub-configuration. For q=16q=16, n=4n=4, r=2r=2 the minimum over 3173^{17} choices is 194194, against the bound 18113181\frac13; whether κ2(16)=194\kappa_2(16)=194 is open.

7.3 Pencils of conics for odd qq

Theorem 7.3. Let qq be odd, βFq\beta\in\mathbb F_q a non-square, Q(x,y)=x2βy2Q(x,y)=x^2-\beta y^2, and for λ0\lambda\ne0 let Cλ={(x,y):Q(x,y)=λ}C_\lambda=\{(x,y):Q(x,y)=\lambda\}. Then:

(a) Each CλC_\lambda has q+1q+1 points, and the CλC_\lambda partition AG(2,q){O}AG(2,q)\setminus\{O\}, O=(0,0)O=(0,0).

(b) Every line through OO meets CλC_\lambda in 00 or 22 points. The set of lines through OO that meet CλC_\lambda depends only on whether λ\lambda is a square; the two classes give complementary sets of q+12\frac{q+1}2 lines each.

(c) Let rr be even and let Λ\Lambda consist of r/2r/2 non-zero squares and r/2r/2 non-squares. Then S=λΛCλS=\bigcup_{\lambda\in\Lambda}C_\lambda has exactly rr points on every line through OO, so its dual (with respect to OO) is an rr-fold configuration.

Proof. (a) QQ is the norm form of Fq2=Fq(β)\mathbb F_{q^2}=\mathbb F_q(\sqrt\beta) over Fq\mathbb F_q: N(x+yβ)=x2βy2N(x+y\sqrt\beta)=x^2-\beta y^2. The norm is a surjective homomorphism Fq2×Fq×\mathbb F_{q^2}^\times\to\mathbb F_q^\times with fibres of size q+1q+1, and Q(x,y)=0Q(x,y)=0 only at OO since β\beta is a non-square.

(b) On the line y=mxy=mx, Q=x2(1βm2)Q=x^2(1-\beta m^2), and 1βm201-\beta m^2\ne0 since β\beta is a non-square; x2(1βm2)=λx^2(1-\beta m^2)=\lambda has two solutions if λ(1βm2)\lambda(1-\beta m^2) is a non-zero square and none otherwise. On x=0x=0, Q=βy2=λQ=-\beta y^2=\lambda has two solutions iff λβ-\lambda\beta is a square. In both cases the condition is "λγ\lambda\cdot\gamma\in\square" for a fixed γ\gamma depending on the line, and replacing λ\lambda by λν\lambda\nu with ν\nu a non-square negates it. Since CλC_\lambda has q+1q+1 points and each line through OO meets it in 00 or 22, exactly q+12\frac{q+1}2 lines through OO meet it.

(c) By (b), each line through OO meets every conic of one of the two classes in 22 points and no conic of the other class; with r/2r/2 conics in each class it meets SS in 2r2=r2\cdot\frac r2=r points. The conics are disjoint by (a) and OSO\notin S. ∎

Remark 7.4. A line y=ax+cy=ax+c with c0c\ne0 meets CλC_\lambda iff the quadratic (1βa2)x22βacxβc2λ(1-\beta a^2)x^2-2\beta acx-\beta c^2-\lambda has a root, i.e. iff its discriminant 4(βc2+λ(1βa2))4\big(\beta c^2+\lambda(1-\beta a^2)\big) is a square or zero. Writing u=βc2u=\beta c^2 and v=1βa2v=1-\beta a^2, the line is skew to all of SS iff u+λvu+\lambda v is a non-square for every λΛ\lambda\in\Lambda. For rr "random-looking" values of λ\lambda one expects about a 2r2^{-r} fraction of the q2qq^2-q lines y=ax+cy=ax+c, c0c\ne0, to be skew, so U(12r)q2|U|\approx(1-2^{-r})q^2. For r=2r=2 this is 34q2\frac34q^2, above the bound 23q2\frac23q^2 by a constant factor: the construction is not asymptotically optimal. It is, however, exactly optimal at q=7q=7.

The minimum of U|U| over all admissible choices of Λ\Lambda was computed (Appendix A.3):

qqrrchoicesminU\min\lvert U\rvertbr(q)b_r(q)κr(q)\kappa_r(q)
524242021
72940371337\frac1340
74948444544\frac4546
9216706065\le65
9436807279\le79
112259688?
11410010810535105\frac35?
1323614012113121\frac13?

For q=7q=7, r=2r=2, the optimal pencil configuration has spectrum {2:16, 3:16, 4:8}\{2{:}16,\ 3{:}16,\ 4{:}8\}, different from the spectrum {1:2, 2:10, 3:22, 4:6}\{1{:}2,\ 2{:}10,\ 3{:}22,\ 4{:}6\} of the search's optimum; both have deficiency 83\frac83. So the 22-fold extremal problem for q=7q=7 has at least two inequivalent solutions.

7.4 The reach of the polynomial method

Proposition 2.7 says that no polynomial of degree less than qq vanishes on an rr-fold Kakeya set. The counting lemma of the polynomial method says that a nonzero reduced polynomial of degree at most dd vanishing on UU exists as soon as md>Um_d>|U|, where mdm_d is the number of reduced monomials of degree at most dd; write d(U)=min{d:md>U}d^\ast(U)=\min\{d:m_d>|U|\} for this trivial threshold and δ(U)\delta(U) for the least degree of a nonzero reduced polynomial vanishing on UU. Then qδ(U)d(U)q\le\delta(U)\le d^\ast(U). In this form the method proves Umd|U|\ge m_d for every dd with δ(U)>d\delta(U)>d: Proposition 2.7 gives δ(U)q\delta(U)\ge q, hence Umq1=(q+12)|U|\ge m_{q-1}=\binom{q+1}2, and more would follow only from a proof that δ(U)>q\delta(U)>q.

Proposition 7.5 (Theorem D). For each minimal configuration in the following table, δ(U)=d(U)\delta(U)=d^\ast(U); that is, the points of UU impose independent conditions on reduced polynomials of every degree below d(U)d^\ast(U).

qqrrU\lvert U\rvertmq1m_{q-1}mdm_d around dd^\astd(U)d^\ast(U)δ(U)\delta(U)kernel dimension at δ\delta
522115m5=19, m6=22m_5=19,\ m_6=22661
713228m7=34m_7=34772
724028m8=39, m9=43m_8=39,\ m_9=43993
734428m9=43, m10=46m_9=43,\ m_{10}=4610102
744628m10=46, m11=48m_{10}=46,\ m_{11}=4811112

Proof. Computation (Appendix A.5): for each dd the rank of the evaluation matrix of the reduced monomials of degree d\le d at the points of UU was computed over Fq\mathbb F_q; it equals min(U,md)\min(|U|,m_d) in every case, so the kernel is nonzero exactly when md>Um_d>|U|, with dimension mdUm_d-|U|. ∎

(The r=1r=1 row uses a 3232-point 11-fold configuration, one point above the minimum; the minimal 3131-point configurations behave the same way, with δ=7\delta=7 and kernel dimension 33.) The table shows that the "zeros" half of the polynomial method, which must show that UU cannot be the zero set of a polynomial of the degree that counting provides, has nothing to work with for r2r\ge2 beyond degree qq: the polynomials of degree d(U)>qd^\ast(U)>q that vanish on the optimal rr-fold unions exist for the trivial reason and there are exactly as many of them as counting predicts. To reproduce Theorem A by the polynomial method one would have to prove δ(U)dr(q)\delta(U)\ge d_r(q) with mdr(q)rr+1q2m_{d_r(q)}\approx\frac r{r+1}q^2. Since md=q2(2q1d2)m_d=q^2-\binom{2q-1-d}2 for dq1d\ge q-1, this means dr(q)q(22/(r+1))d_r(q)\approx q\big(2-\sqrt{2/(r+1)}\big), about 1.18q1.18q for r=2r=2: a statement about polynomials of degree well above qq vanishing on unions of lines, which is the territory of Rédei-type and lacunary polynomials (Rédei 1970; Ball, Blokhuis, and Mazzocca 1997) rather than of the restriction-to-lines argument.

7.5 Table of exact values

The following table collects everything known to me about κr(q)\kappa_r(q) for q9q\le9 and q=16q=16. Provenance: A equality in Theorem A (Corollary 4.3); C Theorem 5.4; E exhaustive search (Appendix A.1); BM Blokhuis and Mazzocca (2008); D dual Denniston sub-configuration (§7.2); a range gives the best lower and upper bounds known.

qqrrκr(q)\kappa_r(q)br(q)b_r(q)provenancespectrum of an extremal configuration
3176BM, E{1:3,2:3,3:1}\{1{:}3,2{:}3,3{:}1\}
3288A, E{3:8}\{3{:}8\}
411010A, E{2:10}\{2{:}10\}
4214131313\frac13C, E{2:4,3:8,4:2}\{2{:}4,3{:}8,4{:}2\}
431515A, E{4:15}\{4{:}15\}
511715BM, E{1:6,2:9,3:2}\{1{:}6,2{:}9,3{:}2\}
522120E{2:6,3:12,4:3}\{2{:}6,3{:}12,4{:}3\}
5323221222\frac12C, E{3:5,4:15,5:3}\{3{:}5,4{:}15,5{:}3\}
542424A, E{5:24}\{5{:}24\}
713128BM
7240371337\frac13E{1:2,2:10,3:22,4:6}\{1{:}2,2{:}10,3{:}22,4{:}6\}
734442E{2:3,3:11,4:21,5:9}\{2{:}3,3{:}11,4{:}21,5{:}9\}
7446444544\frac45C, E{3:3,4:9,5:25,6:9}\{3{:}3,4{:}9,5{:}25,6{:}9\}
7547462346\frac23C
764848A{7:48}\{7{:}48\}
813636A{2:36}\{2{:}36\}
825148E, D{1:2,2:12,3:30,4:7}\{1{:}2,2{:}12,3{:}30,4{:}7\}
835454A, D{4:54}\{4{:}54\}
8458 or 59573557\frac35A; C
8561602360\frac23C
8662615761\frac57C
876363A{8:63}\{8{:}63\}
914945BM
9261 to 6560Thm 6.1; search{1:2,2:24,3:26,4:13}\{1{:}2,2{:}24,3{:}26,4{:}13\} (65)
9368\ge68671267\frac12Thm 6.1
9473\ge7372Thm 6.1
957775C
9678771777\frac17C
9779783478\frac34C
988080A
161136136A
162182 to 19418113181\frac13A; D
163204204A, D
167238238A
1611–15256s256-sC (s5s\le5)
1616256256A

Every equality case of Theorem A in the table is one predicted by Corollary 4.3, and every case predicted by Corollary 4.3 with q9q\le9 or q=16q=16 appears as an equality (the trivial case r=qr=q, where UU is the whole plane, is listed only for q=16q=16).