Permutations minimizing the number of collinear triples

Using the inclusion-exclusion principle, we first add the number of lines through each point, then subtract off the number of lines through pairs of points, then add back number of lines through three points. By Bonferroni’s Inequality, the result is an upper bound for k^*(\Omega) (for any \Omega consisting of q+2 points, including an internal nucleus). So we end up with

On the other hand, by [blokhuis], the above quantity is also a lower bound, so we have equality.

If b\neq 0, then the discriminant of the quadratic obtained by solving for y is (cx+e)^2-4b(ax^2+dx+f), which must be 0, or else there are 0 or 2 solutions to ax^2+by^2+cxy+dx+ey+f=0 for some x, contradicting that \Gamma' is the graph of a function.

Expanding gives

This quadratic would need to hold for every x\in\mathbb{F}_q, except the x corresponding to the external point ignored in \Gamma'. That cannot be the case however when q\geq4, since quadratics only have at most 2 unique solutions, unless all coefficients are 0, i.e., 4ab=c^2, 2bd=ce, and 4bf=e^2. Multiplying the conic by 4b, the conic becomes

which is reducible, a contradiction. Thus, b=0. By an identical argument swapping x and y, we conclude that a=0.

For q \geq 5, we have that \Gamma' = \{(x,y): cxy + dx +ey + f = 0\} from Lemma 1. Simple manipulation gives

It is easy to check that the result also holds for q=3.

By Lemma 1, we have that \Omega = \{[x:y:z] : cxy + dxz +eyz + fz^2 = 0,\ (x,y,z)\neq(0,0,0)\} \cup \{ [x':y':z'] \}, where [x':y':z'] is the external point not on the conic. Clearly, [0:1:0] and [1:0:0] both are points that satisfy the equation cxy + dxz +eyz + fz^2 = 0, which therefore means [0:1:0] and [1:0:0] were on the conic, i.e., neither is the external point.

By Corollary 1.1 and Corollary 1.2, we have that

For simplicity, we can divide through by c, and then just relabel the coefficients to look like Definition 3. That is,

Then we are left with

where \alpha \gamma \neq \beta. We claim that (x',y') = (-\gamma,\alpha). The domain of (\alpha x+\beta)(x+\gamma) does not include -\gamma, so x'=-\gamma. We then can manipulate the fraction to show

Using the fact that \alpha\gamma \neq \beta, we have that \alpha + (\beta-\alpha\gamma)/(x+\gamma) is never equal to \alpha. Since \Gamma is the graph of a permutation, there must be a point (x,y)\in\Gamma where y=\alpha, so our external point is (-\gamma,\alpha).

Therefore \sigma is an ELFTP.

Let \sigma(x) be a permutation that admits (q-1)/2 collinear triples; then, by Theorem 2, it is an ELFTP. We argue that there is exactly one such permutation satisfying \sigma(0)=0, \sigma(1)=1, and \sigma(2)=2, from which it follows that it must be the lexicographic-least. There are 4 cases: \gamma\notin\{0,-1,-2\}, \gamma =0, \gamma =-1, and \gamma =-2.

Case 1: \gamma \notin\{0,-1,-2\}
Then \alpha x+\beta=\sigma(x)(x+\gamma ) for x \in \{0,1,2\}. That \sigma(0)=0 implies \beta =0. Since \sigma(1)=1 and \sigma(2)=2, it follows that \alpha =1+\gamma and 2\alpha=4+2\gamma, a contradiction.

Case 2: \gamma =0
Then \alpha x+\beta = x\sigma(x) for x \in \{1,2\}. Since \sigma(1)=1 and \sigma(2)=2, it follows that \alpha +\beta=1 and 2\alpha+\beta=4. This implies \beta =-2 and \alpha =3. But since \sigma(-\gamma)=\alpha, we may conclude that \alpha =\sigma(0)=0, a contradiction.

Case 3: \gamma =-2
Then \alpha x+\beta = \sigma(x)(x-2) for x \in \{0,1\}. Since \sigma(0)=0 and \sigma(1)=1, it follows that \beta =0 and \alpha +\beta=-1. This implies \alpha =-1. But since \sigma(-\gamma)=\alpha, we may conclude that \alpha =\sigma(2)=2, a contradiction.

Case 4: \gamma =-1
Then \alpha x+\beta = \sigma(x)(x-1) for x \in \{0,2\}. Since \sigma(0)=0 and \sigma(2)=2, it follows that \beta =0 and 2\alpha+\beta=2. This implies \alpha =1. Since \sigma(-c)=a, we may conclude that \sigma(1)=1. So \alpha =1, \beta =0, \gamma =-1 yields the unique permutation of form Definition 3 with (\sigma(0),\sigma(1),\sigma(2))=(0,1,2).

Thus, there is exactly one ELFTP \sigma that satisfies \sigma(0)=0, \sigma(1)=1, and \sigma(2)=2, and it is the desired lexicographically-least permutation.

Let \{L_1,\ldots,L_{q-1},H,V\} be a partition of \mathcal{P} into q+1 parallel classes of q lines each. Consider a generalized permutation S and its set \binom{S}{2} of unordered pairs. Since no such pairs are contained in an element of H or V, for all \{p_1,p_2\} \in \binom{S}{2}, there exists exactly one j \in [q-1] so that there exists an \ell \in L_j with \{p_1,p_2\} \subset L_j. By the pigeonhole principle, some L_j contains at least \lceil \binom{q}{2}/(q-1) \rceil= \lceil q/2 \rceil = (q+1)/2 pairs. These pairs cannot all be contained in distinct (and therefore disjoint) elements of L_j, or else we would have \geq 2 \cdot (q+1)/2 = q+1 points in S, a contradiction. Thus, two pairs are contained in some \ell \in L_j, which implies that \ell \cap S contains at least three points.