Finite Geometry and

Combinatorial Applications

Anurag Bishnoi
Preface
Incidence geometry is the study of abstract structures satisfying certain geometric axioms
inspired from the incidence properties of points, lines, planes, etc. For example, the
axiom that through any two points there is a unique line, or that any two lines intersect
in at most one point. In this course we will introduce several finite incidence structures
and explore how these structures interact with combinatorics. In particular, we will
study finite projective and affine spaces, generalized polygons and polar spaces. On the
combinatorial side we will study Latin squares, blocking sets, strongly regular graphs,
finite field Kakeya and Nikodym problems, etc.

Recommended Reading

References for the course

• “Finite Geometry and Combinatorial Applications” by Simeon Ball.

• “An Introduction to Incidence Geometry” by Bart De Bruyn.
• “Incidence Geometry” by G. Eric Moorhouse, http://math.ucr.edu/home/baez/
qg-fall2016/incidence_geometry.pdf.
• “Projective Geometry” by Rey Casse.
• “Algebraic Graph Theory” by Chris Godsil and Gordon Royle.

Further reading

• “Finite Geometries” by P. Dembowski.

• “Points and Lines” by E. Shult.
• “Projective Geometries over finite fields” by Hirschfeld and Thas.
• “Foundations of Incidence Geometry” by Johannes Ueberberg.
• “Projective Geometry: From Foundations to Applications” by Albrecht Beutelspacher
and Ute Rosenbaum.
• “Combinatorics of finite geometries” by Lynn Margaret Batten.
• “Distance Regular Graphs” by Brouwer, Cohen and Neumaier.
• “Designs, Graphs, Codes and their Links” by Peter J. Cameron and J. H. van Lint.
• “Spectra of Graphs” by Brouwer and Haemers.

Page iii
Background required

1 Linear Algebra
The reader should be familiar with the following notions from linear algebra: vector
spaces over a field, linear independence, span, basis of a vector space, the lattice of
subspaces, linear maps between vector spaces, linear forms, dual spaces, ranks of matrices,
determinants, quotient spaces, eigenvalues and eigenspaces.
The following basic theorems will often be directly useful to us.

Theorem 0.1. Let U1 , U2 be two finite dimensional subspaces of a vector space V . Then

dim U1 + dim U2 = dim(U1 ∩ U2 ) + dim(U1 + U2 ).

Theorem 0.2 (Rank-Nullity Theorem). Let L be a linear map from a vector space V to
a vector space W . Then

dim ker(L) + dim im(L) = dim V.

Theorem 0.3. Let U be a subspace of an n dimensional vector space V , and let V /U be

the quotient space with respect to U . Then dim V /U = n − dim U .

Theorem 0.4. Let A be a square matrix of order n. Then the eigenvalues of A are the
solutions of the degree n polynomial equation det(A − λI) = 0, where I is the identity
matrix.

Corollary 0.5. The sum of all eigenvalues of A is equal to the trace of A, that is, the sum
of the diagonal entries of A, and the product of all eigenvalues is equal to the determinant
of A.

Theorem 0.6. Let A be a real symmetric matrix of order n. Then A has n eigenvalues,
λ1 ≥ · · · ≥ λn , counted with multiplicity. Moreover, there exists an orthonormal basis of
Rn , x1 , . . . , xn , such that xi is an eigenvector of M with eigenvalue equal to λi .

2 Finite Fields
The reader should be familiar with the notions of groups, rings, ideals, fields, quotient
rings, isomorphism of rings/fields and field extensions.
The simplest example of a finite field is Fp = Z/pZ, that is, the integers modulo p,
for a prime number p. The main thing to check is that every element has an inverse,
which follows from Euclid’s theorem: for any integers a, b with GCD(a, b) = 1, there exist

Page v
integers x and y such that ax + by = 1. Let F be an arbitrary field which is finite. By the
repeated addition of 1, we see that there exists a smallest integer m such that 1 + · · · + 1
(m times) is equal to 0 in F . This smallest integer must necessarily be a prime p, since
there are no zero divisors in F , and this prime number is known as the characteristic of
the finite field F . The set {0, s1 , . . . , sp−1 } ⊆ F where si is the i-fold sum of 1, is a subfield
of F , which is known as the prime subfield, and it is isomorphic to Fp .

Theorem 0.7. For every finite field F , there exists a prime number p and a positive
integer n such that |F | = pn .
Proof. Let p be the characteristic of F , and let n be the dimension of F when seen as a
vector space over its prime subfield. The number of elements in an n-dimensional vector
space over Fp is equal to pn .

Therefore, every finite field must have order equal to a prime power. In fact, we have the
following converse.

Theorem 0.8. For every prime power q, there exists a unique field of order q (up to
isomorphism).
Proof. Let q = pn where p is a prime. Any finite field of characteristic p is algebraic over
the subfield Fp , since it is finite, and hence contained in the algebraic closure F of Fp .
The unique field of order q is then just the set of zeros of the polynomial xq − x in F .

This finite field of order q = pn is denoted by Fq .1 The proof above is not really construc-
tive. The most concrete way to construct the finite field Fq , and to do computations in
it, is to take an irreducible polynomial f (x) ∈ Fp [x] of degree n (they always exist), to
get the field Fq as Fp [x]/(f (x)).

We now enumerate some basic properties of Fq that will be useful to us.

(a) For every α ∈ Fq we have αq = α. Moreover, the elements of the prime subfield are
precisely the zeros of the polynomial xp − x ∈ Fq [x].

(b) Fpm is a subfield of Fpn if and only if m divides n. Moreover, for every divisor m of n,
there exists a unique copy of Fpm in the subfield lattice of Fpn , given by the elements
satisfying αp = α in Fpn .
m

(c) The additive group of Fq is isomorphic to (Z/pZ)n . The multiplicative group F∗q of
Fq , is generated by a single element of order q − 1, and therefore it’s isomorphic to
Z/(q − 1)Z. Any generator of this group is called a primitive element of Fq .

(d) The map x 7→ xp is an automorphism of the field Fq . The full automorphism group
of Fq is a cyclic group of order n consisting of the maps x 7→ xp , for i = 1, . . . , n.
i

(e) For q odd, there are precisely (q −1)/2 squares (quadratic residues) in F∗q and (q −1)/2
non-squares. Multiplying F∗q by a square fixes these two sets while multiplying by a
non-square permutes the two of them. For q even, every element of Fq is a square
since x 7→ x2 is a field automorphism.
1
In finite geometry literature it is sometimes denoted by GF(q), where GF stands for Galois Field.

Page vi
Lemma 0.9. For all i ∈ N, i 6= 0, the sum
X
ai
a∈Fq

is −1 if i is a multiple of q − 1 and 0 otherwise.

Proof. Let S be the sum and α be the generator of the group F∗q . Then
X X X
S= ai = (αa)i = αi ai = αi S.
a∈Fq a∈Fq a∈Fq

Now, αi is equal to 1 if and only if i is a multiple of q −1, and hence for all i 6∈ (q −1)N, we
must have S = 0 for the equality above to hold. When i ∈ (q − 1)N, then since aq−1 = 1
for all a 6= 0, S is the (q − 1)-fold sum of the identity 1, and hence it’s equal to −1.

Definition 0.10. Let Fqn be a field extension of Fq . Then the trace function, with respect
to this extension, is defined as
2 n−1
Tr(x) = x + xq + xq + · · · + xq .

The norm function with respect to this extension is defined as

n−1 n −1)/(q−1)
Norm(x) = xxq · · · xq = x(q .

If Fq is the prime subfield, then these maps are known as absolute trace and absolute
norm, respectively.

Lemma 0.11. Tr is an additive surjective function from Fqn to Fq with |Tr−1 (a)| = q n−1
for all a ∈ Fq .

Lemma 0.12. Norm is a multiplicative function from F∗qn to F∗q , with |Norm−1 (a)| =
(q n − 1)/(q − 1) for all a ∈ F∗q .

Page vii
Contents

Preface iii

Background required v
1 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
2 Finite Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

1 Affine and Projective Spaces 3

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Affine planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Mutually Orthogonal Latin Squares . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Projective planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.6 Projetive spaces from vector spaces . . . . . . . . . . . . . . . . . . . . . . 13
1.7 Desarguesian planes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Conics and Ovals 19

2.1 Coordinates and Hypersurfaces . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Conics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Ovals and Hyperovals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3 Some Combinatorial Questions 29

3.1 Blocking Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Kakeya and Nikodym Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Nikodym sets and large minimal blocking sets . . . . . . . . . . . . 36
3.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4 Strongly Regular Graphs 41

4.1 Introdction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Moore graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Spectral Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Two-intersection sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Generalized Quadrangles 53
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.2 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.3 Some examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.4 An application to Ball Packings . . . . . . . . . . . . . . . . . . . . . . . . 58
5.5 Generalized polygons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

Page 1
Contents

5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6 Polar Spaces 69
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Collinearity Graph of a Polar Space . . . . . . . . . . . . . . . . . . . . . . 71
6.3 Quadrics and Symmetric Bilinear Forms . . . . . . . . . . . . . . . . . . . 72
6.4 Sesquilinear Forms and Classical Polar Spaces . . . . . . . . . . . . . . . . 77
6.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

Page 2
1 Affine and Projective Spaces

1.1 Introduction
Incidence geometry is the study of mathematical structures that are formed by abstracting
out the notion of incidence from Euclidean geometry. In this course we will be studying
some finite incidence structures, and how they interact with combinatorics. Our main
objects of study are the following.
Definition 1.1 (Point-Line Geometries). A point-line geometry is a triple S = (P, L, I)
where P is a non-empty set, L is a set disjoint from P, and I is a subset of P × L such
that for every ` ∈ L, there exists at least two x ∈ P such that (x, `) ∈ I. The elements of
P are referred to as points of S, elements of L as lines of S and I is the incidence relation
between points and lines.
Example 1.2. The Euclidean plane is one of the classical examples of a point-line geometry,
where P = R2 is the set of ordered pairs (x, y) of real numbers, L is the set of all straight
lines, that is, subsets of P of the form {(x, mx + c) : x ∈ R} or of the form {(c, y) : y ∈ R}
where m, c ∈ R, and the incidence relation I is simply set containment.
Example 1.3. Here is a finite example of a point-line geometry. Let P = [4], L = [4] and


2
I set containment (see Figure 1.1).

While these two examples might look quite different, they both belong to the family of
affine planes, which will be discussed in the next section.

4 3

1 2

Figure 1.1: Affine plane of order 2

Example 1.4. Every multi-graph without a loop is a point-line geometry, where each line
is incident with precisely two points. Every hypergraph1 where each edge contains at least
two vertices is a point-line geometry.

We will often use geometrical language and write the statement “(x, `) ∈ I” as “the point
x lies on the line `”, or “` is a line through x”, etc. Two points will be called collinear
if there is a common line through both of them, and two lines will be said to meet each
other if there is a common point incident to both of them.
1
A hypergraph is a pair (V, E) where V is a non-empty set and E is a collection of subsets of V .

Page 3
Affine and Projective Spaces

Definition 1.5. A point-line geometry is called a partial linear space if through every
pair of distinct points there is at most one line. It is called a linear space if there is exactly
one line through every pair of points.

Note that two distinct lines in a partial linear space intersect each other in at most one
point, and hence these lines do behave like the lines that we are used to in Euclidean
geometry. In a linear space, we will denote the unique line through two distinct points
x, y by xy.
For partial linear spaces, we can uniquely identify each line with the subset of points it
is incident with. Thus, we can think of these point-line geometries as a hypergraph, with
the incidence relation as set containment. In fact, partial linear spaces are equivalent to
the so-called linear hypergraphs.

1.2 Affine planes

Abstracting out the incidence properties of the Euclidean plane, we get the following
structure.

Definition 1.6. An affine plane is a linear space with the following properties:
(A1) (Playfair axiom) For every point x and a line ` not through x, there exists a unique
line `0 through x that does not meet `.
(A2) There exist three non-collinear points.

It can be easily checked that the point-line geometries from Example 1.2 are both affine
planes.

Definition 1.7. In a affine plane, we say that two lines ` and m are parallel, denoted by
` k m, if either ` = m or ` and m do not meet each other.

Proposition 1.8. Parallelism is an equivalence relation on the lines of an affine plane,

and each parallel class is a partition of the set of points.
Proof. The fact that it is an equivalence relation follows easily from the definition. Now
take a parallel class S and a point x of S. Let ` be a line of S. Then there exists a unique
line through x which is parallel to `, and hence a unique line of S through x. Since x was
arbitrary, this shows that the point set is partitioned by S.

Theorem 1.9. For every finite affine plane there exists an integer n ≥ 2, called the order
of the plane, such that:
(a) Each point is incident with exactly n + 1 lines.
(b) Each line is incident with exactly n points.
(c) Each parallel class has n lines.
(d) There are n + 1 parallel classes.
(e) There are n2 points.

Page 4
 Affine and Projective Spaces

(f ) There are n2 + n lines.

Proof. (a) Since we have a finite affine plane, the number of lines through each point is
finite. Let x 6= y be two points of the plane. From (A1) it follows that for every line
` through x there is a unique line m through y which is parallel to `. This gives a
bijection between lines through x and lines through y. Define n to be such that n + 1
is the common number of lines through every point. We now show that n ≥ 2.

Pick three non-collinear points x, y, z, which exist by (A2). Besides the lines xy and
xz through x, there is the unique line parallel to yz through x, giving us at least three
lines through x. Therefore n + 1 ≥ 3.

(b) Let ` be a line and x a point not in `. Such a point exists since each line has at least
two points on it, and through each point of ` there is at least one other line (recall
that n ≥ 2). There is a unique line through x which is parallel to `, whereas the
remaining n lines through x intersect ` in a unique point. These are all the points
on ` since for every point y on ` the line joining x and y is a line through x that
intersects `.

(c) Let S be a parallel class and ` ∈ S. Take a point x on ` and a line m 6= ` through x.
Through each of the n − 1 points on m other than x, there is a unique line parallel
to `, giving us |S| ≥ n. Moreover, every line in S must intersect m in a point since
m 6∈ S, giving us |S| ≤ n.

(d) Let x be a point. For every parallel class S, there exists a unique line through x which
is contained in S. Moreover, there are n + 1 lines through x, giving us n + 1 different
parallel classes.

(e) Since a parallel class partitions the set of points, and there are n lines in a parallel
class with each line containing n points, we have n2 points in total.

(f ) The n + 1 parallel classes, each containing n lines, partition the set of lines, giving us
n(n + 1) lines in total.

Example 1.10. Let F be any field. The point-line geometry AG(2, F ), with point set equal
to F 2 and lines as sets of the form {(x, y) : x = c, y ∈ F } and {(x, y) : y = mx + c, x ∈ F }
for m, c ∈ F , is an affine plane. A parallel class is either the set of “vertical” lines or the
set of lines with a common slope m. If F is a finite field of order q, then we get a finite
affine plane, which will be denoted by AG(2, q). See Figure 1.2 for a drawing of AG(2, 3),
where for example, the lines AIE, CF H and BDG form a parallel class.
Remark 1.11. An alternate way of describing the affine plane AG(2, F ) is as follows: the
points are the elements of the vector space F 2 and the lines are the cosets (with respect
to vector space addition) of 1-dimensional subspaces of F 2 .

With this background on affine planes, we are ready to see our first connection between
finite geometry and combinatorics.

Page 5
Affine and Projective Spaces

A H G

B I F

C D E

Figure 1.2: Affine plane over F3

1.3 Mutually Orthogonal Latin Squares

Given a set S of n elements, a Latin square L is a function L : [n] × [n] → S, i.e., an n × n
array with elements in S, such that each element of S appears exactly once in each row
and each column. For example,
1 2 3
3 1 2
2 3 1
is a 3 × 3 Latin square with [3] as its set of elements. Sudoku puzzles are examples of
9 × 9 Latin squares, with the extra constraints that in each of the nine 3 × 3 subgrids all
the symbols are distinct. The multiplication table of a finite group is also a Latin square.
Puzzle: From a standard deck of playing cards, take the aces, kings, queens and jacks.
Can you arrange these 16 cards in a 4 × 4 grid such that in each row/column no two cards
share the same suit or the same face value?
Solution: Here is one possible arrangement.

J♥ Q♦ K♠ A♣
Q♣ J♠ A♦ K♥
K♦ A♥ J♣ Q♠
A♠ K♣ Q♥ J♦

Definition 1.12. Let L1 and L2 be two Latin squares over the ground sets S1 , S2 ,
respectively. They are called orthogonal if for every (x1 , x2 ) ∈ S1 × S2 there exists a
unique (i, j) ∈ [n] × [n] such that L1 (i, j) = x1 and L2 (i, j) = x2 .

Page 6
 Affine and Projective Spaces

For example, the following are two orthogonal Latin squares of order 3.

1 2 3 2 3 1
3 1 2 3 1 2
2 3 1 1 2 3

Can you find another Latin square of order 3 which is orthogonal to both of the Latin
squares above? At most how many mutually orthogonal Latin squares (MOLS) of order
n can there be, in terms of n? This is somewhat similar the question “how many pairwise
orthogonal (non-zero) vectors can there by in Rn ?”.
Lemma 1.13. Let L be a Latin square with S as its ground set, and let π : S → S 0 be a
bijection from S to another set S 0 . Then π(L) defined by π(L)(i, j) = π(L(i, j)) is a Latin
square on the ground set S 0 .
Lemma 1.14. Let L1 , L2 be two orthogonal Latin squares on base sets S1 and S2 . Let
π1 : S1 → S10 , π2 : S2 → S20 be two bijections. Then π1 (L1 ) and π1 (L2 ) are also orthogonal
Latin squares.
Proof. Let (x, y) ∈ S10 × S20 . Let (i, j) be the unique index for which L1 (i, j) = π1−1 (x)
and L2 (i, j) = π2−1 (y). Then (i, j) is also the unique index for which π1 (L1 )(i, j) = x and
π2 (L2 )(i, j) = y.
Theorem 1.15. For every n, there exist at most n − 1 mutually orthogonal Latin squares
of order n.2
Proof. Let L1 , . . . , Lk be a set of MOLS of order n. Since changing the ground set
does not affect orthogonality, we may assume that all of them have elements from [n].
Moreover, applying a permutation to the symbols of any of the Latin squares does not
affect orthogonality of the system. So, we may assume that the first row of each Li is 1
2 · · · n. Now say k ≥ n. Since Li (2, 1) ∈ {2, . . . , n} for all i, by the pigeonhole principle
there exist i, j ∈ [k] such that Li (2, 1) = Lj (2, 1) = a for some a ∈ {2, . . . , n}. But then
(a, a) appears more than once in the superimposition of Li and Lj , showing that they are
not orthogonal to each other.

We now have the natural extremal question, is this bound sharp? Can we find n − 1
mutually orthogonal Latin squares for every n? The answer to the former is yes, while to
the latter is no, in the following sense.
Theorem 1.16 (Bose 1938). There exist n−1 mutually orthogonal Latin squares of order
n if and only if there exists an affine plane of order n.
Proof. We will show the sharpness of the bound on MOLS whenever there exists an
affine plane of order n. The other direction is left o the reader.
Let A be an affine plane of order n, and let S1 , . . . , Sn+1 be its parallel classes of lines.
Enumerate each Si as {`i,1 , . . . , `i,n }. Denote the unique point in the intersection of `n,i
and `n+1,j by (i, j). This gives us the domain [n] × [n] for the Latin squares we will
construct. For 1 ≤ k ≤ n − 1, define the k’th Latin square Lk by Lk (i, j) = m where m is
the unique index for which a line `k,m of the k-th parallel class contains the point (i, j).
We claim that this gives us n − 1 MOLS of order n (all of them with the ground set equal
to [n]).
2
Compare this to the statement that in Rn there are at most n mutually orthogonal vectors.

Page 7
Affine and Projective Spaces

Firstly, let Lk be one of the arrays. If there was a repetition in a row of Lk , then the line
of Sk corresponding to this repeated entry intersects the line of Sn corresponding to the
row in more than one points, which is a contradiction to the fact that an affine plane is a
partial linear space. We get a similar contradiction for repetitions in columns, and hence
Lk is a Latin square.

Now let Lk and Lk0 be two Latin squares, corresponding to the parallel classes Sk and Sk0 ,
respecitvely. Let (a, b) ∈ [n] × [n]. Then the unique point of intersection of the lines `k,a
and `k0 ,b gives the unique coordinate (i, j) for which Lk (i, j) = a and Lk0 (i, j) = b.

Example 1.10 gives us an affine plane of order q for each prime power q. This is equivalent
to saying that we for all prime powers q, there exists q − 1 mutually orthogonal Latin
squares of order q. Are there more orders for which we have an affine plane? This is one
of the oldest and central open problems in finite geometry.

Conjecture 1.17 (Prime power conjecture). Every finite affine plane has order equal to
a prime power.

The only known restriction that is known on the order of an affine plane, that rules out
infinitely many values, is as follows.

Theorem 1.18 (Bruck-Ryser 1949). If a finite affine plane of order n congruent to 1 or 2

modulo 4 exists, then n is the sum of two integral squares. In particular, any prime factor
of n of the form 4k + 3, for some k ∈ N, has an even exponent in the prime factorisation
of n.

Corollary 1.19. There exist no affine planes of order 6, 14, 21, 22, 30, 33 . . . .

Beyond this, there is a celebrated result of Lam, Thiel and Swiercz from 1989 which
showed that there is no affine plane of order 10. 3

The non-existence of an affine plane of order 6 was already known before the Bruck-Ryser
theorem. It was a famous problem of Euler to find two orthogonal Latin squares of order
6, who conjectured in 1782 that this can’t be done.4 Tarry proved in 1901 that Euler
was correct, and his result in particular implies that there are no affine planes of order
6. Euler in-fact made a stronger conjecture, that for every n ≡ 2 (mod 4), there does
not exist a pair of orthogonal Latin squares. In the homework you’ll see that for other
values of n one can always construct orthogonal Latin squares of order n. This conjecture
(whose only evidence was the trivial n = 2 case, and an incomplete proof of n = 6) was
proven to be false in 1959 by Parker, Bose, and Shrikhande, a.k.a., the Euler’s Spoilers 5 .

Theorem 1.20 (Bose-Shrikhande-Parker 1960). For every n > 6, there exist two orthog-
onal Latin squares of order n.
3
See http://web.thu.edu.tw/wang/www/emcc_Helly/Lam_finite_Proj_plane_order_10.pdf for
the story behind this result.
4
It is known as the thirty-six officers puzzle.
5
A phrase coined by the New York Times, who published the news of this result on their front page in
a Sunday edition.

Page 8
 Affine and Projective Spaces

1.4 Projective planes

As we saw before, one of the main features of affine planes is the notion of parallelism. In
some situations this feature can be a deficiency. This is “fixed” by the point-line geometries
known as projective planes, which play a key role in Finite Geometry, Combinatorics, and
Algebraic Geometry.

Definition 1.21. A projective plane is a linear space with the following properties.

(P1) Every two lines intersect in a unique point.

(P2) There exists a set of four points, no three of which are collinear.

Say we have an affine plane and let Π be set of parallel classes of the plane. For a line ` of
the affine plane, let Π(`) denote the unique parallel class it belongs to. For each parallel
class S ∈ Π, introduce a single new point σ(S), with different parallel classes getting
different points. Extend each of the original line ` by adding to it the new point σ(Π(`)).
Introduce a new line `∞ = {σ(S) : S ∈ Π}6 consisting of all the new points that have
been introduced. Then we claim that the point-line geometry that we get is a projective
plane.
Proof. Let (P, L, I) be the point line geometry that we get by the process above, so that
the set of points of the affine planes is P \`∞ and the set of lines is {`\`∩`∞ : ` ∈ L\{`∞ }}.
Every two lines of L corresponding to affine lines in the same parallel class S intersect
at the point σ(S). Every two lines corresponding to two non-parallel affine lines intersect
each other in their point of intersections in the affine plane and nowhere else, as the new
points on them are not the same. Every line corresponding to a line of the affine plane
intersects `∞ intersect in a single point, corresponding to the unique parallel class that
the affine line belongs to. Therefore, (P, L, I) is a linear space which satisfies (P1).

Now let x, y, z be three points of the affine plane which do not lie on a common line. Let
` be the line in L corresponding to the unique line through x parallel to the line yz, and
let w be any other point on `. Then no three points in the set x, y, z, w are incident to
the same line. Therefore, (P2) is also satisfied.

The projective plane that we get from the affine plane AG(2, F ), is denoted by PG(2, F ).
The points of PG(2, F ) are the points (x, y) of AG(2, F ), the points corresponding to the
lines of slope m ∈ F , which we denoted by (m), and the point (∞) corresponding to the
parallel class of vertical lines. The lines are the lines of AG(2, F ), extended with the point
corresponding to the slope of the line, and a new line `∞ = {(m) : m ∈ F ∪ {∞}}.

The finite projective plane of order q that we get when F = Fq is denoted by PG(2, q).
The projective plane PG(2, 2) is known as the Fano plane, named after the Italian math-
ematician Gino Fano, and it turns out that it is the unique projective plane of order 2
(see Figure 1.4).

We can reverse the process of constructing a projective plane from an affine plane, as
follows.
6
We tend to denote this line as `∞ because of the case of real perspective geometry, where one can think
of the parallel lines meeting each other at a far away point, that is “at infinity”.

Page 9
Affine and Projective Spaces

Figure 1.3: The Fano plane

Proposition 1.22. For any projective plane, removing a line of the plane and all points
on this line gives rise to an affine plane.
Proof. Let ` be the line which is removed. We still have a linear space, so all that needs
to be checked are the axioms (A1) and (A2). Let x be a point not on `, and m a line
distinct from ` such that x 6∈ m. Let y = ` ∩ m. Then every line through x intersects m
in a point outside `, except for the unique line xy, which is then the unique line through
x parallel to m.
Let x, y, z, w be four points in the projective plane, such that no three of them are incident
to the same line. If ` contains at most one of these four points, then (A2) is satisfied by
the remaining three points. WLOG say z, w ∈ `. Let p be the point of intersection of the
lines xz and yw. Then p 6∈ `, and p 6∈ xy. Therefore, p, x, y satisfy the condition in (A2),
thus proving that we have an affine plane.

We can use this relationship between affine and projective planes to obtain the following
useful result on finite projective planes.

Proposition 1.23. For every finite projective plane, there exists an integer n ≥ 2, called
the order of the plane, such that:
(a) Each line is incident with n + 1 points.
(b) Each point is incident with n + 1 lines.
(c) There are n2 + n + 1 points.
(d) There are n2 + n + 1 lines.
Proof. Remove a line `∞ from the projective plane to get a finite affine plane. Let n ≥ 2
be the order of the affine plane, from Theorem 1.9. Each affine line is incident with n
points, and hence adding the point at infinity gives us n + 1 points on each line in the
projective plane. Through every affine point there are n + 1 lines, and if we have a point
at infinity then there are n affine lines through it, corresponding to a parallel class, and
the line `∞ . There are n2 affine points and n + 1 points at infinity. There are n2 + n affine
lines and one `∞ .

By the discussion above the existence question of finite projective planes of a given order
n is equivalent to the existence of an affine plane of order n, in particular, we again have
the prime power conjecture which states that all finite projective plane must have their
order equal to a prime power, and the Bruck-Ryser theorem.

Page 10
 Affine and Projective Spaces

Figure 1.4: Degenerate projective plane

Projective planes play an important role finite geometry and combinatorics, as they often
turn out to be extremal objects in many theorems. We will see one such examples below.

If we remove (P2) from the axioms of a projective plane, and add the condition that we
have at least two lines, then it can be shown that the only new possibility we get is the
following point line geometry, which we call the degenerate projective plane (see Figure
1.4). The points set contains a point p and the line set contains a line ` such that every
point except p lies on ` and every line except ` passes through p.

Theorem 1.24 (De Bruijn-Erdős). Let (P, L, I) be a finite linear space, which has at
least 2 lines. Then |P| ≤ |L|, with equality if and only if the linear space is a (possibly
degenerate) projective plane.
Proof. (due to Conway/Motzkin) Note that in both degenerate projective planes and
non-degenerate projective planes the number of points is equal to the number of lines.
We will show that if |L| ≤ |P|, then |L| = |P| and we have a possibly degenerate projective
plane. For this it suffices to show that every two lines intersect.

Let b = |L| and v = |P| and assume that b ≤ v. For every x ∈ P let nx denote the
number of lines through x, and for every ` ∈ L let m` denote the number of points on `.
Let A = (P × L) \ I be the set of point-line pairs (x, `) such that x is not incident with
`.7 . Then from the fact that we have a linear space, we get nx ≥ m` for all (x, `) ∈ A, as
for every point y on the line ` we get a unique line xy through x. This implies
1 1
≥ ,
b − nx b − m`
for all (x, `) ∈ A. Note that since there are at least two lines, and each line contains
at least two points, the set A is non-empty. By summing over all anti-flags, we get the
inequality
X 1 X 1
≥ ,
b − nx b − m`
(x,`)∈A (x,`)∈A

which can be simplified to the following by noting that for each x there are exactly b − nx
lines not incident with x and for each ` there are exactly v − m` points not incident with
`,
X b − nx X v − m`
≥ .
x∈P
b − n x
`∈L
b − m `

7
These are called anti-flags in the point-line geometry

Page 11
Affine and Projective Spaces

The left hand side is equal to v whereas each term in the right hand side is at least v/b
since we assumed v ≥ b, and hence the right hand side is greater than or equal to v. We
must have equality everywhere, which implies that v = b and for all (x, `) ∈ A we have
nx = m` . In particular this means that any two pair of lines must intersect, as otherwise
we’ll get an (x, `) ∈ A for which nx > m` . Therefore, the linear space must be a possibly
degenerate finite projective plane.

This result also has the following consequence on discrete sets of points in Euclidean
spaces.

Corollary 1.25. Let S be a set of n points in R2 . Then S determines at least n lines,

where a line is said to be determined by S if it contains at least two points of S.
Proof. The set S along with the lines it determines forms a finite linear space. Moreover,
it is known that no finite projective plane can be embedded in R2 using straight lines.8
Therefore, we must have a degenerate projective plane in case of equality.

1.5 Duality

Another reason to study projective planes is that there is a “duality” between the points
and lines, in the following sense.

Definition 1.26. Let S = (P, L, I) be a point-line geometry such that every point
is incident with at least two lines. Then the dual of of S is the point-line geometry
S D = (P D , LD , I D ) where P D = L, LD = P and (`, x) ∈ I D if and only if (x, `) ∈ I.

Theorem 1.27. Let π be a projective plane. Then π D is also a projective plane.

Proof. The first axiom is satisfied in π D . For the second one, let x, y, z, w be four points
in π such that no three of them are collinear. Then the lines xy, yz, zw and wx are four
lines of π such that no three of them are concurrent, giving us the second axiom in π D .

Theorem 1.28 (Principle of duality). If a theorem is valid for all projective planes, then
the dual theorem obtained by interchanging the notions of point and line is also valid for
all projective planes.

Definition 1.29. An isomorphism between point-line geometries S = (P, L, I) and S 0 =

(P 0 , 0 L, I 0 ) is a pair of bijections f1 : P → P 0 and f2 : L → L0 such that (x, `) ∈ I if and
only if (f1 (x), f2 (`)) ∈ I 0 . We say that S is isomorphic to S 0 , denoted by S ∼ = S 0 if there
exists an isomorphism between S and S 0 .

Definition 1.30. A point-line geometry S is called self-dual if S ∼

= SD.

For example, the graph Cn gives a self-dual point-line geometry. We will soon see that
PG(2, F ) is self-dual for every field F .
8
Try to prove it!

Page 12
 Affine and Projective Spaces

1.6 Projetive spaces from vector spaces

Projective geometry pre-dates the concept of a vector spaces, but since linear algebra
has become ubiquitous in the world of mathematics it has become standard to define
projective planes (and higher dimensional spaces) using vector spaces. However, this only
defines a particular class of projective planes, whereas for higher dimensional spaces this
definition is equivalent to the axiomatic definition of a projective space.

Theorem 1.31. Let V be the 3-dimensional vector space over a field F . Let P be the set
of 1-dimensional subspaces of V , and L the set of 2-dimensional spaces. Then (P, L, I)
with I as the subspace relation, is a projective plane.
Proof. Any two distinct 1-dimensional subspaces are contained in a unique two dimen-
sional subspace, which is the subspace spanned by them. Let U1 , U2 be two 2-dimensional
subspaces. Since U1 + U2 is a subspace of V , we have dim(U1 + U2 ) = dim U1 + dim U2 −
dim U1 ∩ U2 ≤ 3, which implies that dim U1 ∩ U2 ≥ 1 since dim U1 = dim U2 = 2. More-
over, since U1 6= U2 , we must have dim U1 ∩ U2 < 2, and hence U1 ∩ U2 is the unique
1-dimensional subspace contained in both U1 and U2 .
The 1-dimensional subspaces spanned by the vectors (1, 0, 0), (0, 1, 0), (0, 0, 1) and (1, 1, 1)
satisfy (P2), since the matrix formed by taking these vectors as columns has rank 3 over
any field F .

Lemma 1.32. The projective plane constructed above from the vector space F 3 is iso-
morphic to PG(2, F ).
Proof. We adopt a geometrical model of F 3 where the 1-dim subspaces are the lines
through origin and the 2-dim subspaces are the planes through origin. Let π be the plane
defined by the equation z = 1 in AG(3, F ). We know that π is isomorphic to AG(2, F ).
Every 1-dim subspace, except for those lying in the plane z = 0, intersect π in a unique
point. Each 1-dim subspace contained in z = 0 naturally corresponds to a parallel class in
π, and hence to the point at infinity in the projective completion of PG(2, F ), as it gives
a direction for the lines in π. Every 2-dim subspace of F 3 , except for the space z = 0,
intersects π in a line.

Corollary 1.33. PG(2, F ) is self-dual.

Proof. Any isomorphic between the vector space F 3 and its dual gives us an isomorphism
between PG(2, F ) and PG(2, F )D .

The projective planes arising from 3-dimensional vector spaces immediately suggest the
following higher dimensional point-line geometry known as a projective space.

Definition 1.34. Let V be an n + 1 dimensional vector space over a field F . Then the n
dimensional projective space over F , PG(n, F ) is the point-line geometry where the points
are the 1-dimensional subspaces of V and the lines are the 2-dimensional subspaces of V .9

We also have an axiomatic definition of projective spaces.

9
Note that this n has nothing to do with the earlier n which was the order of a finite projective plane.

Page 13
Affine and Projective Spaces

Definition 1.35. An abstract projective space is a linear space satisfying the following
properties:
(PS1) Every line is incident with at least three points.
(PS2) Let `1 , `2 be two lines through a point z, and for i ∈ {1, 2} let xi , yi be two distinct
points on `i \ {z}. Then the lines x1 x2 and y1 y2 meet in a point.

Starting from this definition, one can build up the notion of subspaces, basis, and dimen-
sion, without relying on any linear algebra. See for example the book by Beutelspacher
and Rosenbaum, “Projective Geometry: From Foundations to Applications". We will not
do so here, and we justify that by referring to the following important result.

Theorem 1.36 (Veblen-Young, 1910). An abstract projective space is either a point, a

line, a projective plane, or the n dimensional projective space over a (skew) field10 F , for
some n ≥ 3.

In particular, in an abstract projective space if we assume that there exist two disjoint
lines then it must be isomorphic to PG(n, F ) for some (skew) field F and n ≥ 3. So,
all we have to deal with are the projective planes, which we defined axiomatically, and
projective spaces over vector spaces of dimension at least 4.
Even though the point-line incidences are sufficient to describe higher dimensional projec-
tive spaces, they have more structure than just points and lines. A k-dimensional subspace
of PG(n, F ) is the set of points, i.e., 1-dimensional subspaces of the underlying vector
space F n+1 , contained in a (k + 1)-dimensional subspace of F n+1 . The 0, 1, 2, 3, . . . , n − 1
dimensional subspaces of PG(n, F ) are known as points, lines, planes, solids, . . . , hyper-
planes, respectively, and the correspond to the 1, 2, 3, 4, . . . , n dimensional subspace of the
underlying vector space F n+1 .
If we remove a hyperplane from PG(n, F ) and all points incident with it, then we get the n-
dimensional affine space AG(n, F ). The point-line geometry AG(n, F ) can also be defined
without any reference to the projective spaces as the point-line geometry whose points are
the elements of the vector space F n and lines are the cosets of the 1-dimensional subspaces.
The k-dimensional subspaces are then the cosets of the k-dimensional subspaces of the
vector space F n . The equivalence relation of parallelism defined on the hyperplanes, i.e.,
the n − 1 dimensional subspaces of AG(n, F ), gives us the projective space PG(n, F ) by
adding a hyperplane at infinity, in the same was as it did for n = 2. When F = Fq , we
denote these geometries by AG(n, q) and PG(n, q), respectively.

1.7 Desarguesian planes

We can characterise the projective planes PG(2, F ) among the class of abstract projective
planes via the following theorem.

Theorem 1.37 (Desargues’ Theorem). Let F be a field. Let a1 b1 c1 and a2 b2 c2 be two

triangles in PG(2, F ) such that the lines a1 a2 , b1 b2 and c1 c2 all pass through a common
point. Then the points x = b1 c1 ∩ b2 c2 , y = c1 a1 ∩ c2 a2 and z = a1 b1 ∩ a2 b2 are collinear.

Page 14
 Affine and Projective Spaces

a2

a1

b2
b1

c2 c1

x y
z

Figure 1.5: Desargues’ theorem

Theorem 1.38. Any projective plane in which the Desargues’ theorem holds true is iso-
morphic to PG(2, F ) for some division ring F .
Remark 1.39. It is a nice exercise to prove that in any abstract projective space that
contains two disjoint lines, Desargues’ theorem is true. This then implies Theorem 1.36.

It is a classical result of Wedderburn, from 1907, which then states that any finite division
ring is a finite field, giving us the following.
Corollary 1.40. Any finite projective plane in which Desargues’ theorem holds is iso-
morphic to PG(2, q) for some prime power q.

Therefore, the planes PG(2, q) defined over the finite field Fq are known as the Desargue-
sian projective planes. There are many known families of finite projective planes which
are non-Desarguesian. In fact there exists an infinite sequence of n for which the number
of non-isomorphic projective planes of order n grows with at least a super polynomial
function of n.
While the constructions of finite non-Desarguesian planes are fairly involved, and you
will see some once we have developed more theory, it is fairly easy to construct a non-
Desarguesian projective plane over R: Let P = R2 , and let L consist of the same vertical
and positively sloped lines of the Euclidean plane, and the negatively sloped lines replaced
by the following, {(x, y) : y = mx + c if x ≥ 0 and y = 21 mx + c if x ≤ 0}. Then it can
be checked that (P, L) is a non-Desarguesian affine plane11 , and its projective completion
10
Skew fields, a.k.a. division rings, satisfy all axioms of a field except possibly the commutativity of
multiplication.
11
What is the corresponding statement of Desargues’ theorem for affine planes?

Page 15
Affine and Projective Spaces

gives us a non-Desarguesian projective plane.

1.8 Exercises

1. Prove that up-to isomorphism there are unique projective plane of order 2 and 3.12
2. Prove that a degenerate projective plane is self-dual.
3. (a) Let S be a linear space on n2 points where each line is incident with exactly n
points, for some integer n ≥ 2. Prove that S is an affine plane.
(b) Prove that if a system of n − 1 mutually orthogonal Latin squares of order n
exists, then there exists an affine plane of order n.
4. (a) Given two Latin squares of orders m and n, construct a Latin square of order mn.
(b) Given two orthogonal Latin squares of order m, and two orthogonal Latin squares
of order n, construct two orthogonal Latin squares of order mn.
(c) Prove that for every n 6≡ 2 (mod 4), there exist two orthogonal Latin squares of
order n.
(Hint: In part (a) try to construct the new Latin square over the ground set S1 × S2 ,
where S1 is the ground set of the first Latin square and S2 of the second Latin square.)
5. For n ≥ 2 an integer, define a graph Gn whose vertex set is equal to the set of all
n × n Latin squares on the ground set [n], and two Latin squares are adjacent if they
are orthogonal to each other. Prove that for all n, the chromatic number χ(Gn ) is at
most n − 1. For what values of n is the chromatic number equal to n − 1?
6. A subplane of a projective plane (P, L, I) is a projective plane (P 0 , L0 , I0 ) such that
P 0 ( P, L0 ( L and I0 = I ∩ (P 0 × L0 ).
(a) Prove that if a projective plane has order n and m is the order of a subplane of
the projective plane, then n ≥ m2 .
(b) Prove that the bound above is tight by constructing a projective plane of order
m2 and a subplane in it that has order m, for some m (preferably an infinite
family of m’s).
(c) Show that if n > m2 , then n ≥ m2 + m.
7. For a prime power q, and integers 0 ≤ k ≤ n, the Gaussian binomial coefficient is
defined as
(q n − 1)(q n−1 − 1) · · · (q n−k+1 − 1)
 
n
= .
k q (q k − 1)(q k−1 − 1) · · · (q − 1)

(a) Prove that the number of k-dimensional subspaces of PG(n, q) is equal to n+1 .
 
k+1 q

(b) Give combinatorial proofs of the following identities:

12
It has been conjecture that for every prime p, there is a unique projective plane of order p.

Page 16
 Affine and Projective Spaces

(i)      
n n−1 k n−1
= +q .
k q k−1 q k q

(ii)    
n n
= .
k q n−k q

(iii)      
n n−k n − 1 n−1
=q + .
k q k−1 q k q

8. For integers 1 ≤ k ≤ n, prove the points of PG(n − 1, q) can be partitioned into a

collection of pairwise disjoint k − 1 dimensional subspaces if and only if k divides n.
9. Determine the number of k-dimensional subspaces in AG(n, q), for 0 ≤ k ≤ n.
10. Prove the following claims for PG(n, F ), n ≥ 2:
(a) Let H be a hyperplane and S a k-dimensional subspace, for k < n − 1. Then
either S is contained in H or S intersects H in a (k − 1)-dimensional subspace.
(b) Any subspace that intersects every line in at least one point must be a hyperplane.
(c) Let `1 , `2 be two lines through a point z, and for i ∈ {1, 2} let xi , yi be two distinct
points on `i \ {z}. Then the lines x1 x2 and y1 y2 meet in a point.13

13
This shows that indeed these are abstract projective spaces.

Page 17
2 Conics and Ovals

2.1 Coordinates and Hypersurfaces

Let PG(n, F ) be the n-dimensional projective space with V as the underlying n-dimensional
vector sapce over the field F . After picking a basis of V , we can give coordinates to each
element of V . These coordinates can also be used for the points of PG(n, F ), i.e., the
1-dimensional subspaces of V . A 1-dimensional subspace is determined by a non-zero vec-
tor, upto scalar multiplication by a non-zero element of F . Thus, a point P of PG(n, F )
corresponding to hvi for some v ∈ V \ {0} is given the coordinates (x0 , . . . , xn ) of v, where
we assume that (x0 , . . . , xn ) = λ(x0 , . . . , xn ) for all λ ∈ F ? . In other words, we put an
equivalence relation on the non-zero elements of F n+1 given by (x0 , . . . , xn ) ∼ (x00 , . . . , xn
if ∃λ ∈ F ? such that (x0 , . . . , xn ) = λ(x00 , . . . , x0n ), and then the coordinates of the points
of P G(n, F ) correspond to the equivalence classes of this relation.
If we right-normalise, that is, make the first non-zero coordinate from the right equal to 1
by dividing every coordinate by that element, then we get a unique representative of each
of these equivalence classes, and hence a unique coordinate for every point of PG(n, F ).
For example, the right normalised coordinates of the points of PG(2, F ) are

{(λ, µ, 1) : λ, µ ∈ F } ∪ {(λ, 1, 0) : λ ∈ F } ∪ {(1, 0, 0)}.

We can think of the points with z = 0 as the points lying at infinity, and {(λ, µ, 1) : λ, µ ∈
F } as the points of the affine plane we get by removing the line z = 0. The point (λ, 1, 0)
denotes the point at infinity where the lines x = λy + c meet the line at infinity, whereas
the point (1, 0, 0) denotes the point where all horizontal lines of the affine plane meet.
There are other ways to give unique coordinates to each point as well. For example, we
could have right normalised the points with z 6= 0 and left normalised the points with
z = 0, to get the points at infinity corresponding to the slopes of the affine lines.
Once we have coordinates, we can consider sets of points defined using algebraic equations.

Definition 2.1. Let d be a positive integer. A homogenous polynomial of degree d in

variables x0 , . . . , xn is a non-zero polynomial
P φ ∈ F [x0 , . . . , xn ] whose each term is a
monomial of the form ax0 x1 · · · xn with
e0 e1 en
ei = d and a ∈ F .

For example, x20 x1 +x2 x23 is a homogeneous polynomial of degree 3 in variables x0 , x1 , x2 , x3

whereas x0 x21 + x32 + x3 is not a homogeneous polynomial.
In PG(n, F ), identifying the non-zero solutions to a polynomial equation φ(x) = 0 with
points of PG(n, F ) make sense if φ is a homogeneous polynomial in F [x0 , . . . , xn ] since
for such a polynomial we have φ(λx) = λdeg φ φ(x), and hence the condition φ(u) = 0 is
well defined on the equivalence class of the non-zero scalar multiples of u.

Page 19
Conics and Ovals

Definition 2.2. A hypersurface of degree d in PG(n, F ) is the set of points satisfying a

degree d homogeneous polynomial equation. A hypersurface in PG(2, F ) is known as a
curve.

In particular, every line of PG(2, F ) is a curve as it is given by the points satisfying an

equation of the form ax + by + cz = 0, for some (a, b, c) ∈ F 3 \ {0} (recall that a line is
just a 2-dimensional subspace of F 3 .)
Remark 2.3. A projective algebraic variety is defined as the set of common zeros of a
system of homogeneous polynomial equations. In particular, every k-dimensional subspace
of PG(n, F ) is a variety. What are the homogeneous polynomial equations that define the
set of points of such a subspace?

2.2 Conics
Definition 2.4. Let φ ∈ F [x, y, z] be a homogeneous polynomial of degree 2. Then the
set C = {x ∈ PG(2, F ) : φ(x) = 0} is called a conic. A conic is called irreducible,
or nondegenerate, if the polynomial φ cannot be written as a product of two degree 1
polynomials over F and every extension of F .
Example 2.5. 1. In PG(2, R), the conic C defined by the equation x2 + y 2 + z 2 + 2xy = 0
is degenerate, since the polynomial is equal to (x + y)2 + z 2 , which splits into two
linear factors (x + y + iz)(x + y − iz) over the extension C of R.
2. The conic C defined by y 2 − zx = 0 in PG(2, F ) is non-degenerate and it can be
parametrically written as C = {(λ2 , λ, 1) : λ ∈ F } ∪ (1, 0, 0). Therefore, if F = Fq ,
then C has q + 1 points.
2 2 2 2
3. The parabolas, hyperbolas, and ellipses given by y 2 = 4ax, xa2 − yb2 = 1 and xa2 + yb2 = 1,
in R2 , can be homogenized to give us the following conics in PG(2, R): y 2 = 4axz,
2 2
x2 2
a2
− yb2 = z 2 and xa2 + yb2 = z 2 . How many points at infinity, that is, the points with
z = 0, do each of these conics have?

Our aim in this section is to understand conics in Desarguesian projective planes and then
look at their combinatorial generalizations to arbitrary finite projective planes.
Lemma 2.6. Let C be a conic in PG(2, F ) and ` a line. Then either all points of ` are
contained in C or ` intersects C in at most two points.
Proof. Let C be defined by φ(x, y, z) = ax2 + by 2 + cz 2 + f yz + gxz + hxy = 0. Say
` is not contained in C, and let P = (x1 , y1 , z1 ) ∈ ` \ C. Then φ(x1 , y1 , z1 ) 6= 0. Let
Q = (x0 , y0 , z0 ) be any other point of `. Then an arbitrary point of ` \ {P } is given by
Q + λP = (x1 , y1 , z1 ) + λ(x0 , y0 , z0 ) for some λ ∈ F (where λ = 0 corresponds to the point
Q). This point in ` \ {P } is contained in C if and only if

φ(Q + λP ) = φ(x0 , y0 , z0 )
+ ((2ax0 + gz0 + hy0 )x1 + (2by0 + f z0 + hx0 )y1 + (2cz0 + f y0 + gx0 )z1 ) λ
+ φ(x1 , y1 , z1 )λ2 = 0.
(2.1)
Since this is a quadratic equation in λ with a non-zero leading term, it has at most two
zeros in F .

Page 20
 Conics and Ovals

Remark 2.7. Note that the quadratic equation in the proof above always has zeros in
a degree 2 extension of F ; in fact, it has exactly two zeros counted with multiplicity.
Moreover, when F = Fq , there is a unique quadratic extension, Fq2 , and so every line not
contained in C intersects the conic C 0 obtained by using the same equation as that of C
but over F2q , in two points of PG(2, q 2 ).
If E is a quadratic extension of a field F , obtained by adjoining the square √ root of a
non-square k in F , then every element of E can be written uniquely as √ λ + µ k, with
λ, µ ∈ F . √ The conjugate a of an element a ∈ E, with a = λ + µ k is defined as
1

a = λ − µ k. Note that aa ∈ F and a + a ∈ F for all a ∈ E. For a line ` in PG(2, E)

given by the equation ax + by + cz = 0, we define ` to be the line given by the equation
ax + by + cz = 0, which is equivalent to taking the ` to be the set of conjugates of points
on `. Note that if ` 6= `, then the point ` ∩ ` always lies in the subplane PG(2, F ) of
PG(2, E).
Lemma 2.8. Let C be a conic in PG(2, F ) defined by φ ∈ F [x, y, z]. Then φ is reducible
if and only if C contains all the points of a line in F , or in a quadratic extension E of F .
Moreover, if it’s reducible and one of the linear factors does not have all of its coefficients
in F then the two linear factors define conjugate lines in PG(2, E).
Proof. Left to the reader.

If Q = (x0 , y0 , z0 ) is a point on the conic and P = (x1 , y1 , z1 ) is an arbitrary point outside

the conic then Equation 2.1 is of the form rλ + sλ2 = 0, where s 6= 0. Therefore it has
roots λ = 0 and λ = −r/s. In particular, λ = 0 is a double root, i.e. the line P Q
intersects C in only one point, if and only if r = 0, where

r = (2ax0 + gz0 + hy0 )x1 + (2by0 + f z0 + hx0 )y1 + (2cz0 + f y0 + gx0 )z1 = 0.

It then follows that if the three equations

∂φ
:= 2ax + hy + gz = 0
∂x
∂φ
:= hx + 2by + f z = 0
∂y
∂φ
:= gx + f y + 2cz = 0
∂x
are not simultaneously satisfied by the point (x0 , y0 , z0 ) of the conic, then the line given
by
(2ax0 + gz0 + hy0 )x + (2by0 + f z0 + hx0 )y + (2cz0 + f y0 + gx0 )z = 0
is the unique line through Q = (x0 , y0 , z0 ) that intersect C in only Q. It is called the
tangent line at Q. 2
Definition 2.9. Let C be a curve in PG(2, F ) defined by the equation φ = 0. Then a
point P on C is called a singular point of C if
∂φ ∂φ ∂φ
= = =0
∂x ∂y ∂z
1
Think of how we get C from R.
2
The partial derivatives above are purely formal algebraic operations, where we define the derivative of
xr to be rxr−1 and extend this linearly to define the derivative of any univariate polynomial.

Page 21
Conics and Ovals

when evaluated at P . The curve is called singular if it has a singular point on it, and
non-singular otherwise.

For conics we have the following nice characterisation of irreducibility, which we will use
to count the total number of points on a conic in PG(2, q).

Proposition 2.10. A conic in PG(2, F ) is irreducible, if and only if it is non-singular.

Proof. Let C be the conic defined by the polynomial φ ∈ F [x, y, z]. Say φ is reducible
and φ = (ax + by + cz)(a0 x + b0 y + c0 z), possibly over a quadratic extension E of F . If
(a, b, c) ∈ E 3 \ F 3 , then a0 = λa, b0 = λb and c0 = λc, for some λ ∈ F ? (by Lemma 2.8).
Let P be the point of intersection of the lines ` and m defined by the zero sets of these
two linear factors of E (take P to be any point on the line if ` = m). Then P is in
PG(2, F ) and it can be easily checked that all the partial derivatives vanish at P , making
C singular.

Say φ = ax2 + by 2 + cz 2 + f yz + gxz + hxy and say C is singular. After a change of basis,
we can assume that the singular point has coordinates (1, 0, 0). Since φ(1, 0, 0) = 0 and
all the three partial derivatives vanish at this point, we get a = g = h = 0, and hence
φ = by 2 + cz 2 + f yz. This has linear factors
! !
√ √
p p
f − f 2 − 4bc f + f 2 − 4bc
( b)y + √ z and ( b)y + √ z,
2 b 2 b

over an extension of F where b and f 2 − 4bc are perfect squares. Therefore, C is non-
degenerate.
Example 2.11. Using this proposition, we can easily check that the conic defined by x2 +
y 2 + z 2 = 0 is non-degenerate whenever the characteristic of the field is not equal to 2,
as then the only common solutions that the three equations from the partial derivatives
have is (0, 0, 0). The polynomial xy + yz + zx gives us an irreducible conic over fields of
characteristic 2.

Lemma 2.12. If an irreducible conic C of PG(2, q) contains a point, then it has exactly
q + 1 points in it.
Proof. Since it is irreducible, it is non-singular. Say P is a point of C, and let ` be the
unique line through P which is tangent to C. Each of the q lines through P , other than
the line `, intersect C in exactly one other point, and conversely every point of the conic
other than P defines one of these lines. Therefore, C has exactly q + 1 points in it.
Remark 2.13. Note that in this proof we only used the fact that every irreducible conic
is also non-singular; the other direction is not needed.

It will follow from the Chevalley-Warning theorem, which we will see in the next chapter,
that every conic (whether irreducible or not) in PG(2, q) contains a point, i.e., there are
no “point-less” conics over a finite field.3
3
In PG(2, R) the conic given by x2 + y 2 + z 2 = 0 is point-less. Note that it is also irreducible.

Page 22
 Conics and Ovals

2.3 Ovals and Hyperovals

Definition 2.14. An arc in a projective plane is a set of points in which no three points
are collinear. An arc of size k is called a k-arc.

Clearly, any subset of an irreducible conic is an arc, as otherwise the conic will contain a
line. The conic itself gives us a (q + 1)-arc in PG(2, q). Is this the largest possible arc we
can have in a finite projective plane of order q?

Theorem 2.15. An arc in a projective plane of order n has at most n + 2 points in it.
Moreover, if n is odd then it has at most n + 1 points.
Proof. Let S be an arc in a projective plane π of order n. Let x ∈ S. For each of the
n + 1 lines of π through x, we have at most one point of S \ x contained in the line, and
thee must be all points of S since every point y ∈ S \ {x} defines the line yx through x.
Therefore, |S \ {x}| ≤ n + 1, and hence |S| ≤ n + 2.
Let S be an arc of size n + 2. Then every line must intersect S in 0 or 2 points, as if we
have a line intersecting S in exactly one point, then looking at the n + 1 lines through
this point we will get |S| ≤ n + 1. Now consider the lines through a point y 6∈ S. They
give rise to a partition of S into pairs, showing that n + 2 is even, and hence n is even.
Therefore, for n odd, every arc has size at most n + 1.

Definition 2.16. An arc of size n + 1 in a projective plane of order n is called an oval,

and an arc of size n + 2 is called a hyperoval.

From the proof of Theorem 2.15 it follows that through every point of an oval there is a
unique line that intersects the oval in a single point. This gives us an alternate proof of
the fact that every point of a nondegenerate conic in PG(2, q) is contained in a unique
tangent, if we can show that an nondegenerate conic has q + 1 points in it, without using
the fact that each point is in a unique tangent.
The nondegenerate conics in Desarguesian projective planes show that ovals exist. But
what about hyperovals? Can you find a hyperoval in PG(2, q) for any even prime power
q? It turns out that whenever there exists an oval in a an arbitrary projective plane of
even order, there is also a hyperoval. In fact, something stronger holds true.

Theorem 2.17. Let π be a projective plane of even order n, and O an oval in π. Then
there is a unique hyperoval in π that contains O.
Proof. We are looking for a unique point p 6∈ O such that all lines through p are tangents
to the oval, i.e., p must be the intersection of all n + 1 tangents. For 0 ≤ i ≤ n + 1, let
ei denote the number of points in the plane, outside O, that are incident with exactly i
tangents. We want to show that en+1 = 1.4
Noting that each of the n2 points outside can be incident to at most n + 1 tangents of O,
we have
n+1
X
ei = n2 .
i=0

4
In fact, showing that en+1 > 0 would suffice.

Page 23
Conics and Ovals

Double counting the pairs (x, `) where x is a point of π outside O and ` is a tangent of
O through x, we get
Xn+1
iei = (n + 1)n.
i=1

Counting the triples (x, `, m) where `, m are two distinct tangents of O and x is their
point of intersection, we get
n+1
X
i(i − 1)ei = (n + 1)n.
i=2

Now observe that since n + 1 is odd, each point outside must be incident to at least one
tangent, since otherwise we will get aPpartition of O into pairs using the lines through
this point. Thus we have e0 = 0, and n+1i=1 ei = n .
2

Consider the sum

n+1
X
S= (i − 1)(n + 1 − i)ei ,
i=1

which is always ≥ P0 with equalityPif and only ifPei = 0 for all i ∈ {2, . . . , n}. We can
simplify it to S = (n + 1)iei − (n + 1)ei − i(i − 1)ei . Using the equations above
we get S = (n + 1)2 n − n2 (n + 1) − (n + 1)n = 0.

We can now compute en+1 from the third equation where we get (n +1)n · en+1 = (n +1)n,
and hence en+1 = 1.
Remark 2.18. The counting trick above is known as the variance trick, or the second
moment method. Over the years it has seen many applications, especially when studying
“regular” combinatorial structures.

Definition 2.19. Given a set S of points in a projective plane, a line ` is called a tangent
of S if |` ∩ S| = 1, a external if |` ∩ S| = 0 and a secant if |` ∩ S| > 1.

Proposition 2.20. Let π be a projective plane of order n, O an oval in π and H a

hyperoval in P.

(1) O has n + 1 tangents, n+1 secants and n2 externals.

2

(2) H has n+2 secants and n

externals.



2 2

Proof. In an oval there is a unique tangent through each of the n + 1 points, a unique
secant through every pair of points on the oval, and the rest are external. In a hyperoval
there are no tangent lines.

The fact that there are n + 1 tangents, and n + 1 points in an oval, might suggest some
sort of duality between these objects. Indeed, for n odd we have the following result.

Theorem 2.21. Let O be an oval in a projective plane of odd order, and O∗ the set of
tangents to O. Then O∗ forms an oval in the dual projective plane.
Proof. There are n + 1 tangents of O, one through each point, giving us n + 1 points in
the dual projective plane. All we need to show is that no three tangents are concurrent.

Page 24
 Conics and Ovals

As above, we get the following three equations, where ei denotes the number of points
outside O which are incident to exactly i tangents of O.
n+1
X
ei = n2 .
i=0

n+1
X
iei = (n + 1)n.
i=1
n+1
X
i(i − 1)ei = (n + 1)n.
i=2

We want to show that ei = 0 for all i ≥ 3. Note that since n is odd, no point outside
O can
P be incident to P a unique tangentPof O, giving usPe1 = 0. Now consider the sum
S = n+1
i=3 i(i−2)e i = n+1
i=0 i(i−2)ei =
n+1
i=0 i(i−1)ei −
n+1
i=0 iei = (n+1)n−(n+1)n = 0.
Therefore, ei = 0 for all i ≥ 3.

Corollary 2.22. Let O be an oval in a projective plane π of odd order. Then the points
of π are partitioned into the following three types:
(1) On points: the point through which there is a unique tangent to O, that is, the points
lying on O; there are n + 1 one such points.
(2) Exterior points: the points through which there are exactly 2 tangents to O; there are
n+1
such points


2

(3) Interior points: the points through which there are no tangents of O; there are n2

such points.

We saw that we can construct ovals in finite Desarguesian planes by taking an irreducible
conic. In fact, we used the property of irreducible conics being a set of q + 1 points no
three of which collinear to motivate the definition of ovals. A deep result of Segre shows
that when q is odd, the converse also holds true.

Theorem 2.23 (Segre 1955). Every oval in PG(2, q), for q odd, is a nondegenerate conic.

This statement was conjectured by Jarnefelt and Kustaanheimo in 1949. In a review,

Marshall Hall Jr. said that “The reviewer finds this conjecture implausible.” He was again
a reviewer of the paper of Segre, where he then said “The fact that this conjecture seemed
implausible to the reviewer seems to have been at least a partial incentive to the author
to undertake this work. It would be very gratifying if further expressions of doubt were
as fruitful.”.
For q even it is not the case that every oval/hyperoval comes from a conic. In the exercise
you will prove that for every even prime power q ≥ 8 there exists an oval in PG(2, q)
which is not a conic.
Classifying hyperovals in PG(2, q), for q even, is a major open problem in finite geometry.
This problem is further motivated by various connects that hyperovals have with other
important objects like generalized quadrangles and certain permutation polynomials over

Page 25
Conics and Ovals

finite fields. Complete classification of hyperovals is only known for q < 64, and there are
various infinite families which do not come from extending a conic by a point.
In non-Desarguesian planes, even the existence of ovals and hyperovals is a widely open
problem in general. It was long conjectured that they always exist, until 1996 when
Penttila, Royle and Simpson showed that there are projective planes of order 16 that do
not contain any ovals, using computer aided search.

2.4 Exercises
1. (a) Prove that the Fano plane is not a subplane of PG(2, R).
(b) Give a characterisation of all fields F , in terms of the characteristic of F , for
which the Fano plane is a subplane of PG(2, F ).
2. Let f ∈ Fq [x, y, z] be an irreducible homogeneous polynomial of degree 2, defining the
non-degenerate conic C in PG(2, q). Let q be odd.
(a) Prove that the function β(u, v) = f (u + v) − f (u) − f (v) is a symmetric bilinear
form on the vector space F3q such that for any non-zero vector u, the set u⊥ =
{v ∈ F3q : β(u, v) = 0} is a 2-dimensional subspace of F3q , and hence a line of
PG(2, q).
(b) Prove that the map x 7→ x⊥ gives rise to an isomorphism between PG(2, q) and
its dual PG(2, q)D . To which lines do the points of C get mapped to by this map?
3. In this exercise we give an alternate proof of the fact any oval in a finite projective
plane of even order can be uniquely extended to a hyperoval. Let O be an oval in a
finite projective plane of even order.
(a) Prove that through every point of the plane which lies on a secant of O, i.e. a
line intersecting O in exactly two points, there is a unique line which is tangent
to O.
(b) Deduce that all tangents of O intersect in a common point, and thus show that
O can be uniquely extended to a hyperoval.
4. Let K6 denote the complete graph on the vertex set [6] := {1, 2, 3, 4, 5, 6}. Define a
point line geometry (P, L, I) as follows:
• P = P1 ∪ P2 where P1 = [6] and P2 is the set of all distinct 1-factors (perfect
matchings) of K6 .
• L = L1 ∪ L2 where L1 = [6] and L2 is the set of all distinct 1-factorizations of


2
K6 . 5

• a vertex x ∈ P1 is incident to an edge e ∈ L1 if x ∈ e and it is never incident to

an element of L2 ; a 1-factor f ∈ P2 is incident to an edge e ∈ L1 if e ∈ f and it
is incident to a 1-factorization F ∈ L2 if f ∈ F .

(a) Prove that this point-line geometry is a projective plane of order 4.

5
A 1-factorization is a partition of the edges of K6 into pairwise disjoint 1-factors.

Page 26
 Conics and Ovals

(b) Prove that this projective plane is isomorphic to PG(2, 4). (Hint: use the existence
of a hyperoval in PG(2, 4).)
(c) Prove that every finite projective plane of order 4 is isoomrphic to PG(2, 4), by
showing that every such plane has a hyperoval.

5. (a) For an arbitrary field F , let S be a set of 5 points in PG(2, F ) such that no
three of them are collinear. Prove that there exists a unique conic containing S.
Moreover, show that the conic is non-degenerate.
(b) Find the number of distinct non-degenerate conics in PG(2, q).
(c) For every even prime power q ≥ 8, show that there exists an oval in PG(2, q)
which is not a conic.
6. Every conic in PG(2, Fq ) is one of the following five types:
a) Conics made up of two coincident lines of PG(2, q).
b) Conics whose equations split into two distinct linear equations over Fq .
c) Conics whose equations split over two distinct conjugate linear equations over
Fq 2 .
d) Irreducible conics that have a point in PG(2, q).
e) Irreducible conics that have no point in PG(2, q).
Count the total number of conics in PG(2, q), the number of conics of each type, and
deduce that every irreducible conic in PG(2, q) has a point on it.
7. Let q be an odd prime power and let S be a set of points in PG(3, q) such that no
three of them are collinear.
(a) Prove that |S| ≤ q 2 + 1 with equality if and only if every plane of PG(3, q)
intersects S in 0, 1 or q + 1 points.
(b) Prove that if |S| = q 2 + 1 then through every point x of S, there exists a unique
plane πx such that S ∩ πx = {x} and moreover, for any line ` through x, we have
` ∩ S = {x} if and only if ` ∈ πx .
(c) Construct such a set S with |S| = q 2 + 1, for every odd prime power q.
8. Let F be a finite field of characteristic 2, and let C be a conic given by ax2 + by 2 +
cz 2 + f yz + gzx + hxy = 0.
(a) Prove that C is equal to a repeated line if and only if f = g = h = 0.
(b) Say C is not equal to a repeated line, and let N be the point of PG(2, F ) with
coordinates (f, g, h). Prove that C is singular if and only if N lies on C.
(c) Say C is irreducible. Then prove that every line tangent to C passes through the
point N .
√
9. (a) Prove that the set {(1, t, t) : t ∈ Fq } ∪ {(0, 0, 1), (0, 1, 0)} is a hyperoval in
PG(2, q), if q is an even prime power.6
6
Note that for q even every element of Fq is a square.

Page 27
Conics and Ovals

(b) Does the hyperoval constructed in (a) contain a conic in it?

Page 28
3 Some Combinatorial Questions

In this chapter we study some combinatorial questions arising in finite geometry. We will
see how beyond simple counting arguments, algebraic methods also play an important
role in answering some of these questions when we are working over Desarguesian spaces.

3.1 Blocking Sets

Definition 3.1. A subset B of points in a projective plane is called a blocking set if B

meets all lines of the plane non-trivially, i.e., for every line ` we have ` ∩ B 6= ∅.

Remark 3.2. The concept of vertex covers in hypergraphs is a generalization of blocking

sets.
Remark 3.3. The notion of blocking sets goes back to a 1955 paper of Richardson “On
Finite Projective Games”, where such a set was called a blocking coalition, if we have the
extra condition that B contains no-lines.

If B is a blocking set, then clearly every set containing B is also a blocking set. Moreover,
a line in a projective plain of order n always forms a blocking set of size n + 1, since any
two lines intersect each other. It turns out that these are the smallest possible blocking
sets.

Theorem 3.4. Every blocking set in a projective plane of order n has size at least n + 1,
with equality if and only if the blocking set is a line.
Proof. The bound itself easily follows from the fact that each point blocks at most n + 1
lines, and hence if we have ≤ n points, then they blocks at most n(n + 1) = n2 + n lines,
which is less than the total number of lines. To characterise the sharp examples, we prove
a stronger claim that for any 1 ≤ k ≤ n + 1, the number of lines meeting a set of k points
is at most kn + 1, with equality if and only if the k points are contained in a line.

This is clearly true for k = 1 since a point meets exactly n + 1 lines. Say the statement
is true for some k ≤ n. We will prove it for k + 1.

Let S be a set of k + 1 points and x ∈ S. Then S \ {x} meets at most kn + 1 lines with
equality if and only if S \ {x} is contained in a line `. There are n + 1 lines through x, and
at least one of them also meets another point of S \ {x}, and hence it has already been
counted in the lines meeting S \ {x}. Therefore, S meets at most kn + 1 + n = (k + 1)n + 1
lines. Equality can only occur if there is exactly one line ` through x which meets a point
of S \ {x}, and all the points of S \ {x} lie on `. The line ` also contains x, and hence S
is a subset of `. Conversely, k + 1 collinear points meet precisely (k + 1)n + 1 lines.

Page 29
Some Combinatorial Questions

Inspired by this result, one might ask what is the corresponding result for a finite affine
plane. In any affine plane of order n, a pair of intersecting lines forms a blocking set,
since every other line is parallel to at most one of these lines, and hence intersects the
other. This gives us a blocking set of size 2n − 1. But is this the best that we can do? It
turns out that if the affine plane is Desarguesian, then indeed this size is optimal (though
there are many other configurations giving us the same bound), but proving this is highly
non-trivial! Moreover, it is known that the following result is not true if the planes are
non-Desarguesian.

Theorem 3.5 (Jamison/Brouwer-Schrijver). Every blocking set in the affine plane AG(2, q)
has size at least 2q − 1.
Proof. Let B be a blocking set in AG(2, q). Let B 0 be the set of points corresponding
to B in the projective plane PG(2, q) we obtain by adding `∞ . Then B 0 is a set of points
in PG(2, q) that meets every line except one, `∞ . In the dual projective plane, which is
isomorphic to PG(2, q), we get a set L0 of lines in PG(2, q) that cover all points except one.
Remove a line ` ∈ L0 to get a set L of lines in AG(2, q) that cover all points except one. We
will show that |L| ≥ 2(q − 1), which would imply that |B| = |B 0 | = |L0 | = |L| + 1 ≥ 2q − 1.
Every line of L is given by the zero set of a degree one polynomial in Fq [x, y]. Multiply
these polynomials together to get a polynomial f of degree |L| with the property that
f vanishes on all points of F2q , except one. From the Lemma below we see that |L| =
deg f ≥ 2(q − 1), and hence |B| = |L| + 1 ≥ 2q − 1.

Lemma 3.6. Let f ∈ Fq [x1 , . . . , xn ]. If there exists a point u ∈ Fnq such that f (u) 6= 0
and f (a) = 0 for all a =
6 u, then deg f ≥ n(q − 1).
Proof. Say deg f < n(q − 1). Then in everyPmonomial m of f , there exists an exponent
which is less than q − 1. This implies that a∈Fnq m(a) = 0, since we can split the sum
according to the n variables, and the variable whose power is less than q − 1 and greater
than 0 gives us the 0 factor by Lemma 0.9. If the exponent of a variable is 0 then this
also gives us
Pa 0 sum, as then we have q times the sum of the evaluation of the monomials
Therefore, a∈Fnq f (a) = 0. But this is impossible since the sum is equal to f (u) 6= 0.

As mentioned before, the lower bound of 2q − 1 on size of a blocking set does not hold
true for non-Desarguesian affine planes. In fact, recently it has been proved by De Beuele,
Héger, Szönyi and Van de Voorde, that for every prime power q ≥ 3, there exist non-
Desarguesian affine planes of order q 2 that contain a blocking set of size b4q 2 /3 + 5q/3c.
Another interesting consequence of Lemma 3.6 is the following classical result in number
theory, which in particular implies that every conic in PG(2, q) has at least one point in
it.

Theorem 3.7 (Chevalley-Warning Theorem). Let f ∈ Fq [x1 , . . . , xn ] be a polynomial of

degree d < n. If the set Z = {a ∈ Fnq : f (a) = 0} is non-empty, then |Z| > 1.
Proof. Let g = 1 − f q−1 . Then g is a polynomial of degree d(q − 1) < n(q − 1), such
that g(a) 6= 0 for all a ∈ Z and g(a) = 0 for all a 6∈ Z. If |Z| = 1, then g vanishes on
all points except one, and hence by Lemma 3.6, it must have degree at least n(q − 1), a
contradiction. Therefore, either Z = ∅ or |Z| > 1.

Corollary 3.8. Every conic in PG(2, Fq ) contains a point.

Page 30
 Some Combinatorial Questions

Proof. Let φ ∈ Fq [x, y, z] be the homogeneous polynomial of degree 2 defining the conic.
Since φ is homogeneous, (0, 0, 0) ∈ Z := {a ∈ F3q : φ(a) = 0}. Therefore, |Z| > 1, and
hence there exists a non-zero vector at which φ vanishes, giving us a point in C.

While Theorem 3.4 might suggest that there is nothing that interesting about blocking
sets in finite projective planes, since we can completely characterise the minimal ones,
and in fact we can easily construct blocking sets of any larger size by just adding extra
points. Things start getting more interesting if we disallow our blocking sets to contain
lines.

Definition 3.9. A blocking set in a projective plane is called non-trivial if it does not
contain any line.

What are some examples of non-trivial blocking sets that you can think of? What size
does your set have, if the order of the projective plane is n? One of the simplest examples
is the set of size 3n − 3 consisting of three sides of a triangle except the vertices, the
so-called vertex-less triangle, which gives a non-trivial blocking set for all n > 2. Is this
the smallest possible non-trivial blocking set? We will construct better ones for some
families of finite projective planes.

Consider PG(2, q 2 ). Restricting the coordinates to the subfield Fq , we get a copy of

PG(2, q) inside PG(2, q 2 ). This is known as a Baer subplane. In fact, in general if
n = m2 , then a Baer subplane in a projective plane of order n is defined as a subplane of
order m.

Theorem 3.10 (Bruen 1970). Baer subplanes are non-trivial blocking sets.
Proof. Let B be a collection of m points in a projective plane π of order n = m2 such that
the unique line through any 2 points of B contains exactly m + 1 points of B. Moreover,
any two lines which contain m+1 points of B meet inside B. For the sake of contradiction,
let ` be a line of π that does not contain any point of B. Let x ∈ `. There can be at
most one line through x that contains m + 1 points of B, since B forms a Baer subplane,
and every other line through x contains at most one point of B. Since the lines through
x must partition the points of B, we get |B| ≤ m + 1 + (n − 1) = m2 + m, which is a
contradiction since B is a Baer subplane of order m and hence has m2 + m + 1 points.

Note that from the

√ proof it follows that every line intersects the set of points of a Baer
subplane in 1 or n√+ 1 points, where n is the order of the plane. In fact, it can be shown
that any set of n + n + 1 points with these intersection properties with respect to lines
must correspond to a Baer subplane.

Theorem 3.11 (Bruen√1970). Every non-trivial blocking set in a projective plane of order
n has size at least n + n + 1, with equality if and only if it is the set of points of a Baer
subplane.
Proof. Let B be a non-trivial √ blocking set of the projective plane. We will assume that
B is of size n + m, with m ≤ n + 1 and show that it must be a Baer subplane. First we
claim that every line intersects a such a set B in at most m points. For this, let ` be a
line and let x be a point in ` \ B. Every line through x contains at least one point of B,
giving us n + |` ∩ B| ≤ |B| = n + m, and hence |` ∩ B| ≤ m.

Page 31
Some Combinatorial Questions

Let the lines of the plane be `1 , . . . , `n2 +n+1 , and define ki = |`i ∩ B|. By double counting
the pairs (x, `) where x is a point of B and ` a line through x, we get
X
ki = |B|(n + 1).

By double counting the triplets (x, y, `) where x 6= y ∈ B and ` is the unique line joining
x and y, we get X
ki (ki − 1) = |B|(|B| − 1).
√ √
Consider the sum S = (ki − 1)( n + 1 − ki ). Since 1 ≤ ki ≤ n√+ 1 for all i, we
P
have S ≥ 0, with equality if and only if every line intersects B in 1 or n + 1 points. By
expanding the sum and using the previous identities we get
√ √ √ √
S = −|B|(|B|−1)+( n+1)(n+1)|B|−( n+1)(n2 +n+1) = (|B|−n− n−1)(n n+1−|B|).
√
Since√S ≥ 0, we must have |B|√ ≥ n + n + 1, and since we assumed |B| = n + m ≤
n + n +√ 1, we get |B| = n + n + 1. Therefore, S = 0 and hence every line intersects
B in 1 or n + 1 points.

Remark 3.12. Instead of working with ki ’s, we can define ei to be the number of lines
which intersect the blocking set in exactly i points, which will then give us the following
equations: √
n+1
X
ei = n2 + n + 1,
i=1
√
n+1
X
iei = |B|(n + 1),
i=1
√
n+1
X
i(i − 1)ei = |B|(|B| − 1).
i=1

This choice of variables is similar to our earlier proofs on ovals, but √

it basically gives the
same proof as above when we try to show that ei = 0 for all i 6∈ {1, n + 1}.

What happens when the order of the projective plane is not a square? What’s the smallest
size of a non-trivial blocking set? This is a much more difficult question. It is known that
in PG(2, p) the smallest non-trivial blocking set has size 3(p + 1)/2. Proving this requires
another surprising application of the polynomial method, due to Blokhuis.

Theorem 3.13 (Blokhuis 1994). Let S be a non-trivial blocking set in PG(2, p). Then

|S| ≥ 3(p + 1)/2.

The construction however can be described using directions determined by certain sets
of points in AG(2, p), as you will see in the exercises. We now look at another direction
problem, which was solved quite recently using polynomial method.

Page 32
 Some Combinatorial Questions

3.2 Kakeya and Nikodym Sets

The classical Kakeya’s needle problem asks for the “smallest” subset S of R2 such that S
contains a unit line segment (needle) in every direction. Clearly a semi-circle of radius
1 is such an example with area π/2. A circle of diameter 1 is a smaller example of such
a set, as it has area √
π/4. An even smaller example is an equilateral triangle of height
1, which has area 1/ 31 . The reader is invited to come up with further examples that
have smaller area. Perhaps surprisingly, Besicovitch proved (in 1919) there is no δ > 0
such that every Kakeya set in R2 has area at least δ. Therefore if area was a measure of
how small a Kakeya set be in R2 , then the answer is 0. Over the years, mathematicians
used other “measures” (like Minkowski dimension and Hausdorff dimension), and made
conjectures regarding lower bounds. For example, it was conjectured that any Kakeya
set in Rn has Hausdorff/Minkowski dimension at least n (any subset of Rn has Hausdorff
dimension at most n, and hence Kakeya sets are conjectured to have the highest possible
Hausdorff dimension). This is known for n = 1, 2 but it is still wide open for all n ≥ 3.
To solve these conjectures, many connections between Kakeya conjecture and some prob-
lems from areas like Additive Combinatorics, and Fourier Analysis were found.2 In 1999,
Wolff suggested that perhaps looking at the same question over the finite field setting
might provide us some tools to make progress in the original setting. He made the fol-
lowing conjecture regarding Kakeya sets in AG(n, q):
Conjecture 3.14. Let S be a set of points in the affine space AG(n, q) with the property
that for every direction, there exists a line in that direction that is completely contained
in S. Then
|S| ≥ cn q n ,
for some constant cn that only depends on n.
Remark 3.15. In AG(n, q), the line joining points x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) is
given by {x + λ(y − x) : λ ∈ Fq }, where y − x = (y1 − x1 , . . . , yn − xn ).

Note that the “direction” of a line ` in AG(n, q) corresponds to a point p in the hyperplane
at infinity, H∞ , in the projective completion PG(n, q) of the affine space, such that in
PG(n, q) the line ` contains the point p. Therefore, a Kakeya set can alternately be
defined as a set S of points in PG(n, q) such that there exists a hyperplane H disjoint
from S and for every point x ∈ H, there exists a line `x such that `x \ {x} ⊆ S. This
definition leads to the concept of Nikodym sets.
Definition 3.16. A Nikodym set in PG(n, q) (or AG(n, q)), is a set S of points such that
for all x 6∈ S, there exists a line `x for which (`x \ {x}) ⊆ S.

A similar conjecture for the size of Nikodym sets was also proposed. So, let us explore how
small Kakeya and Nikodym can be in the finite field spaces. We will start with Kakeya
sets in the plane.
Example 3.17. Let H be a hyperoval in PG(2, q) for q even. In the dual plane, which is
also isomorphic to PG(2, q), H corresponds to a set L of q +2 lines such that no-three lines
are concurrent. Let ` be a line of L and define K to be the set of points

contain onq+1 the
lines in L, except those lying on `. Then K is a Kakeya set of size 2 − (q + 1) = 2 .
q+2

1
This is in fact the minimum possible area if we only allow convex sets.
2
See for a survey.

Page 33
Some Combinatorial Questions

Theorem 3.18. A Kakeya set in AG(2, q) has size at least q+1

.


2
Proof. Let K be the Kakeya set. We have q + 1 distinct lines, corresponding to each of
the directions, completely contained in our set. The first line gives us q points of K, the
second one gives rise to at least q − 1 further points, the third line q − 2 points, and so on,
because any two of these lines intersect in at most (in fact, exactly) one point. Therefore,
|K| ≥ q(q + 1)/2. Equality occurs if and only if K is the union of these q + 1 lines, and
no three lines are concurrent,i.e., the lines along with the line at infinity form a hyperoval
since all these lines are in different direction.
Example 3.19. Let O be an oval in PG(2, q), with q odd. Take a point x on O and let `
be the tangent line through x to O. Through every point y on `, except for x, there is a
unique line `y 6= ` that is a tangent to O. Pick any line `x through x which is not equal
to `. Let !
[
K= `y \ `.
y∈`

Then in the affine plane that one obtains by treating ` as the line at infinity, K is a
Kakeya set in AG(2, q) of size q(q + 1)/2 + (q − 1)/2.

Theorem 3.20 (Blokhuis-Mazzocca 2008). A Kakeya set in AG(2, q), for q odd, has size
at least q(q + 1)/2 + (q − 1)/2, with equality if and only if it is equivalent to the example
given above.

The proof is beyond the scope of our course, but it’s interesting to note that its main
ingredients are Theorems 2.23 and 3.5. Blokhuis and Mazzocca also managed to give a
linear algebraic proof of the fact that in AG(n, q), the size of a Kakeya set is at least
q n−1 /(n − 1)!, which was quite close to the conjecture and substantially improved the
previously known best bound of cn q (n+2)/2 . But soon after they proved their results, Dvir
settled the conjecture completely, using a different (and much simpler) linear algebraic
argument. We will now see his proof, first for Nikodym sets (since it’s simpler) and then
for Kakeya sets. The basic idea behind the proof is as follows:

(1) Say the set S has smaller cardinality then what we want to prove.

(2) Show that for such a small S there must be a non-zero polynomial f that vanishes
on all points of S.

(3) Show that such a polynomial f then vanishes on too many points, and hence must
be the 0 polynomial, giving us a contradiction.

We start with the algebraic lemmas that formalise steps (2) and (3) above.

Lemma 3.21. Let F
be a field, d, n positive integers, and E ⊆ F n a set of cardinality
strictly less than d+n
n
. Then there exists a non-zero polynomial f ∈ F [x1 , . . . , xn ] of
degree at most d such that f (x) = 0 for all x ∈ E.
Proof. The vector space F E of functions from E to F , is of dimension |E|. Consider the
vector space V of all polynomials in
F [x1 , . . . , xn ] with degree at most d. Then by a simple
counting argument, dim V = d+n n
. Consider the linear map ev from V to F E , defined
by taking ev(f ) to be the function x 7→ f (x). If |E| < dim V , then this map cannot be
injective, and in particular the kernel of ev contains a non-zero polynomial f .

Page 34
 Some Combinatorial Questions

Lemma 3.22 (Ore, DeMillo-Lipton-Schwartz-Zippel). Let f ∈ Fq [x1 , . . . , xn ] be a non-

zero polynomial of degree d < q. Then f has at most dq n−1 zeros in Fnq .
Proof. We prove this by induction on n. For n = 1, this follows from the fact that any
univariate polynomial of degree d over an arbitrary field has at most d zeros. We now
proceed by induction. Let n > 1, and assume that the statement is true for n − 1. Write
f as
dn
X
f (x1 , . . . , xn ) = gi (x1 , . . . , xn−1 )xin ,
i=1

where dn = degn f is the highest exponent of xn appearing in f . Note that deg gn ≤ d−dn ,
since d = deg f , and gn 6= 0. Let (a1 , . . . , an−1 ) ∈ Fqn−1 such that gn (a1 , . . . , an−1 ) 6= 0,
then f (a1 , . . . , an−1 , xn ) is a non-zero univariate polynomial of degree dn , and hence
has at most dn zeros. We can partition the set Z of zeros of f in Fnq into the set
Z1 = {(a1 , . . . , an−1 , an ) : gn (a1 , . . . , an−1 ) = 0, f (a1 , . . . , an ) = 0} and the set Z2 =
{(a1 , . . . , an−1 , an ) : gn (a1 , . . . , an−1 ) 6= 0, f (a1 , . . . , an ) = 0}. By the discussion above,
|Z2 | ≤ dn q n−1 and by the induction hypothesis, |Z1 | ≤ q(d − dn )q n−2 since deg gn = d − dn .
Therefore, |Z| ≤ dn q n−1 + dq n−1 − dn q n−1 = dq n−1 . Let S ⊆ Fq be the set of values t for
which f (x1 , . . . , xn−1 , t) is the 0 polynomial in Fq [x1 , . . . , xn−1 ].

We can now easily prove the conjectured bound for Nikodym sets.
Theorem 3.23. A Nikodym set in AG(n, q) has size at least q−2+n .


n
Proof. Let N be a Nikodym set. Say |N | < q+n−2
. Then by Lemma 3.21 there exist


n
a non-zero polynomial f of degree at most q − 2 which vanishes on N . Let x ∈ Fnq \ N .
Since N is a Nikodym set, there exists a line ` through x such that ` \ {x} ⊆ N . The
points on ` are given by {x + λv : λ ∈ Fq } for some non-zero vector v (which is the
direction of the line). Evaluating f at an arbitrary point of `, we get f (x + λv) which is
univariate polynomial in λ of degree at most q − 2. Since it vanishes on the q − 1 points
{x + λv : λ 6= 0}, it must be identically 0. In particular, this means that f (x) = 0.
Since x was an arbitrary point outside N , and f vanishes on every point of N , we have f
vanishing entirely on Fnq , which is a contradiction to Lemma 3.22.

Note that q−2+n = (q + n − 2) · (q + n − 3) · · · (q − 1)/n! ≥ q n /(2n!) for all prime powers

n
q and n ≥ 2. For Kakeya sets, the proof above needs to be modified a bit.
Theorem 3.24. A Kakeya set in AG(n, q) has size at least
 
q−1+n
≥ q n /n!.
n
Proof. As before, say there is a Kakeya set K with a smaller size. Let f ∈ Fq [x1 , . . . , xn ]
be a non-zero polynomial of degree d ≤ q−1 which vanishes on K. Let f h ∈ Fq [x1 , . . . , xn , xn+1 ]
be the homogenized form of f obtained by making each monomial in f of degree exactly
d by multiplying it with a suitable power of xn+1 . Note that f h (x1 , . . . , xn , 0) is the ho-
mogeneous part of the polynomial f , that is, the degree d terms in f , and hence this
polynomial is non-zero. Also note that f h (x1 , . . . , xn , 1) = f (x1 , . . . , xn ). Embed the set
K in PG(n, q) formed by adding the hyperplane H∞ given by xn+1 = 0. The fact that K
is a Kakeya set is equivalent to saying that for all points in H∞ , there is a line through this
point whose affine points are all contained in K. Let (a1 , . . . , an , 0) be such a point, with
(a1 , . . . , an ) 6= (0, . . . , 0), and ` the line through it. Let (b1 , . . . , bn , 1) be an affine point

Page 35
Some Combinatorial Questions

on `. Then the affine points of ` are given by {(b1 + λa1 , . . . , bn + λan , 1) : λ ∈ Fq }. As

before, f h restricted to ` is a univariate polynomial in λ of degree at most d ≤ q − 1 which
has at least q zeros since ` ⊆ K, implying that it is the zero polynomial. The leading
coefficient of f h (b1 + λa1 , . . . , bn + λan , 1), is equal to f h (a1 , . . . , an , 0), which must then be
equal to 0. Therefore, g(x1 , . . . , xn ) = f h (x1 , . . . , xn , 0) vanishes on all non-zero vectors
of Fnq . Since g is homogeneous, it in fact vanishes on all elements of Fnq . By Lemma 3.22
g should be identically zero since deg g = deg f h = d ≤ q − 1. This is the contradiction
since g = f h (x1 , . . . , xn , 0) is non-zero.

Therefore, Kakeya (and Nikodym sets) are of cardinality at least cn q n , where cn =

Θ(1/n!). Is this bound tight, in terms of the expression for cn ? Dvir, Kopparty, Saraf and
Sudan later showed that we can improve cn to 1/2n , and they also showed that for Kakeya
sets this bound is tight, up-to a factor of 2 and some lower order terms.3 Interestingly,
for Nikodym sets even the bound of q n /2n is far from being tight. We explore this in the
next section.

3.2.1 Nikodym sets and large minimal blocking sets

We know that every Nikodym set N in AG(n, q) has size at least q n /(2n!). In this
section we will see how this bound is far from being sharp for n = 2. Surprisingly, the
improvement in the lower bound will follow from a combinatorial argument, instead of
polynomial method.
If N is a Nikodym set in AG(n, q), then the set N ∪ H∞ is clearly a Nikodym set in
PG(n, q) which has size |N | + (q n−1 + · · · + q + 1) = |N | + Θ(q n−1 ). Conversely, if N is
a Nikodym set in PG(n, q), then N \ H∞ is a Nikodym set in AG(n, q) of size at least
|N | − Θ(q n−1 ). Therefore, any lower bound of the form cn q n , in projective/affine setting,
implies a lower bound of the form (cn + o(1))q n , in the other setting. So it doesn’t make
much difference in studying Nikodym sets over projective spaces instead of affine spaces
(at least as far as the asymptotic are concerned).
Let’s fix our n to be 2 and study Nikodym sets in PG(2, q). Recall that a set N in PG(2, q)
is a Nikodym set iff for every x 6∈ N , there exists a line `x such that `x \ {x} ⊆ N .
Therefore, the complement N c has the property that through every point x of N c there
exists a line that intersects N c in precisely one point. Such sets have been studied in finite
geometry at least since 1999, when they were introduced by Bruen and Drudge.4 This is
about the same time when Wolff introduced the finite field Kakeya sets.

Definition 3.25. A tangency set in a finite projective plane is a set T such that for all
x ∈ T , there exists a line ` such that ` ∩ T = {x}.

For example, a line is a tangency set of size q + 1, and so is a conic, in PG(2, q). A Baer
√
subplane in PG(2, q 2 ) is also a tangency set, that has size q + q + 1. The motivation for
the study of tangency sets in fact came from the study of minimal blocking sets.
3
The used a more involved polynomial interpolation, that took into account the multiplicities of the
zeros.
4
An earlier concept of strong representative systems in projective planes due to Illés, Szönyi, Wettle
from 1991 is equivalent to tangency sets.

Page 36
 Some Combinatorial Questions

Definition 3.26. A minimal blocking set in a finite projective plane is a blocking set B
such that no proper subset of B is also a blocking set.

Note that through every point x of a minimal blocking set B, there must be a line ` such
that ` ∩ B = {x}, as otherwise B \ {x} will also be a blocking set. In particular, every
minimal blocking set is a tangency set, and hence the complement of a Nikodym set. It is
then possible to answer the question about how small a Nikodym can be via the max-min
question: How large can a minimal blocking set be?

Theorem 3.27 (Bruen-Thas 77). Let B be a tangency set in a projective plane of order
√ √
q. Then |B| ≤ q q + 1, with equality if and only if every line intersects B in 1 or q + 1
points.
√
Proof. Say |B| = tq + 1. We will prove that t ≤ q with equality if and only if every line
√
intersects B in either 1 point or q+1 points. Let `i be the i-th line, for i = 1, . . . , q 2 +q+1,
and ki = |`i ∩ B|. By the same double counting as before, we have
X
ki = |B|(q + 1),
X
ki (ki − 1) = |B|(|B| − 1).
Since through each point of B there is a line that intersects B in exactly 1 point, we have
at least |B| lines that intersect B in exactly one point. Say these are `1 , . . . , `|B| , and
hence ki = 1 for 1 ≤ i ≤ |B|. Let
q 2 +q+1
X
S= (ki − t − 1)2 .
i=1

Then we get S ≥ |B| i=1 t = |B|t , with equality if and only if ki = t + 1 for all i > |B|,
2 2
P
that is, there is exactly one tangent line through each point of S and every other line of
the plane intersects S in exactly t + 1 points. From the equations above, we get
X X
S= (ki2 − ki ) − (2t + 1)ki + (t − 1)2 (q 2 + q + 1).

Putting everything together, we get the inequality

(t + 1)q(q − t2 ) ≥ 0,

which proves the result.

Note that in case of equality, we also get that through each point there is a unique tangent
line. Can there be equality in the bound? Indeed, is sharp, as shown by the so-called
Hermitian curve xq+1 + y q+1 + z q+1 = 0 in PG(2, q 2 ). Since x 7→ xq is an automorphism of
Fq2 , and in fact the only non-trivial automorphism, we can think of xq as the conjugate of
x, just like a − ib is the conjugate of a + ib in C, which is a quadratic extension of R. The
product xxq is always an element of the subfield Fq , and the analogue of the Hermitian
curve over PG(2, C)
xx + yy + zz = 0,

Page 37
Some Combinatorial Questions

which in fact has no solutions, just like the conic defined by x2 + y 2 + z 2 = 0 had no
points in PG(2, R). We will later see how Hermitian curves and conics, are both related
to polarities of a projective plane.
√
Theorem 3.27 implies that Nikodym set in PG(2, q) has size at most q 2 + q + 1 − q q − q =
(1−o(1))q 2 , where o(1) is a function of q that approaches 0 as q approaches ∞. This clearly
improves the polynomial method bound of q 2 /4. In fact, for every n ≥ 2, it has been
conjectured that a Nikodym set in PG(n, q) (or AG(n, q)) has size at least (1 − o(1))q n .
The conjecture is true for n = 2 but widely open for n > 2. The best known general lower
bound of (1 − o(1))q n /2n comes from the polynomial method with multiplicities.
√
We have shown that a minimal blocking set in PG(2, q) has size at least q + q + 1 and at
√
most q q + 1. Determining the possible sizes of a minimal blocking sets, is a challenging
problem in finite geometry. Most of the progress here is based on algebraic methods, but
sometimes even probabilistic methods gives us existence proofs of certain large minimal
blocking sets.

3.3 Exercises

1. Prove that every blocking set of the Fano plane is trivial. For every projective plane
of order n ≥ 3, construct a non-trivial blocking set of size 2n.
√
2. Let n be a square. Prove that if B is a collection of n + n + 1 points √
in a projective
plane of order n such that every line intersects B in either 1 point or n + 1 points,
then B is the point set of a Baer subplane.

3. Let π be a projective plane of order n and L be a subcollection of lines of π. Then

a set B of points of π is called a blocking set relative to L if it for every line ` in L,
we have ` ∩ B 6= ∅. Let O be an oval of π and assume that n is odd. For each of
the following subcollections L of lines, find the smallest possible size of a blocking set
relative to L.

a) L is the set of all external lines to O.

b) L is the set of all secants to O.

c) L is the union of all tangents and externals of O.

d) L is the union of all tangents and secants of O.

In each case, characterise the extremal examples, assuming that you have a Desar-
guesian projective plane.

4. For a subset S of points in AG(2, q) we say that S determines a direction m if there

exist two points in S such that the line joining them has slope m, where m ∈ Fq ∪{∞}
(vertical lines given by the equation x = c are assumed to have slope ∞).

(a) Prove that if |S| > q, then S determines all directions.

Page 38
 Some Combinatorial Questions

(b) Let f be a function from Fq to Fq . Consider the set S = {(x, f (x)) : x ∈ Fq } of

points in AG(2, q) and let D be the set of directions determined by S. Prove that
if |D| > 1, then the set Bf = S ∪ D is a non-trivial blocking set in the projective
completion PG(2, q) of AG(2, q).
(c) Determine the size of the blocking set Bf when f (x) = x(q+1)/2 , assuming q to be
odd.
(d) For every even q, construct a minimal blocking set of size (3q + 2)/2.

5. Let S be a set of points in AG(2, q) such that S blocks every line in each direction
except at most k directions where at most m lines remain unblocked. Prove that
|S| ≥ 2q − k − m.
6. A blocking set in a projective space is a set of points such that every hyperplane
meets the set non-trivially. For all n ≥ 2, and all prime powers q, find the minimum
possible size of a blocking set in PG(n, q), and classify all minimal examples.
7. Let s(q) be the smallest possible size of a set S of points in AG(3, q) such that every
line intersects S non-trivially.
(a) Determine s(2) and s(3).
(b) Prove that
2q 2 − q ≤ s(q) ≤ 3q 2 − 3q + 1
for all q.
(c) Prove that s(q) ≥ 2q 2 − 1 for all q, and s(q) ≤ 3q 2 − 3q for all q ≥ 3.
(d) For q square, prove that s(q) ≤ 2q 2 + o(q 2 ), by using the fact that for such a q
there are two disjoint Baer subplanes in PG(2, q).
(e) (Open Problem) Determine s(q) for all q, at least asymptotically. In particular,
show that lim sup s(q)/q 2 exists and determine its value.
8. Let S be a set of points in PG(n, q) with the property that through each point x of S
there exists a hyperplane Hx such that Hx ∩ S = {x}, a.k.a., a tangent hyperplane.
(a) Prove that |S| ≤ q (n+2)/2+1 , and equality in the bound implies that through each
point there is a unique hyperplane.
(b) Prove that |S| < q (n+2)/2 + 1 for all n ≥ 4.

9. Let A be an affine plane of order q 2 , and let π be the projective plane of order q 2
obtained from A by adding a line `∞ . Let D be a set of q + 1 points on `∞ with the
following property: for any two distinct points x, y of A, if the line xy meets `∞ in
a point of D, then there exists a Baer subplane B of π containing x and y such that
B ∩ `∞ = D. Define the following point-line geometry AD = (P, L, I):
(a) P is equal to the the point set of A.
(b) L = L1 ∪ L2 , where L1 is the set of lines of A that meet `∞ in a point outside D,
and L2 is the set of Baer subplanes B of π that satisfy B ∩ `∞ = D.
(c) I is the natural incidence of containment.

Page 39
 Some Combinatorial Questions

Prove that AD is an affine plane of order q 2 . 5

10. (a) Let S be be a subset of points in a projective plane of order q, and let k1 , . . . , kq2 +q+1
be the intersection sizes of the q 2 + q + 1 lines with the set S. Prove that
q 2 +q+1  2
X (q + 1)|S|
ki − 2 ≤ q|S|.
i=1
q +q+1

(b) Let S be a set of points and T a set of lines in a projective plane of order q. Define
i(S, T ) to be the number of incidences between the points of S and the lines of
T , that is, i(S, T ) = |I ∩ (S × T )|, where I is the point-line incidence relation of
the projective plane. Prove that
 
i(S, T ) − q + 1 |S||T | ≤ q|S||T |.
  p
 q2 + q + 1 

11. Let H = {(x, y, z) ∈ PG(2, q 2 ) : xq+1 + y q+1 + z q+1 = 0}.

a) Determine |H| by considering the following three types of points in the plane:
{(x, 1, 0) : x ∈ Fq2 }, {(0, y, 1) : y ∈ Fq2 } and {(x, y, 1) : x, y ∈ Fq2 , x 6= 0}.
(Hint: The function x 7→ xq+1 is the norm function from Fq2 to Fq .
b) Let P = (x0 , y0 , z0 ) ∈ H. Assuming that no line is completely contained in H,
show that the line `P defined by the equation xq0 x + y0q y + z0q z = 0, satisfies
`P ∩ H = {P }. Deduce that every line meets H in exactly 1 or q + 1 points.
c) (Bonus) Show that H does not contain any line completely.
12. (a) Let S be be a subset of points in a projective plane of order q, and let k1 , . . . , kq2 +q+1
be the intersection sizes of the q 2 + q + 1 lines with the set S. Prove that
q 2 +q+1  2
X (q + 1)|S|
ki − 2 ≤ q|S|.
i=1
q +q+1

(b) Let S be a set of points and T a set of lines in a projective plane of order q. Define
i(S, T ) to be the number of incidences between the points of S and the lines of
T , that is, i(S, T ) = |I ∩ (S × T )|, where I is the point-line incidence relation of
the projective plane. Prove that
 

i(S, T ) − q + 1  p
|S||T | ≤ q|S||T |.
 q2 + q + 1 

5
By choosing such sets D in PG(2, q 2 ) one can construct non-Desarguesian planes of order q 2 . You can
try to find such a set and show that the plane you construct is non-Desarguesian, for all q ≥ 3.

Page 40
4 Strongly Regular Graphs

4.1 Introdction
In finite geometry, the kind of graphs that we usually encounter are often highly symmet-
ric. In particular, they are regular graphs, i.e., the degree of each vertex is the same. In
fact, often the graphs have an even higher form of regularity, captured by the following
notion due to R. C. Bose.

Definition 4.1. A non-complete and non-empty graph G is called a strongly regular graph
with parameters n, k, λ, µ, or more concisely an srg(n, k, λ, µ), if it is a k-regular graph on
n vertices such that
• every pair of adjacent vertices have exactly λ common neighbours, and
• every pair of non-adjacent vertices have exactly µ common neighbours.

We will develop some theory of these graphs, which will be applied to finite geometry in
later chapters. Using this theory, we will also see a proof of the famous friendship theorem
of Erdős, Rényi and Sós, which says that in any finite society where every two individuals
have a unique common friend, there must be a person who is everybody’s friend. This
result also has a surprising connection to finite projective planes.
Some “trivial” examples of strongly regular graphs are the disjoint union of m Kr ’s, which
is an srg(mr, r − 1, r − 2, 0), and its complement, the m-partite complete graph with each
part of size r, which is an srg(mr, (m − 1)r, (m − 2)r, (m − 1)r). These are trivial because
of the following reason:

Proposition 4.2. Let G be an srg(n, k, λ, µ). Then 0 ≤ µ ≤ k, with µ = 0 if and only

if G is a disjoint union of equally sized cliques and µ = k if and only if G is a complete
multipartite graph with equal sized parts.
Proof. Exercise.

For non-trivial examples, consider C5 , which is an srg(5, 2, 0, 1) and the Petersen graph,
an srg(10, 3, 0, 1). For infinite families of non-trivial examples, take the line graphs L(Kn )
which is an srg(n(n − 1)/2, 2(n − 2), n − 2, 4), and L(Kn,n ) which is an srg(n2 , 2(n − 1), n −
2, 2).1 Many more examples of strongly regular graphs are given in the exercises, and in
later chapters.
Our first results on strongly regular graphs says that the parameters are not entirely
independent of each other. In fact, it gives us the first condition for the feasibility of the
parameters.
1
As a graph theory exercises, find all graphs G for which L(G) is a strongly regular graph.

Page 41
Strongly Regular Graphs

Proposition 4.3. Let G be an srg(n, k, λ, µ). Then

k(k − λ − 1) = (n − k − 1)µ.
Proof. Fix a vertex x of G. Double count the set of pairs of vertices (y, z) such that x
is adjacent to y, y is adjacent to z and x, z are non-adjacent. There are k choices for y,
and then since x, y have λ common neighbours, there are k − λ − 1 remaining choices for
z. There are n − k − 1 choices for z, since x has exactly k neighbours, and then for every
such z there are µ choices for y.

If try to generalise the C5 and the Petersen graph, both of which had λ, µ = 0, 1 by asking
for an srg(n, k, 0, 1), then from the proposition above we see that n must be equal to
k 2 + 1.

Proposition 4.4. The complement of an srg(n, k, λ, µ) is an srg(n, n − k − 1, n − 2k +

µ − 2, n − 2k + λ).
Proof. Let G be an srg(n, k, λ, µ). The complement of G also has n vertices. The number
of non-neighbours in G of a vertex is n − k − 1. Let x, y be two non-adjacent vertices in G.
They have µ common neighbours, and therefore the number of vertices that are adjacent
to at least one of x or y is equal to 2k − µ. This implies that the number of common
non-neighbours is equal to n − 2 − (2k − µ). Let x, y be two adjacent vertices in G. They
have λ common neighbours, and hence the set of vertices that is adjacent to at least one
of x or y, has cardinality 2(k − 1 − λ). Therefore, among the n − 2 vertices other than
x, y there are n − 2 − 2k + 2 − λ vertices which are non-adjacent to both x and y.

4.2 Moore graphs

Strongly regular graphs are sometimes extremal examples of natural graph theoretical
questions. One of these is as follows.
In a simple graph G, girth of G is the smallest length of a cycle contained in G. If G is
acyclic, then we say that it’s girth is ∞. The diameter of G is the largest distance between
two vertices of G, where the distance is measured by the length of a shortest path. As one
might expect, the graphs with large (but finite) girth, must also have a large diameter.
This is captured by the following relation.

Lemma 4.5. For any connected graph G of diameter d and girth g < ∞, we have:

g ≤ 2d + 1.
Proof. Say g ≥ 2d + 2, and let C be a cycle of length g. Let x, y be two opposite vertices
on C. Since the two paths along C from x to y both have length at least d + 1, and d is
the diameter of the graph, there must be a path P of length at most d from x to y that
does not share all of its edges with C. By combining P and any path from x to y along
C, we get a closed walk of length at most d + g/2 < g, which must contain a cycle. This
is a contradiction to the fact that g is the length of the smallest cycle in G.

A trivial example of graph in which the bound above is tight, is the complete graph on
n ≥ 3 vertices for which g = 3 and d = 1.

Page 42
 Strongly Regular Graphs

Definition 4.6. A Moore graph is a graph of diameter d and girth 2d + 1, for some d > 1.

For example, a C2d+1 is a Moore graph for all d ≥ 2. Can you think of any other examples?
The Petersen graph should be a good guess, and indeed it is a Moore graph of diameter
2. Are there more examples?

Lemma 4.7. Every Moore graph is k-regular, for some k ≥ 2, and has 1 + k + k(k − 1) +
· · · + k(k − 1)d−1 vertices, where d is the diameter of the graph.
Proof. Let d be the diameter of the Moore graph G. The girth of G is then 2d + 1.
We first show that any two vertices x, y which are at distance d from each other have
equal degrees. Let P the unique path of length d joining x and y. There is a neighbour
x1 of x on P and a neighbour y1 of y on P . Let xi be any other neighbour of x. Then
d(xi , y) = d, and the unique path of length d between them gives rise to a neighbour yi of
y. All yi ’s must be distinct, as otherwise we will get a cycle of length ≤ 2d. This shows
that deg(y) ≥ deg(x), and a similar argument shows that deg(x) ≥ deg(y).
Now let C be a cycle of length 2d + 1. Let x, y be two adjacent vertices on C. Then
there exists a vertex z on C at distance d from both x and y. This implies that deg(x) =
deg(y) = deg(z), and hence every vertex on C has the same degree. Let k be this degree.
Let w be a vertex not on C, at distance i from C. Then there exists a vertex on C which
has distance d from w,obtained by taking d − i steps from the nearest vertex in C to w,
implying that deg(w) = k.
The number of vertices follows from an easy count that shows that there are exactly
k(k − 1)i−1 vertices at distance i from a fixed vertex.

Lemma 4.8. Let G be a Moore graph of diameter 2, and k its regularity. Then G is an
srg(k 2 + 1, k, 0, 1).
Proof. We know that G is a k-regular graph on k 2 + 1 vertices. Let x, y be two adjacent
vertices. If they have any common neighbour, then G will contain a triangle, which is
impossible since the girth of any Moore graph is at least 5. Now let x, y be two non-
adjacent vertices. There exists a path of length 2 between them, since the diameter of
the Moore graph is 2. This should be unique path of length 2 as otherwise we will get a
cycle of length 4.

One can prove that Moore graphs of diameter greater than 2 are also “highly regular” in
some sense (look up the notion of distance regular graphs), but we will not delve into
that and just use the theorem of Bannai and Ito from 1971 states that there are no Moore
graphs of diameter d > 2, except for the odd cycles C2d+1 . Therefore, we are only left
with understanding the diameter 2 case, for which the fact that it is a strongly regular
graph with the parameters above, turns out to be an extremely strong condition.

4.3 Spectral Methods

Let A be the adjacency matrix of a (simple undirected) graph G on n vertices, i.e., the
n × n real matrix obtained by ordering the vertex set of G and then defining aij = 1 if the
i-th vertex is adjacent to the j-th vertex, and 0 otherwise. Since A is a real symmetric
matrix of order n, it has n real eigenvalues λ1 ≥ · · · ≥ λn , counted with multiplicity,

Page 43
Strongly Regular Graphs

where λ is an eigenvalue of A if there exists a non-zero vector x such that Ax = λx.

These are precisely the zeros of the degree n polynomial φ(t)
Qn= det(tI − A) = 0, known
as the characteristic polynomial
Q of A. So, we have φ(t) = i=1 (t − λi ). Note that this
implies λi = Tr A = 0 and λi = det A.
P

Two graphs G1 , G2 on the same vertex set V = {v1 , . . . , vn } are isomorphic if and only
if there exist a permutation matrix P such that P T A1 P = A2 , where A1 and A2 are the
adjacency matrices of G1 and G2 . Since P T = P −1 for permutation matrices, we get
det(tI − A1 ) = det(P −1 (tI − A1 )P ) = det(tI − A2 ). Therefore, isomorphic graphs have
the same characteristic polynomial, and hence the same eigenvalues. Moreover, this shows
that the eigenvalues of a graph do not depend on the ordering of the vertices we choose
to write its adjacency matrix.

Definition 4.9. The spectrum of a graph G is the set of distinct eigenvalues of G, along
with their corresponding multiplicities. If the eigenvalues of G are λ1 > λ2 > · · · > λs ,
and their multiplicities are m1 , m2 , . . . , ms , then we write
 
λ1 λ2 . . . λs
Spec(G) = .
m1 m2 . . . ms

By direct computation, or by using some of the results that follow, one can easily compute
the following spectra:  
n − 1 −1
Spec(Kn ) = .
1 n−1
√ √ 
mn 0 − mn
Spec(Km,n ) = .
1 m+n−2 1

The multiplicity of an eigenvalue λ is also equal to the dimension of the eigenspace Vλ =

{x ∈ Rn : Ax = λx}. Moreover, the real spectral theorem asserts that there exists an
orthonormal basis of Rn consisting entirely of eigenvectors, or in other words, the matrix
A can be diagonalised via a change of basis. In particular, it says that for two eigenvalues
λ 6= µ, any vector in Vλ is orthogonal to any vector in Vµ , a fact that can be shown easily.

Lemma 4.10. Let x and y be two eigenvectors of a real symmetric matrix A with Ax =
λx, Ay = µy and λ 6= µ. Then x · y = 0.
Proof. From the definition of the dot product and the fact that A is symmetric, we get
that (Ax) · y = y T Ax = xT AT y = xT Ay = x · (Ay). Since Ax = λx and Ay = µy, we get
λ(x · y) = µ(x · y) which implies that x · y = 0 since λ 6= µ.

Therefore, if  
λ1 λ2 . . . λs
Spec(G) = ,
m1 m2 . . . ms
then
Rn = Vλ1 ⊥ Vλ2 ⊥ · · · ⊥ Vλs .

While the definition of eigenvalues might suggest that to find the eigenvalues of a graph
on n vertices we will be finding the zeros of a degree n polynomial, in practice, this is
something we never do. The way we will find eigenvalues is by looking at the eigenvectors
graph. One can interpret the equation Ax = λx, with x = (x1 , . . . , xn ) as labelling the n

Page 44
 Strongly Regular Graphs

vertices of G by real numbers x1 , . . . , xn such that for any vertex of the graph the sum
of the labels on its neighbours is λ times the label of the vertex. Therefore, finding an
eigenvector is the same as finding such a labelling. This immediately tells us that for any
k-regular graph, the all-one vector (1, . . . , 1) is an eigenvector with eigenvalue k. We can
prove something stronger.
Lemma 4.11. For any k-regular graph G, k is the largest eigenvalue and it has multi-
plicity 1 if and only if G is connected.
Proof. Let V (G) = [n], A the n × n adjacency matrix and let (x1 , . . . , xn ) be an eigen-
vector of A with eigenvalue λ. We first show that |λ| ≤ k. Let m be the vertex for which
|xi | is takes the maximum value. Then looking at the m’th coordinate of the equation
Ax = λx, we get X
xi = λxm ,
i∈N (m)

where N (m) is the neighbourhood of the vertex m, which by k-regularity has size k.
Taking absolute value on both sides, and using the triangle inequality we get
X
|λ||xm | ≤ |xi | ≤ k|xm |,
i∈N (m)

which implies |λ| ≤ k. Therefore, every eigenvalue of G lies in the range [−k, k], and in
particular, since k is an eigenvalue of the graph, it is the largest eigenvalue.
Let G be a disconnected graph, with C ⊂ [n] as one of its components. Then the vector
(x1 , . . . , xn ) with xi = 1 for i ∈ C and xi = 0 for i 6∈ C is an eigenvector with eigenvalue
k, but it is not a scalar multiple of the all-one vector which implies that the multiplicity
of k is at least 2.
Now let G be a connected graph, and let (x1 , . . . , xn ) be any eigenvector with eigenvalue
k. We will show that x1 =P x2 = · · · = xn . Let m be the vertex for which xm is the
largest. Then since kxm = i∈N (m) xi , and xi ≤ xm for all m, we must have xi = xm
for all i ∈ N (m). By repeating this we get that xi = xm for all i, because the graph is
connected.

Therefore, for connected k-regular graphs on n vertices the eigenvalues are k = λ1 > λ2 ≥
λ3 ≥ · · · ≥ λn ≥ −k. The eigenvalues of a graph can be used to characterise several
properties of graphs. You will see some of these in the exercises. For example, you will
show that for connected k-regular graphs, λn = −k if and only if the graph is bipartite.
But now let’s get back to strongly regular graphs, where we can computer the eigenvalues,
and their multiplicities precisely in terms of the parameters of the graphs.
Proposition 4.12. Let G be an srg(n, k, λ, µ), with k ≥ µ > 0. Then G has three distinct
eigenvalues k, θ1 and θ2 with multiplicities 1, m1 and m2 , respectively, where
1 √ 
θ1 , θ2 = λ−µ± ∆
2
and  
1 2k + (n − 1)(λ − µ)
m1 , m2 = n−1∓ √
2 ∆
with ∆ = (λ − µ)2 + 4(k − µ).

Page 45
Strongly Regular Graphs

Proof. Since µ > 0, G must be connected, and hence k is an eigenvalue of multiplicity

1. Let θ be an eigenvalue of A = A(G) with θ 6= k, and let x be an eigenvector with
Ax = θx. The fact that G is an srg(n, k, λ, µ) implies the following:

A2 = kI + λA + µ(J − A − I),

where I is the n × n identity matrix and J is the n × n all-one matrix. Evaluating at x

on both sides, and using the fact that Jx = 0 (since x is orthogonal to the eigenspace
corresponding to the eigenvalue k), we get that

θ2 x = kx + λθx − µ(θx + x).

Since x is non-zero, this implies that

θ2 − (λ − µ)θ − (k − µ) = 0.

The discriminant ∆ of this quadratic equation is (λ − µ)2 + 4(k − µ), and thus we get the
solutions θ1 and θ2 as mentioned above.

Let m1 , m2 be the multiplicities. We know that

1 + m1 + m2 = n,

since there are n eigenvalues counted with multiplicities. Since the sum of all eigenvalues
is 0, we get
k + θ1 m1 + θ2 m2 = 0.
Solving these two equations we get the values of m1 and m2 .

From this we see that the spectrum of the Petersen graph, which is an srg(10, 3, 0, 1), is
 
3 1 −2
.
1 5 4

One of the ways in which we can apply this result to give restrictions on parameters of a
strongly regular graph is by observing that the multiplicities m1 and m2 must be integers.

Corollary 4.13. Let G be a Moore graph of diameter 2, with regularity k. Then k ∈

{2, 3, 7, 57}.
Proof. We know that G is an srg(k 2 + 1, k, 0, 1). The multiplicities of the eigenvalues of
G are
k 2 − 2k
 
1 2
k ±√ .
2 4k − 3
√
One necessary condition for these numbers to be an integer is that (k 2 − 2k)/ 4k − 3
is an integer. One possibility is that k 2 − 2k = 0, i.e, k = 2, which is feasible as shown
by C5 . Say k 2 − 2k 6= 0, then we must have 4k − 3 = m2 for some integer m and
k 2 − 2k = (m2 + 3)2 /16 − 2(m2 + 3)/4 ≡ 0 (mod m). Multiplying both sides by 16, we
get that (m2 + 3)2 − 8(m2 + 3) ≡ −15 ≡ 0 (mod m). Therefore m = 1, 3, 5 or 15, which
gives us k = 1, 3, 7 or 57. Since k > 1, as we assume that the graph has finite girth, the
only possibilities we are left with are k = 2, 3, 7, 57.

Page 46
 Strongly Regular Graphs

Theorem 4.14 (Friendship Theorem). Let G be a graph in which any two vertices have
a unique common neighbour, then G has a vertex which is adjacent to all other vertices,
and as a consequence it is the windmill graph Wn on 2n + 1 vertices for some n ≥ 1.
Proof. Say there is no vertex in G that is adjacent to every other vertex. We claim that
G is regular. Note that G is C4 -free. Let u, v be two non-adjacent vertices, which exist
by our assumption. We know that u, v have a unique common neighbour, call it w. For
every other neighbour of u, we get a unique neighbour of v, and these neighbours must be
distinct since otherwise we will have C4 in the graph. Therefore, deg(v) ≥ deg(u) and by
symmetry we get deg(u) ≥ deg(v). Now let x be any vertex in V (G) \ {u, v, w}. Then x
is non-adjacent to at least one of u or v, since otherwise u and v will have more than one
common neighbours. Hence deg(x) = deg(u) = deg(v). Finally, since w can’t be adjacent
to all vertices, there exists a y ∈ V (G) \ {x, u, v} with deg(w) = deg(y). This shows that
G is regular.
We now derive a contradiction. Let n be the number of vertices in G and k its regularity.
Then G is an srg(n, k, 1, 1), which implies that n = k 2 − k + 1. The multiplicities of the
eigenvalues of G are  
1 2 k
k −k± √ .
2 k−1
Therefore, k − 1 = m2 for some integer m and m divides k. This is only possible if m = 1,
that is k = 2. But then G is a triangle, which does have a vertex adjacent to all other
vertices, a contradiction.
Therefore, G has a vertex adjacent to all other vertices. We leave it to the reader to now
deduce that now G must be the windmill graph.

An interesting corollary of the Friendship theorem is a classical result of Baer on finite

projective planes.

Definition 4.15. Let π = (P, L, I) be a projective plane. A polarity of π is a bijective

map σ : P ∪ L → P ∪ L such that, σ(P) = L, σ(L) = P, σ 2 = id and (x, `) ∈ I if and
only if (σ(`), σ(x)) ∈ I. A point x of π is called an absolute point of σ if x ∈ σ(x).

Remark 4.16. Clearly, a polarity of a projective plane π gives us an isomorphism between

π and π D . An arbitrary isomorphism between π and π D is known as a duality of π, and
then a polarity is a duality which is inverse of itself.

Theorem 4.17 (Baer 1946). Every polarity of a finite projective plane has an absolute
point.
Proof. Let n ≥ 2 be the order of the plane, and assume that there are no absolute points
of a polarity σ. Let G be the graph defined on the n2 + n + 1 points by making x adjacent
to y if y ∈ σ(x). Since there are no absolute points, this defines a simple graph which is
(n + 1)-regular, as the neighbourhood of a point x is equal to the set of points on σ(x).
Moreover, since any two distinct lines in the plane intersect in a unique point, G has
the property that any two distinct vertices have a unique common neighbour. By the
Friendship theorem, G must be the triangle, which is a contradiction to n ≥ 2.
Remark 4.18. Baer also proved that in a finite projective plane of order n every polarity
has at least n + 1 absolute points. The map (a, b, c) 7→ {(x, y, z) : ax + by + cz = 0} is
a polarity of the Desarguesian projective plane PG(2, q), that has exactly q + 1 absolute

Page 47
Strongly Regular Graphs

points, which form a conic if q is odd and a line if q is even. The theorem of Baer can
also be used to show that for odd q every irreducible conic in PG(2, q) has a point.
Remark 4.19. The absolute points form a tangency set because if y ∈ σ(x) for some
absolute point x, and y 6= x, then since σ(y) must contain x it can’t contain y, that is, y
is not an absolute point.
√ Therefore, the number of absolute qpointsq in aq projective plane
of order n is at most n n + 1, with the polarity (a, b, c) 7→ {a x + b y + c z : x, y, z ∈ Fq2 }
in PG(2, q 2 ) showing that equality is achieved for infinitely many values of n.

We end this section with a classic result in spectral graph theory that gives a non-trivial
upper bound on the independence number of a graph, and hence a lower bound on the
chromatic number of the graph.

Theorem 4.20 (Delsarte-Hoffman bound). Let G be a k-regular graph on n vertices and

k = λ1 ≥ · · · ≥ λn its eigenvalues. Then
−λn
α(G) ≤ · n.
k − λn
Proof. Let v1 , . . . , vn be the vertices of G, and A the adjacency matrix. Let S be an
independent set of G, and χ the characteristic vector of S, that is, χ(i) = 1 if vi ∈ S and
0 otherwise. Since S is an independent set, we must have χT Aχ = 0.2 Let u1 , . . . , un be
an orthonormal basis consisting eigenvectors of A, with λ1 , . . . √ , λn as the √
corresponding
eigenvalues. We take u1 to be the normalised all 1 vector, (1/ n, . . . , 1/√ n). Let χ =
, for some c1 , . . . , cn ∈ R. ThenPci = χT ui , and hence c1 = |S|/ n. Moreover,
P
c u
P i2 i
ci = ||χ|| = χT χ = |S|. Since Aχ = ci λi ui , using χT Aχ = 0 we get
n
X n
X
2
0 = |S| λ1 /n + c2i λi 2
≥ λ1 |S| /n + λn c2i = k|S|2 /n + λn (|S| − |S|2 /n),
i=2 i=2

which implies the inequality

−λn
· n.
|S| ≤
k − λn
Since S was an arbitrary independent set, this proves the result.

4.4 Two-intersection sets

In this section we give a finite geometric construction of strongly regular graphs.

Definition 4.21. A set S of points in PG(n − 1, q) is called a two-intersection set3 if

there exist two non-negative integers k1 < k2 such that for every hyperplane π, we have
|S ∩ π| ∈ {k1 , k2 }.

Example 4.22. A hyperoval in PG(2, q) is a two-intersection set with k1 = 0 and k2 = 2.

Example 4.23. A Baer subplane in PG(2, q) is a two-intersection set with k1 = 1 and
√
k2 = q + 1.
2
In general, if S, T are subsets of vertices, and χS , χT are their characteristic vectors, then χTS AχT
counts the total number of edges between S and T .
3
These are also referred to has two-character sets in the literature.

Page 48
 Strongly Regular Graphs

More examples in higher dimensions will be given later.

Theorem 4.24 (Delsarte 1972, Calderbank-Kantor 1986). Let H be a hyperplane in

PG(n, q), n ≥ 3, and S a two-intersection set in H. Then the graph G(S), with vertex
set equal to PG(n, q) \ H, and a point x adjacent to a point y if the line joining x and y
meets H in a point of S, is a strongly regular graph.
Proof. Let S be a two-intersection set in H, with parameters k1 and k2 . The graph is
clearly |S|(q − 1)-regular since there are |S| lines through a point x that meet H in a
point of S, and each of them contains q − 1 points other than x and the point of S the
line contains. Let u, v be two non-adjacent vertices of G(S). Then the line uv intersects
H in a point t 6∈ S. Let w be a common neighbour of u and v. Then the lines uw and
vw meet H in points x and y of S, respectively. The planes spanned by the lines uw and
vw meets H in the line xy, and hence t must be collinear with both x and y. Conversely,
if x and y are two vertices in S contained on a line through t, then the intersection of ux
and vy gives a common neighbour of u and v. This gives a bijectionbetween ordered pairs
(x, y) in S for which the line xy contains t, and the common neighbours of S. We will
show that the number µ of such ordered pairs is a constant not depending on our choice
of u and v.
We count the triplets (x, y, π) where x, y are distinct points of S and π is a hyperplane
of H containing t, x and y. Let θ1 be the number of hyperplanes through t that meet S
in k1 points and θ2 the number of hyperplanes through t that meet S in k2 points. Then
this number is equal to
k1 (k1 − 1)θ1 + k2 (k2 − 1)θ2 .
For the µ pairs (x, y) where the line xy contains t the number of choices for π is n−3
n−2
q
=
since we are counting the number of hyperplanes through a line and for the remain-
n−2
1 q
ing |S|(|S| − 1) − µ pairs there are n−3 such choices since we are counting the number
 
1 q
of hyperplanes through a plane. Therefore we get
   
n−2 n−3
µ + (|S|(|S| − 1) − µ) = k1 (k1 − 1)θ1 + k2 (k2 − 1)θ2 .
1 q 1 q

It now suffices to show that θ1 and θ2 do not depend on the choice of t.

Counting the total number of hyperplanes through the point t in H ∼
= PG(n − 1, q) we
get  
n−1
θ1 + θ2 = .
1 q
Counting pairs (x, π) where x ∈ S and π is a hyperplane of H containing both x and t,
we get  
n−2
k1 θ1 + k2 θ2 = |S|.
1 q
Solving these linearly independent equations we get values of θ1 and θ2 in terms of the
constants |S|, k1 , k2 , n and q, and hence the value of µ.
Now let u, v be any two adjacent vertices, and let t = H ∩ `, where ` is the line joining u
and v. All the points in ` \ {t, u, v} are common neighbours of u and v giving rise to q − 2
such points. The other common neighbours correspond to pairs (x, y), with x 6= y 6= t

Page 49
Strongly Regular Graphs

such that t lies on the line joining x and y, and x, y ∈ S. If max{k1 , k2 } ≤ 2, then there
are no such neighbours, and we get λ = q − 2. Otherwise, we get
   
n−2 n−3
(λ−(q−2)) +((|S|−1)(|S|−2)−λ+q−2) = (k1 −1)(k1 −2)λ1 +(k2 −1)(k2 −2)λ2 ,
1 q 1 q

where λ is the number of such neighbours, and λ1 , λ2 are the solutions of

 
n−1
λ1 + λ2 =
1 q

and  
n−2
(k1 − 1)λ1 + (k2 − 1)λ2 = (|S| − 1).
1 q

If we take S to be a hyperoval of PG(2, q), then G(S) is an srg(q 3 , (q−1)(q+2), q−2, q+2).
In the next chapter we will see how this graph is in fact a special case of another general
construction of strongly regular graphs coming from the collinearity graph of some partial
linear spaces.

4.5 Exercises

1. Let G be an srg(v, k, λ, µ).

(a) Prove that µ ≤ k, with equality if and only if G is a complete multipartite graph.
(b) Prove that the following are equivalent (i) µ = 0, (ii) k = λ + 1, (iii) G is a
disjoint union of complete graphs, (iv) G is disconnected.
2. Determine the spectrum of the graphs Cn and Km,n for all positive integer m, n.
3. Prove that the following graphs are strongly regular and determine their parameters.
(a) Given a set of m − 2 MOLS of order n, m ≥ 2, the graph Lm (n) with vertex set
[n]2 and two vertices adjacent if they either share a coordinate, or have the same
symbol appearing in the coordinates in one of the Latin squares (if m > 2).
(b) The graph on the lines of PG(n, q), n ≥ 3, with two lines adjacent if they meet
in a point.
(c) Let P1 , . . . , P5 , Q1 , . . . , Q5 be disjoint copies of the cycle graph C5 . Label the
vertices of Pi as {vi1 , vi2 , vi3 , vi4 , vi5 } such that vik is adjacent to vi(k+1) . Label
the vertices of Qj as {uj1 , uj2 , uj3 , uj4 , uj5 } such that ujk is adjacent to uj(k+2) .4
For all 1 ≤ i, j, k ≤ 5 connect vik to ujk0 with an edge where k 0 = ij + k (the
arithmetic on all indices is done modulo 5).
4. Let q be an odd prime power, and define a directed graph on Fq by taking (a, b) as
an edge if a − b is a square in Fq .
4
With this labelling Pi ’s look like pentagons whereas Qj ’s look like pentagrams.

Page 50
 Strongly Regular Graphs

(a) Prove that this graph is an undirected graph if and only if q ≡ 1 (mod 4).
(b) Prove that for q ≡ 1 (mod 4), this graph is a strongly regular graph and compute
its parameters.
(c) Find the independence number of the graph when q ≡ 1 (mod 4) and q is an even
power of a prime.
5. We give further characterisations of the Moore graphs.

a) Prove that any k-regular strongly regular graph with µ 6= 0 has at most k 2 + 1
vertices, with equality if and only if it is a Moore graph of diameter 2.
√
b) Prove that a graph on n vertices with girth ≥ 5 has at most n n − 1 edges, with
equality if and only if it’s a Moore graph.
c) Prove that a k-regular graph of diameter d has at most 1+k d−1 i=0 (k−1) vertices,
i
P
with equality if and only if it’s a Moore graph.

6. (a) Let O be an oval in a finite projective plane of order 5. Prove that the graph
defined on the interior points of the plane, that is, the points through which there
are no tangents to O, with two points adjacent if the line joining them is a secant
of O, is isomorphic to the Petersen Graph.
(b) Use this to identify all points and lines of the projective plane with certain sub-
structures of the Petersen graph (vertices, edges, 1-factors, and independent sets),
and thus prove the uniqueness of the projective plane of order 5 assuming it con-
tains an oval.
(c) Show that every projective plane of order 5 has an oval.
7. Let G be a graph with eigenvalues λ1 ≥ λ2 ≥ · · · ≥ λn .
(a) Prove that
xT Ax
λ1 = max xT Ax = max ,
||x||=1 x6=0 ||x||2
and deduce that λ1 ≥ d(G), that is, the average degree of G.
(b) Suppose G is k-regular. Prove that G is bipartite if and only if λn = −k.
8. Let G be a graph with [n] as its vertex set, and let A be its adjacency matrix. Let
C = [2m] for some 1 ≤ m ≤ n/2, and D = [n] \ C, such that the following two
properties hold: (i) the induced subgraph on the vertex set C is regular and (ii) every
vertex in D is adjacent to 0, m or 2m vertices in C. Let
1 
m
J2m − I2m 0
Q=
0 In−2m

(a) Prove that QAQT is an adjacency matrix of a simple graph G0 such that Spec(G0 ) =
Spec(G).
(b) Use this idea to construct a pair of non-isomorphic graphs on 9 vertices that have
the same spectrum.

Page 51
Strongly Regular Graphs

9. Let S be a two-intersection set in PG(2, q), with every line meeting S in either k1 or
k2 points, k1 < k2 .
(a) Show that k2 − k1 divides q.
(b) Prove that |S| is one of the roots of the quadratic equation

x2 − (q(k1 + k2 − 1) + k1 + k2 )x + k1 k2 (q 2 + q + 1) = 0.

Page 52
5 Generalized Quadrangles

5.1 Introduction

Recall that a partial linear space is a point-line geometry where through every pair of
distinct points there is at most one line. So far we have only seen partial spaces that
are linear spaces, that is, through every pair of points there is a unique line. We now
introduce our first family of proper partial linear spaces, the generalized quadrangles,
which are intimately linked with strongly regular graphs, bipartite Moore graphs, and
many interesting combinatorial questions.

Definition 5.1. A generalized quadrangle is a partial linear space satisfying the following
axioms:
(GQ1) for every non-incident point-line pair x, ` there exists a unique point x0 incident
to ` such that x and x0 are collinear,
(GQ2) every point is incident with at least two lines.

For example, the quadrangle, with P = {1, 2, 3, 4} and L = {12, 23, 34, 41}, is a general-
ized quadrangle. An easy generalisation of this is the m × n grid for integers m, n ≥ 1,
where the lines are either the vertical lines (which have n points on them) or the horizontal
lines (which have m points on them).
It is easy to see that the point-line dual of a generalized quadrangle is also a generalized
quadrangle. The dual of an m × n grid is simply the complete graph Km,n .
We will often abbreviate “generalized quadrangle” to GQ. We already notice that unlike
projective planes, the number of points on a line in GQ can be different from the number
of lines through a point. Even though we will show that the grid and the dual grid are in
some sense degenerate examples, we will also see non-degenerate GQ’s where the number
of points on a line is not the same as the number of lines on a point. Even so, in the
non-degenerate case all lines do still have the same number of points, and all points do
have the same number of lines through them, that is, there is some order.

Definition 5.2. A generalized quadrangle is said to have an order (s, t), if there exists
constants s, t ∈ N such that every line is incident with s + 1 points and every point is
incident with t + 1 lines.

If a GQ has order (s, t), then clearly its dual has order (t, s). For example, the symmetrical
n × n grid has order (n − 1, 1) while its dual, Kn,n has order (1, n − 1). Are there any
GQ’s where all the lines have at least three points on them and all the points have at
least three lines through them? Here is our first such example.

Page 53
Generalized Quadrangles

Example 5.3. Let H be a hyperoval in PG(2, 4). Consider the point-line structure where
the points are the points of PG(2, 4) not contained in H, and the lines are the secants of
H.

The above example can also be described as follows: the points are all the 1-factors of
K6 , the lines are the edges of K6 , and a point is incident to a line if the 1-factor contains
the edge.1 This example even has a pretty drawing, which gives it the name “Doily” (see
Figure 5.1).

Figure 5.1: Doily, a GQ of order (2, 2)

Before we proceed with more interesting examples of GQ’s, let’s prove some basic struc-
tural properties.

5.2 Basic properties

Lemma 5.4. (a) If S is a generalized quadrangle in which every point is incident with
exactly 2 lines, then S is a grid.
(b) If S is a generalized quadrangle in which every line is incident with exactly 2 points,
then S is a dual grid.
Proof. It suffices to prove the second statement from which the first follows by duality.
So let S be a generalized quadrangle in which every line is incident with exactly two
points. We can think of S as a simple graph on the vertex set equal to the point set P
where two vertices are adjacent if they are collinear. This graph has no triangles. Let x
be a point of S. The set N (x) of neighbours of x is an independent set since there are no
triangles. Now let z 6= x be a point which is not in N (x). For every y ∈ N (x), zy must
be a line by (GQ2) applied to the point z and the line xy. No two distinct points z1 , z2
in P \ ({x} ∪ N (x)) can then be collinear as that would give rise to a triangle. Therefore,
S is a complete bipartite graph with N (x) and P \ N (x) as its parts.

We now show that almost every generalized quadrangle has an order.

Proposition 5.5. If S is a generalized quadrangle, then one of the following is true:

1
This incidence structure goes back to the work of Sylvester from 1861, and it can also be used to prove
that the group S6 is the only symmetric group Sn which has an outer automorphism.

Page 54
 Generalized Quadrangles

(a) S is a non-symmetrical grid.

(b) S is a non-symmetrical dual grid.

(c) S has an order (s, t).

Proof. It suffices to show that if S is not a grid or a dual-grid then it has an order. For
a point x, let tx + 1 be the number of lines through x. Let x, y be two non-collinear points
of S. For every line through x there exists a unique point on the line collinear with y,
giving us a unique line through y. This is a bijection proving tx = ty .

Now, since S is not a dual grid, the previous lemma shows that there exists a line ` with
at least three points, say u, v, w. Let z be any point on a second line through w. Then
z is non-collinear with u and v proving that tu = tv = tz . Repeatedly applying the same
argument we get tu = tv = tw = t, and that every point on ` is incident with exactly t + 1
lines. Every point not on ` has at least one point in ` it is non-collinear to, proving that
each point of the generalized quadrangle is incident with t + 1 lines.

Using the same argument in the dual generalized quadrangle, and using the fact that S
is not a grid, we can show that every line is incident to s + 1 points for some constant
s.

Proposition 5.6. Let S be a generalized quadrangle of order (s, t). The collinearity
graph G, defined on the points of S by making two points adjacent if they are incident to
a common line, is an

srg ((s + 1)(st + 1), s(t + 1), s − 1, t + 1) .

Proof. Let x be a point of S. It is incident with t + 1 lines, each of which has s points
besides x, giving us s(t + 1) neighbours of x. Let x, y be two collinear points. Then the
only points collinear with both of them are the points on the line joining x and y, giving
us s−1 of them. Let x, y be two non-collinear points. On each of the t+1 lines through x,
there is a unique point collinear with y, and vice versa, giving us t + 1 common neighbours
between x and y. Therefore, the graph is an srg(n, s(t + 1), s − 1, t + 1), where now n
can be computed using Proposition 4.3. Alternately, let ` be a line of the GQ, then the
set of points not incident to ` is the disjoint union of N (x) \ {`}, for x ∈ `, giving us
s + 1 + (s + 1)st points in total.

Corollary 5.7. The eigenvalues of the collinearity graph of a generalized quadrangle of

order (s, t) are s(t + 1), s − 1 and −t − 1, with respective multiplicities

st(s + 1)(t + 1) s2 (st + 1)

1, , .
s+t s+t
Corollary 5.8. If a generalized quadrangle has order (s, t), then s + t divides st(s + 1)(t +
1).

Theorem 5.9 (Higman 1971). If a generalized quadrangle has order (s, t), with s > 1
and t > 1, then s ≤ t2 , and dually t ≤ s2 .
Proof. (Cameron 1975) Let x, y be two non-collinear points. Let V be the set of points
which are not collinear to either x or y, and let T be the set of points collinear with both
x and y. Then |T | = t + 1 and |V | = (s + 1)(st + 1) − 2 − 2s(t + 1) + t + 1 = s2 t − st − s + t.

Page 55
Generalized Quadrangles

For the i-th point z of V , let ti be the number of points of T that z is collinear with. We
then have X
ti = (t + 1)(t − 1)s
by double counting (z, u) ∈ V × T , z collinear with u, and
X
ti (ti − 1) = (t + 1)t(t − 1)

by double counting (z, u, u0 ) ∈ V × T × T , u 6= u0 and z collinear with both u and u0 . By

Cauch-Schwarz inequality we have
X X 2
|V | t2i ≥ ti ,

which simplifies to t(s − 1)(s2 − t) ≥ 0. Since s > 1, we get t ≤ s2 . Arguing the same
way in the dual GQ we get s ≤ t2 .

As an application of the restrictions to the parameters above, let us obtain all values
of t for which there exists a GQ of order (2, t). Firstly, we must have 2 + t dividing
6t(t + 1) = 6(t + 2 − 2)(t + 2 − 1), which implies that t + 2 is a divisor of 12. This gives
t ∈ {1, 2, 4, 10}, but since t ≤ s2 = 4, we know that t = 10 is not possible. It turns out
that for all of these 3 values of t, there is a unique generalized quadrangle of order (2, t).

5.3 Some examples

Example 5.10 (R. W. Ahrens and G. Szekeres 1969). 2 Let H be a hyperoval in π ∼ =

PG(2, q) embedded inside a PG(3, q), for q a power of 2. The following point-line geometry
is a generalized quadrangle of order (q − 1, q + 1):

• Points are all the points of PG(3, q) that lie outside π.

• Lines are all lines of PG(3, q) that are not contained in π and intersect π in a point
of H.

• Incidence is containment.

The proof that the above example is a GQ is pretty straightforward. The fact that it’s a
partial linear space follows from the fact that the line set is a subset of lines of PG(3, q),
which is a linear space. Also note that each line contains q ≥ 2 points. If we take a line
` and a point p not in `, then the span of p and ` is a plane that intersects π in a line `0 .
This `0 contains the point x where ` meets H. Since H is a hyperoval, `0 must contain
another unique point y of H. The line py of PG(3, q) also lies in the plane spanned by p
and `, and hence intersects ` in a unique point p0 . This point p0 is the unique point on `
which is collinear to p, and hence (GQ2) is satisfied. For (GQ1), we can see that through
each point there are q + 2 > 2 lines.
2
They came up with this example, and another one of order (q − 1, q + 1) which works for all prime
powers q, as a way to generalized the incidence structure of “27 lines on a cubic surface”, which in fact
corresponds to the q = 3 case.

Page 56
 Generalized Quadrangles

Remark 5.11. For q = 2 this examples gives us the complete bipartite graph K4,4 , and for
all even q > 2, we get a non-degenerate generalized quadrangle. Ahrens and Szekeres in
fact gave another construction of a GQ of order (q − 1, q + 1) which works for all prime
powers q.

The next example is a GQ of order (q, q) where q is an arbitrary prime power, and the
construction is slightly more involved. The proof is left to the reader, as it is merely a
careful case analysis.
Example 5.12. Let O be an oval in π ∼ = PG(2, q) embedded inside a PG(3, q). The points
are one of the following three types:

(i) the points of PG(3, q) not contained in π;

(ii) the planes of PG(3, q) which meet O in a unique point;

(iii) a new symbol (∞).

The lines are of two types:

(a) the lines of PG(3, q) not contained in π that meet O in a point;

(b) the points of O.

A point of type (i) is incident with no line of type (b), and with all those lines of type
(a) on which it lies in PG(3, q). A point of type (ii) is incident with those lines of type
(a) and (b) which are contained in it in PG(3, q). A point of type (iii) is incident with no
lines of type (a) and with all lines of type (b).
Example 5.13. Let β : F4q × F4q → Fq be the bilinear form defined by β(x, y) := x0 y0 −
x1 y1 +x2 y2 −x3 y3 . For a set S of points of PG(3, q), define S ⊥ = {y ∈ PG(3, q) : β(x, y) =
0, ∀x ∈ S}. Observe that β(x, x) = 0 for all x, and β(x, y) = −β(x, y), which implies
that x ∈ x⊥ for all x, and x ∈ y ⊥ implies y ∈ x⊥ for all x, y.

Let P be the set of all points of PG(3, q) and L be the set of those lines ` which satisfy
` ⊆ `⊥ , i.e., β(x, y) = 0 for all x, y ∈ `. Then (P, L) is a generalized quadrangle of order
(q, q).

To see this first observe that a line ` joining points x and y satisfies ` ⊆ `⊥ if and only if
β(x, y) = 0. One direction is clear from the definition of `⊥ , and the other direction follows
from the fact that β(λx + µy, λ0 x + µ0 y) = λλ0 β(x, x) + µµ0 β(y, y) + λµ0 β(x, y) + µλ0 β(y, x)
for any λ, λ0 , µ, µ0 ∈ Fq . Therefore, for any point x, the set of points collinear with it is
given by x⊥ , which is a plane π in PG(3, q), that contains x. Let ` be a line not containing
x. Then x⊥ , being a plane in PG(3, q), meets ` in a unique point unless ` ⊆ x⊥ . We show
that the latter cannot happen.

Say x ∈ P and ` ∈ L, such that x 6∈ ` and ` ⊆ x⊥ . Let y, z be two distinct points on `.

Let A be the non-singular 4 × 4 matrix over Fq that defines β, that is,
 
0 1 0 0
−1 0 0 0
A=  0 0 0 1


0 0 −1 0

Page 57
Generalized Quadrangles

and β(x, y) = xT Ay. The vectors xT A, y T A and z T A are linearly independent, since x,
y and z are linearly independent. Therefore, the null space of the 3 × 4 matrix B that
has xT A, y T A and z T A as its rows, is 1-dimensional. Note that this null space is equal to
the set of points w for which β(x, w) = β(y, w) = β(z, w) = 0, which implies that there
is a unique such point. But by our assumption x, y, z all satisfy these three equations, a
contradiction.
Remark 5.14. The last two examples are due to Jacques Tits, from his paper of 1959
where he introduced the notion of generalized polygons (of which generalized quadrangles
is a special case). He gave many other classical examples, some of which we will study
later.

In the homework you will construct a GQ of order (q, q 2 ) for all prime powers q. Despite
60 years of research, the only known values of (s, t), s ≤ t, for which we know the existence
of a non-degenerate GQ of order (s, t) are: (q − 1, q + 1), (q, q), (q, q 2 ) and (q 2 , q 3 ), where
q is an arbitrary prime power. Showing the existence of any new GQ of a different order
would be a big breakthrough.
Another interesting open problem regarding generalized quadrangles is showing the exis-
tence or non-existence of GQ’s in which every line as a finite number of points on it, while
through each point there are infinitely many lines, the so-called semi-finite generalized
quadrangles. It is known that GQ(s, ∞) doesn’t exist for s = 2, 3 and 4, where while
the first two cases (due to Cameron and Brouwer) are proved using direct combinatorial
methods whereas the last one (due to Cherlin) employs tools from model theory, which is
an area of log

5.4 An application to Ball Packings

The celebrated four colour theorem says that the chromatic number of any planar graph
is at most 4. How could one generalise this question to higher dimensions? Perhaps, one
could ask the chromatic number of a graph that can be embedded in Rd , for d > 2, using
straight lines. The answer to this question is that the chromatic number of such graphs
cannot be bounded from above for any d ≥ 3, since any graph can be embedded in R3 !
Consider the following question instead. Let B be a finite collection of disks in R2 such
that the interiors of any two disks are disjoint. Define a graph G, known as the contact
graph of B, with vertices as the centres of the disks in B and and two vertices adjacent
if the corresponding disks touch each other at the boundary. Then clearly G is planar,
and hence χ(G) ≤ 4 by the four colour theorem. Determining the chromatic number of
such graphs then seems like a weaker version of the four colour theorem. Surprisingly,
it is equivalent. A famous result of Koebe, Andreev, and Thurston, known as the circle
packing theorem says that for any planar graph we can find a collection B of disks whose
contact graph is isomorphic to the planar graph.
Therefore, we can ask the following question for higher dimensions:

What is maximum chromatic number χ(d) of the contact graph of a finite collection B
of balls in Rd such that any two balls in B have disjoint interior?

Page 58
 Generalized Quadrangles

Such a collection B is known as a ball packing. From four colour theorem we know that
χ(2) = 4. In general it’s not hard to show that χ(d) ≥ d + 2 by picking balls centred
at the vertices of a regular simplex, and its barycenter. One can prove an upper bound
on χ(d) in terms of the so-called kissing number κ(d), which is the maximum number of
unit spheres that can touch a given sphere. Greedily colouring the balls starting from
the one with the smallest radius and its neighbours shows that χ(d) ≤ κ(d) + 13 , and an
easy crude estimate on the kissing number is κ(d) ≤ 3d − 1.4 But what about the lower
bounds? Can we do better than d + 2?

It was shown by Maehara in 2007 that χ(d) ≥ d + 3, for all d ≥ 3, and it was observed by
Hao Chen in 2014 that some existing results imply that χ(d) ≥ d + 4 for all d = 2k − 2,
k ≥ 3. One might suspect that χ(d) is always d + c for some constant c.5 Chen showed
that this is not true by giving an infinite sequence of d’s for which χ(d)−d ≥ f (d) for some
function f (d) = Θ(d2/3 ). The more interesting part of his result for us is that he used finite
geometries, and in particular generalized quadrangles, to obtain this result! The method
was inspired from the counterexamples to Borsuk’s conjecture found by Bondarenko using
strongly regular graphs.

The ball packings that we will construct use geometric representations of strongly regular
graphs. Let G be an srg(n, k, λ, µ) with adjacency matrix A and
 
k θ1 θ2
Spec(G) =
1 m1 m2
√ √
where θ1 = 21 (λ − µ + ∆) > 0 and θ2 = 21 (λ − µ − ∆) < 0, with ∆ = (λ − µ)2 + 4(k − µ).
We will construct a ball packing in Rm1 whose contact graph is G.

Since G is strongly regular, we have

A2 − (λ − µ)A − (k − µ)I = µJ. (5.1)

We can in fact factorise the left and side to write

(A − θ1 I)(A − θ2 I) = µJ. (5.2)

Consider the matrix

E = (A − θ2 I)(I − J/n)

Lemma 5.15. Columns of E are eigenvectors of A with eigenvalue θ1 .

Proof. It suffices to show that (A − θ1 I)E = 0. Plugging in the value of E, we get

(A − θ1 I)E = (A − θ1 I)(A − θ2 I)(I − J/n)

which we can simplify using Equation 5.2, to get

µJ(I − J/n) = µ(J − J 2 /n) = µ(J − J) = 0,

3
Note that at any step the neighbours of a an uncoloured ball of smallest radius r will have radius at
least r.
4
Better exponential bounds are known for κ(d), and its exact value is known only for d = 1, 2, 3, 4, 8, 24.
There is also a lower bound on κ(d) which is exponential.
5
Though looking at the known upper bounds one might suspect this is far away from the truth.

Page 59
Generalized Quadrangles

where we have used that J 2 = nJ.

Therefore, we can see that the column vectors live in an m1 dimensional subspace of Rn .
These vectors, or points in Rm1 , have the useful property that they form a 2-distance set.

Lemma 5.16. Let u1 , . . . , un be the columns of E. Then


a + c, if i = j

ui uj = b + c, if Aij = 1
T

if Aij = 0 and i 6= j

c,


for some numbers a, b, c such that a, b ≥ 0.

Proof. To check the dot products we look at the matrix E T E, whose ij’th entry is equal
to uTi uj . Since E = (A − θ2 I)(I − J/n) = A − θ2 I − kJ/n + θ2 J/n = A − θ2 I − (k − θ2 )J/n,
we have

E T E = E 2 = A2 + θ22 I + (k − θ2 )2 J/n − 2θ2 A − 2(k − θ2 )kJ/n + 2θ2 (k − θ2 )J/n

Simplifying and using Equation 5.1 for the value of A2 , we get

E T E = kI + λA + µ(J − I − A) − 2θ2 A + θ22 I − (k − θ2 )2 J/n,

where the right hand side can be written as aI +bA+cJ, with a = k−µ+θ22 , b = λ−µ−2θ2 ,
and c = µ − (k − θ2 )2 /n.

This lemma implies that p the set u1 , . . . , un of points define a two-distance set since
d(ui , uj ) = ||ui −uj || = ||ui ||2 + ||uj ||2 − 2ui · uj . Moreover, since b+c ≥ c, the distance
is smaller when the i-th vertex is adjacent to the j-th vertex. Therefore, we can define
a ball packing in Rm1 with balls of radius equal to half of the smaller distance, centred
around ui ’s, such that the contact graph is isomorphic to G. So all we need now is the
chromatic number of G to be much larger than m1 . To estimate this we use Hoffman’s
bound, which says that χ(G) ≥ 1 − k/θ2 . There are several examples of strongly regular
graphs for which 1 − k/θ2 > m1 + 3, as one can check on the online database of Brouwer6 .
We look at the compliment of the collinearity graph of a generalized quadrangle of order
(q, q 2 ), which is a strongly regular graph with parameters

(q + 1)(q 3 + 1), q 4 , q(q − 1)(q 2 + 1), (q − 1)q 3 ),

and spectrum  4 
q q2 −q
1 q3 − q2 + q q4 + q2
The lower bound on the chromatic number turns out to be q 3 + 1, whereas the dimension
m1 is equal to q 3 − q 2 + q. Therefore, χ(d) − d ≥ q 2 − q + 1, for all d = q 3 − q 2 + q, where
q is a prime power.
Question: Can this bound be improved further using strongly regular graphs, perhaps
by finding a family where the chromatic number is much larger than the one given by the
Delsarte-Hoffman bound?
6
https://www.win.tue.nl/~aeb/graphs/srg/srgtab.html

Page 60
 Generalized Quadrangles

The following result puts a limit on the best improvement on the lower bound that one
can hope for by using strongly regular graphs, while also providing a useful restriction on
the parameters.
Theorem 5.17. Let S be a two-distance set in Rd . Then |S| ≤ (d + 2)(d + 1)/2.

This bound can be improved to (d + 2)(d + 1)/2 − 1 if we assume that all the points lie
on a sphere, which happens to be the case with the two-distance set we have constructed
from a strongly regular graph.
Corollary 5.18. For any strongly regular graph on n vertices, with
 
k θ1 θ2
Spec(G) = ,
1 m1 m2

we have n ≤ m1 (m1 + 3)/2.

This in particular implies that the number of vertices, and hence the chromatic number, in
the contact graph that we construct using a strongly regular graph is at most a quadratic
function of the dimension, and hence strongly regular graphs can only be used to prove
χ(d) ≥ Θ(d2 ). It is not clear what the actual asymptotic of χ(d) should be.

5.5 Generalized polygons

We saw earlier that a graph of diameter d has girth at most 2d + 1, and defined the
graphs meeting this bound as Moore graphs. If we add the condition that the graph
under consideration is bipartite, that is, it has no odd cycles, then the girth g cannot be
2d + 1, and hence g ≤ 2d. What can we say about bipartite graphs of diameter d and
girth 2d?
For d = 2, complete bipartite graphs are examples of such graphs, and in fact the only
ones. For any d ≥ 2, the cycle C2d is another such example. Are there some “non-trivial”
examples? We will see that many such bipartite graphs come from finite geometry.
Definition 5.19. The incidence graph of a point-line geometry (P, L, I) is a bipartite
graph with parts P and L with an edge between x ∈ P and ` ∈ L if (x, `) ∈ I.
Lemma 5.20. The incidence graph of a partial linear space has girth at least 6.
Proof. Say there exists a cycle of length 4, given by x, `, y, m where x, y are points, and
`, m are lines. Then x and y are contained in two lines, a contradiction to the fact that
we have a partial linear space.
Proposition 5.21. The incidence graph of a projective plane has diameter 3 and girth 6.
Proof. Any two points have a line through them, and any two lines intersect in a point.
Therefore, two vertices in the same part of the incidence graph have distance 2 between
each other. Let x be a point and ` a line. If x ∈ `, then the distance between them in
the incidence graph is 1, and if x 6∈ `, then since x is at distance 2 from any point on `,
the distance between x and ` is 3, which shows that the diameter is 3. To get a cycle of
length 6, consider a triangle x1 , `1 , x2 , `2 , x3 , `3 in the projective plane.

Page 61
Generalized Quadrangles

Proposition 5.22. The incidence graph of a generalized quadrangle has diameter 4 and
girth 8.
Proof. Left to the reader.

We will now show that being a bipartite Moore graphs of the corresponding diameter char-
acterises projective panes and generalized quadrangles. But first we prove some structural
results on bipartite Moore graphs.
Lemma 5.23. Let G be a bipartite graph of diameter d and girth 2d. Then the following
hold:
(a) For any two vertices u and v at distance < d, there is a unique shortest path joining
them.
(b) For any two vertices u and v at distance d from each other, deg(u) = deg(v).
(c) The minimum degree of G is at least 2.
Proof. For (a) observe that two distinct paths of length < d would then give rise to a
cycle that has length strictly less than 2d.
Now let u, v be two vertices at distance d from each other. Let u0 be a neighbour of v.
If d(u0 , v) = d, then we will get an odd cycle of length 2d + 1 since the girth is 2d. This
contradicts the fact that we have a bipartite graph, and hence d(u0 , v) = d − 1. By (a)
there exists a unique path of length d − 1 between ui and v, which contains a neighbour
v 0 of v as its last vertex. These neigbours of v have to be distinct as otherwise we will get
a cycle of length < 2d. Therefore, deg(v) ≥ deg(u). Similarly we get deg(u) ≥ deg(v).
Let C be a cycle of length 2d. Every vertex on C has degree at least 2. Let x be a vertex
outside C, and let P be the shortest path from x to C, that has length i. By going d − i
steps on C we get a vertex y that has distance d from x, and hence deg(y) = deg(x) by
(b). This shows that deg(x) ≥ 2.
Theorem 5.24. A bipartite graph has diameter 3 and girth 6 if and only if it is the
incidence graph of a possibly degenerate projective plane.
Proof. We have already shown one side of this claim. Now let G be such a graph. By
Lemma 5.23, every vertex in G is of degree at least 2, and hence we can identify the parts
of G as points and lines to get a point-line geometry7 , with G as its incidence graph. Let
u, v be two vertices in one of the parts, that is, two points or two lines. The distance
between them must be even since the graph is bipartite, and it can’t be ≥ 4 since the
diameter is 3. Therefore, d(u, v) = 2. Moreover, they must have a unique common
neighbour as otherwise we will get a cycle of length 4, which is not possible in a graph of
girth 6. Thus, the axioms of a possibly degenerate projective plane are satisfied.
Theorem 5.25. A bipartite graph has diameter 4 and girth 8 if and only if it is the
incidence graph of a generalized quadrangle.
Proof. Let G be a bipartite graph of diameter 4 and girth 8. Form lemma 5.23, we know
that every vertex has degree 2. The fact that there are no cycles of length 4 implies that
we have an incidence graph of a partial linear space. We just need to check (GQ1). Let
u be a vertex in one part and v in the other part, such that v is not adjacent to u. This
corresponds to a non-incidenct point-line pair of the point-line geometry. As the diameter
7
recall that in our definition we required each line to be incidence to at least 2 points

Page 62
 Generalized Quadrangles

is 4, we have d(u, v) ≤ 4, but the distance must be odd and hence d(u, v) ≥ 3. There
must be equality as u is not adjacent to v. This gives a unique neighbour of v which is at
distance 2 from u (by Lemma 5.23), proving (GQ1).

These results motivate the following definition.

Definition 5.26. A generalized d-gon, d ≥ 2, is a point-line geometry whose incidence

graph has diameter d and girth 2d.

These point-line geometries were introduces by Jacques Tits in 1959, whose motivation
came from a geometric study of Lie Groups.
It turns out that generalized d-gons, for d > 4, also have the property that either they
are “degenerate”, and can be easily described, or have an order. We prove this after a
sequence of structural lemmas. In all of them, assume that we have a generalized d-gon,
and we refer to the points/lines as vertices of the incidence graph. All distances between
points, or lines, will be measured in this incidence graph.

Lemma 5.27. Any two vertices lie in a cycle of length 2d.

Proof. Let x, y be two vertices, and let P be the shortest path connecting them. Extend
P , using Lemma 5.23(c) and the fact the girth is at least 2d, to get to a path of length d
with z as the other end point. Let z 0 be a neighbour of z which is not in this path. It has
distance d − 1 from x, and hence a path of length d − 1 between x and z 0 gives a cycle of
length 2d containing x and y.

Definition 5.28. A vertex u is called thick if it has degree at least 3. A generalized

polygon is called thick, every vertex has degree at least 3.

Lemma 5.29. Let C be a cycle of length 2d, then any two vertices on C that are at the
same distance from a thick vertex of C have the same degree.
Proof. Let x be the thick vertex, and y, z vertices on the cycle that are at equal distance
i from x. Let x0 be the point at distance d from x on the cycle. Since x is thick, there is
a neighbour u of x which is not on C. The distance between u and x0 is d − 1 and hence
there exists a path P between them which is disjoint from C except for sharing the vertex
x0 . This gives us three internally disjoint paths between x and x0 . On P take a vertex
w at distance d − i from x. Then d(w, y) = d(w, z) = d, and hence by Lemma 5.23(b),
deg(y) = deg(z).

Lemma 5.30. The minimum distance k between any two thick vertices is a divisor of
d. If d/k is even, then thick vertices have at most two degrees. If d/k is odd, then every
thick vertex has the same degree.
Proof. If k = d, then this is a consequence of Lemma 5.23(b). So assume that k < d.
Let u, v be two thick vertices at distance k, from each other, and let C be a cycle of
length 2d containing them. Applying the previous lemma starting at v, we get that every
k-th vertex on C starting from u is thick. Let x0 = u, x1 = v, . . ., xi , . . ., be the thick
vertices on C obtained this way, with xi being the vertex on C at distance ki from u. The
vertex u0 at distance d from u on C has the same degree as u and hence is thick. Write
d = mk + r, for some 0 ≤ r < k. If r > 0, then xm is at distance 0 < r < k from u0 ,
which contradicts the minimality of k. Therefore, d = mk, that is, xm = u0 . The same
argument yields that x0 = u, x1 = v, . . . , xm = u0 , . . . , x2m−1 are all the thick vertices

Page 63
Generalized Quadrangles

on this cycle. Again from the previous lemma, we see that deg(xi ) = deg(xi+2 ) for all i,
with arithmetic modulo 2m. This implies that any thick vertex of C has degree equal to
deg(u) or deg(v). Moreover, if m = d/k is odd, then deg(xm ) = deg(x0 ) which implies
that deg(u) = deg(v) and hence every thick vertex has the same degree.
We can now finish the proof by showing that for any arbitrary thick vertex x, there is a
cycle of length 2d containing x, u and v. WLOG say u is closer to x than v, and let P be
a shortest path between x and u. Extend P to a path P 0 of length d by moving around
C in the direction from u to v. Let y be the end point of P 0 , and z the point on C after
y. Then z is at distance d − 1 from x, and hence there is a path from x to z which is
internally disjoint from P 0 , giving us a cycle C 0 of length 2d containing P 0 . The cycle C 0
which contains z 6= u can only meet C in a path between two thick vertices, and since
there are no thick vertices between u and v, the vertex v must belong to C 0 .

Corollary 5.31. Every thick generalized d-gon has an order (s, t), s > 1, t > 1. More-
over, if d is odd, then s = t.

The main result for thick generalized polygons is the following theorem of Feit and Hig-
man, that can be proved using the theory of distance regular graphs, which we haven’t
introduced in this course, and hence we will skip the proof.

Theorem 5.32 (Feit-Higman 1964). If G is a thick generalized d-gon, then d ∈ {3, 4, 6, 8}.

We will prove a special case instead, using our knowledge of strongly regular graphs.

Proposition 5.33. There are no thick generalized 5-gons.

Proof. Say such a generalized 5-gon exists, and let (s, s) be its order. We will show that
s must be equal to 1. Let G be the collinearity graph on points. Then it is an srg of
parameters (s4 + s3 + s2 + s + 1, s(s + 1), s − 1, 1). Therefore, the eigenvalues are
1 √
θ1 , θ2 = (s − 2 ± s 5).
2
Since the sum of eigenvalues with multiplicities is equal to 0, we must have m1 = m2 as
otherwise we will have an irrational number equal to a rational number. Therefore, we
have m1 = m2 = (n − 1)/2. By taking the sum of the eigenvalues, we now get
1
s(s + 1) + s(s + 1)(s2 + 1)(s − 2) = 0.
2
If we take s ≥ 2, then the left hand side is > 0.
Alternately, the multiplicity of θ1 is

2s(s + 1) + (s2 + 1)(s + 1)(s − 2)

 
1 2
(s + 1)(s + 1) − √ ,
2 s 5
which is an integer if and only if s = 1.

A k-fold subdivision of a graph G is the graph obtained by replacing each edge of G with
a path of length k. It turns out that the non-thick generalized polygons can be described
via subdivisions of thick generalized polygons.

Page 64
 Generalized Quadrangles

Theorem 5.34. The incidence graph of a generalized polygon that is not thick is either
an even cycle, the k-fold subdivision of a multiple edge, or the k-fold subdivision of the
incidence graph of a thick generalized polygon for some k ≥ 2.
Proof. Say G is the incidence graph of a generalized d-gon. Say there are no thick
vertices, then G is C2d . If there is one thick vertex, then there are at least 2 since there is
a cycle C2d containing this vertex in which the opposite point has the same degree. Let
k be the minimum distance between two thick vertices of G.
Say k = d, and let u, v be two thick vertices at distance d from each other. As in the
proof of the Lemma 5.30, any thick vertex x 6= u, v must be contained in a cycle of length
2d with u and v, which will contradict the minimality of k = d. Hence, there are no other
thick vertices. Both u and v must be contained in every C2d of G, and hence G looks like
a collection of (at least three) internally disjoint paths of length k = d from u to v, that
is, it is a k-fold subdivision of a multiple edge.
Now assume that k < d. Then there exist at least 3 thick vertices. Define a graph G0 on
thick vertices by making two of them adjacent if they are at distance k from each other.
We claim that G0 is the incidence graph of a generalized d0 -gon, with d0 = d/k, and G is
a k-fold subdivision of G0 .
Two distinct paths can only have thick vertices in common, and hence G is indeed a k-fold
subdivision of G0 . We first show that every cycle of length 2d in G is a k-fold subdivision
of a cycle of length 2d0 in G0 . In particular, this will show that every vertex in G0 has
degree at least 3. Let C be a cycle of length 2d. All we need to show is that there are
two thick vertices at distance k from each other on C. There is a vertex on C that is at
distance d from a thick vertex, and hence C contains at least one thick vertex x. The
vertex y opposite to x must also be thick, and hence C is the union of two internally
disjoint paths P1 and P2 , between x and y. Now let u, v be two thick vertices at distance
k from each other, and C 0 a cycle of length 2d containing x, u and v. Any neighbour of
x on C 0 is a vertex at distance d − 1 from y. Let P3 be a shortest path from y to this
vertex. Then P3 must contain the thick vertex of C 0 next to x. P3 ∪ P2 contains two thick
vertices at distance k from each other, and hence P1 ∪ P2 = C also does.
The diameter of G0 is at least d0 since on any two thick vertices at distance d from each
other on a cycle C2d of G are at distance d0 from each other in G0 . Now let x, y be two
thick vertices of G that are at distance > d0 from each other in G0 . Let z be a vertex on
a shortest path from x to y which is at distance d0 from x. The thin neighbour z 0 of z,
closer to y, must be at distance d − 1 from x in G giving rise to a cycle C2d containing
x, z, z 0 . Let w be the thick vertex on the nearest thick vertex on this cycle for which z 0 is
between w and z. Then z, w, y show that G0 is not a subdivision of G, a contradiction.
We know that G0 has a cycle of length 2d0 , and any cycle of smaller length will give rise
to a cycle of length smaller than 2d in G, a contradiction.
All that is left to show now is that G0 is bipartite. Let C be a minimum length odd cycle
in G0 , necessarily of length at least 2d0 + 1. Consider two vertices x, y at distance d0 from
each other in G0 on this cycle. Looking at the two thin neighbours of y in G, which are
at distance d − 1 from G, gives rise to a cycle of length 2d in G, which must then be a
subdivision of a cycle of length 2d0 in G0 , and that gives an odd cycle of smaller length in
G0 , a contradiction.

The following are further restrictions known on the order of thick generalized polygons.

Page 65
Generalized Quadrangles

Theorem 5.35. Let S be a finite generalized d-gon of order (s, t), with s, t ≥ 2.

(a) If d = 4, then s ≤ t2 and t ≤ s2 . (Higman)

(b) If d = 6, then st is a square, s ≤ t3 and t ≤ s3 . (Haemers and Roos)

(c) If d = 8, then 2st is a square, s ≤ t2 and t ≤ s2 . (Higman)

The known generalized hexagons have orders (q, q), (q, q 3 ) and (q 3 , q) where q is a prime
power. The known generalized octagons have orders (q, q 2 ) and (q 2 , q) where q = 22k+1 ,
for some k ∈ N.

5.6 Exercises

1. Let H be a hyperoval in π ∼ = PG(2, q) embedded in PG(3, q), for q even. Define

the following point-line geometry S(H, {x, y}), where x, y are distinct points on H.
The points are the points of PG(3, q) not contained in π, the planes through x not
containing y, and the planes through y not containing x. The lines are the lines of
PG(3, q) not contained in π and meeting π in H \ {x, y}. A point and a line are
incident if they are incident as the objects in PG(3, q). Prove that S(H, {x, y}) is a
generalized quadrangle, and determine its order.

2. Find all possible values of t for which there exists a generalized quadrangle of order
(2, t). Prove that there is a unique generalized quadrangle of order (2, 2).

3. An ovoid in PG(3, q) is a set of q 2 + 1 points, no three of which are collinear. Let

O be an ovoid in a hyperplane of PG(4, q). Give a construction of a generalized
quadrangle of order (q, q 2 ) using O, and prove the correctness of your construction.

4. Let β : F4q2 × F4q2 → F4q2 be defined as follows,

β((x0 , x1 , x2 , x3 ), (y0 , y1 , y2 , y3 )) = x0 y0q + x1 y1q + x2 y2q + x3 y3q .

A subspace S of PG(3, q 2 ) is called totally isotropic with respect to β if β(x, y) = 0

for all x, y ∈ S. Prove that the point-line geometry of the totally isotropic points,
and the totally isotropic lines in PG(3, q) is a generalized quadrangle. Also determine
the order of this generalized quadrangle.

5. Let G be a non-empty finite simple graph with the following property:

Every edge xy of G is contained in a triangle x, y, z with the property that any

vertex u 6∈ {x, y, z} is adjacent to exactly one of x, y, z.

(a) Prove that G is either the Windmill graph, or a k-regular graph with k ∈
{4, 6, 10}.

(b) For each of these values of k, construct such a graph G.

(Bonus) Prove (a) without assuming that the graph is finite.

Page 66
 Generalized Quadrangles

6. Borsuk’s conjecture8 states that any bounded subset S of Rd , consisting of at least

2 points, can be partitioned into d + 1 subsets of smaller diameter, for all d ≥ 2.9
(a) Let G be a strongly regular graph on n vertices with
 
k θ1 θ2
Spec(G) = .
1 m1 m2

Construct a finite set S of n points in Rm1 that cannot be partitioned into less
than n/ω(G) parts of smaller diameter, where ω(G) is the clique number of G.
(b) Use the existence of an srg(416, 100, 36, 20) with clique number equal to 5 to
disprove Borsuk’s conjecture.10
7. A connected graph G of diameter d is called distance regular if there exist constants
ai , bi , ci , with i ∈ {0, 1, . . . , d} such that the following holds for any two vertices x
and y of G at distance i from each other:
• there are ai neighbours of y at distance i from x;
• there are bi neighbours of y at distance i + 1 from x;
• there are ci neighbours of y at distance i − 1 from x.
Clearly a0 = c0 = bd = 0 and G is k-regular with k = a0 +b0 +c0 = · · · = ad +bd +cd .
(a) Prove that the collinearity graph of a generalized n-gon of order (s, t) is a
distance regular graph of diameter b n2 c.
(b) Determine the number of points and lines in a generalized n-gon of order (s, t).
8. (a) Prove that a k-regular graph of diameter d has at most
d−1
X
1+k (k − 1)i
i=0

vertices, with equality if and only if the graph is a Moore graph.

(b) Prove that a k-regular bipartite graph of diameter d has at most
d−1
X
2 (k − 1)i
i=0

vertices, with equality if and only if the graph is the incidence graph of a gen-
eralized d-gon of order (k − 1, k − 1).
9. A near 2d-gon, for d ≥ 2, is a partial linear space whose collinearity graph has
diameter d, and for every point x and every line `, there exists a unique point x0 ∈ `
that is closest to x in the collinearity graph, among all points of `.
(a) Prove that every generalized 2d-gon is a near 2d-gon.
8
now known to be false in general
9
You should try to see why d such subsets would clearly not suffice.
10
This strongly regular graph is related to many interesting finite simple groups and finite geometries,
but it’s difficult to give a well motivated construction here.

Page 67
Generalized Quadrangles

(b) Prove that every near 2d-gon that satisfies the following two properties is a
generalized 2d-gon: (i) every point is incident with at least two lines, and (ii)
for every pair of points x and y at distance i ≥ 2 from each other in the
collinearity graph there is a unique neighbour of y at distance i − 1 from x.
10. Let π be the Fano plane PG(2, 2). Construct the following point-line geometry
S = (P, L).
• P is equal to the set of all points, lines, and all point-line pairs (p, `) of π.
• L is equal to the set of all the following 3-element subsets of P: for every
incident point-line pair (p, `) in π, the sets {p, `, (p, `)}, {(p, `), (p1 , `1 ), (p2 , `2 )}
and {(p, `), (p1 , `2 ), (p2 , `1 )}, where p, p1 , p2 are the points on ` in π and `, `1 , `2
are the lines through p in π.

(a) Determine the number of points and the number of lines in S and prove that it
is a partial-linear space of order (2, 2).
(b) Prove that S is a generalized hexagon.

11. (a) Prove that a bipartite k-regular graph of girth g ≥ 4 has at least
(g−2)/2
X
2 (k − 1)i ,
i=0

vertices, and for every g ∈ {4, 6, 8, 12} the bound is sharp for infinitely many
values of k.
(b) For every prime power q, construct a bipartite q-regular graph of girth 6 with
2q 2 vertices, and of girth 8 with 2q 3 vertices.
(c) (Bonus) Improve the constructions from (b) to get as close as you can to the
lower bounds of 2(q 2 − q + 1) and 2(q 3 − 2q 2 + 2q), respectively.

Page 68
6 Polar Spaces

6.1 Introduction

Definition 6.1. A polar space is a partial linear space satisfying the following axioms:
(PS1) for every point x and every line ` not incident with x, either exactly one point of
` is collinear with x or all points of ` are collinear with x.
(PS2) there is no point which is collinear with all points.

A partial linear space satisfying (PS1) but not (PS2) is called a degenerate polar space.
In particular, any linear space is a degenerate polar space. Generalized quadrangles
are examples of non-degenerate polar spaces. We will see that there are several other
examples, and as the name suggests, many (though not all) of them come from polarities.

Definition 6.2. A duality of PG(n, F ) is a bijection δ on the collection of all subspaces

that maps every k-space to an (n − k − 1)-space, such that S ⊆ T if and only if T δ ⊆ S δ .1
A duality δ is called a polarity if δ = δ −1 .

We denote polarities by ⊥, a notation which will become clearer when we talk about
orthogonal polar spaces. Note that if ⊥ is a polarity then S ⊆ T ⊥ implies that T ⊆ S ⊥ ,
and in particular for two points x and y, x lies in the hyperplane y ⊥ if and only if y lies
in the hyperplane x⊥ .

Lemma 6.3. If S and T are two subspaces of PG(n, F ) and ⊥ is a polarity then (S+T )⊥ =
S ⊥ ∩ T ⊥ , and (S ∩ T )⊥ = S ⊥ + T ⊥ .
Proof. Since S, T are subsets of S + T , we get S ⊥ ⊇ (S + T )⊥ and T ⊥ ⊇ (S + T )⊥ .
Therefore, S ⊥ ∩ T ⊥ ⊇ (S + T )⊥ . Since S ⊥ ∩ T ⊥ is a subset of both S ⊥ , and T ⊥ , we get
(S ⊥ ∩ T ⊥ )⊥ ⊇ S + T , and hence (S + T )⊥ ⊆ S ⊥ ∩ T ⊥ .
The proof of the other claim is obtained by replacing S with S ⊥ And T with T ⊥ .

Corollary 6.4. A polarity is completely determined by its action on the points.

Definition 6.5. Let ⊥ be a polarity of PG(n, F ). A projective subspace S is called

isotropic if S ∩ S ⊥ 6= ∅, totally isotropic if S ⊆ S ⊥ and non-isotropic if S ∩ S ⊥ = ∅.

Note that for a point the notion of being isotropic is the same as being totally isotropic.
Such points are also referred to as absolute points, as we did earlier when we defined
polarities of projective planes.
1
In other words, it’s an incidence reversing bijection on the subspace lattice.

Page 69
Polar Spaces

Lemma 6.6. If S is a totally isotropic subspace, then every subspace contained in S is

also totally isotropic.
Proof. Let T ⊆ S, be a subspace. Then S ⊥ ⊆ T ⊥ . And since S ⊆ S ⊥ , we get T ⊆
S ⊥.

Lemma 6.7. Let x, y be two totally isotropic points with respect to a polarity ⊥. Then
the line ` = xy is totally isotropic if and only if x ∈ y ⊥ .
Proof. We have `⊥ = x⊥ ∩ y ⊥ . If x ∈ y ⊥ , then y ∈ x⊥ and hence {x, y} ⊂ `⊥ . But since
`⊥ is a subspace, the span of x and y must then also be in `⊥ . Conversely, if ` ∈ `⊥ , then
x ∈ `⊥ = x⊥ ∩ y ⊥ , and hence x ∈ y ⊥ .

Proposition 6.8. Let ⊥ be a polarity of PG(n, F ). Then the totally isotropic points
and lines with respect to ⊥ form a possibly degenerate polar space. The polar space is
non-degenerate if and only if there is no isotropic point x such that x⊥ containing all the
isotropic points.
Proof. Let x be a totally isotropic point. From the previous lemma we know that for
any other isotropic point y, the line joining them in PG(n, F ) is a line of this point-line
geometry if and only if x ∈ y ⊥ . Therefore, the points collinear with x are precisely the set
of all isotropic points contained in x⊥ . In particular, the polar space is non-degenerate,
that is, satisfies (PS2), if and only if there is no isotropic point x such that all isotropic
points are contained in x⊥ .
Now let ` be any totally isotropic line that does not contain x. Note that every point on
` in PG(n, F ) is an isotropic point, and hence a point of the point-line geometry. Since
x⊥ is a hyperplane, it either contains the line ` or intersects ` in exactly one point, thus
proving (PS1).

When defining polar spaces using a polarity of PG(n, F ) the following is a natural pa-
rameter associated with it.

Definition 6.9. The rank of the polar space formed by a polarity ⊥ of PG(n, F ) is the
maximum vector space dimension of a totally isotropic subspace with respect to ⊥.

We will see later that the dimension of any maximal totally isotropic subspace, is also
equal to the rank of the polar space.

Lemma 6.10. The rank of a polar space formed by a polarity of ⊥ of PG(n, F ) is at most
(n + 1)/2.
Proof. Let k be the vector space dimension of a totally isotropic space S. Then the
vector space dimension of S ⊥ is n + 1 − k. Since S ⊆ S ⊥ , we get k ≤ n + 1 − k, that is,
k ≤ (n + 1)/2.

Recall that the rank of the polar space associated to a polarity of PG(2, q) is 1, as it
cannot have any totally isotropic lines. The examples of these polarities that we have
seen√are (a,√b, c) 7→√` : ax + by + cz = 0 which has q + 1 isotropic points, and (a, b, c) 7→
√
` : a q x + b q y + c q z = 0 which has q q + 1 isotropic points (when q is a square). We
will see more examples of rank 1 polar spaces later. First we consider rank 2 polar spaces,
and show that these give rise to generalized quadrangles.

Page 70
 Polar Spaces

Proposition 6.11. Any non-degenerate rank 2 polar space over PG(n, F ) is a generalized
quadrangle.
Proof. Let x be a totally isotropic point, and ` a totally isotropic line not containing
x. Since x⊥ is a hyperplane, it either meets ` in one point, and thus (GQ1) is satisfied
for this pair, or it contains `. We show that the latter cannot happen. Say ` ⊆ x⊥ . Let
y, z be two points on `, and let π = x ⊕ y ⊕ z be the plane spanned by x and `. Then
π ⊥ = x⊥ ∩ y ⊥ ∩ z ⊥ = x⊥ ∩ (y ⊥ ∩ z ⊥ ) = x⊥ ∩ `⊥ . Since ` ∈ `⊥ and x ∈ x⊥ , we have
π ∈ x⊥ ∩ `⊥ , and hence π ∈ π ⊥ . But this contradicts the fact that the rank of the polar
space is 2.

Now we show that (GQ2) is satisfied. Let x be a totally isotropic point. Since the rank
of the polar space is 2, there exists a totally isotropic line `. Say x 6∈ `. Then by (GQ1)
there exists a line m through x meeting `. Say x ∈ `, then pick a point not collinear with
x and then a line m through that point that meets `. In both cases we have two lines
meeting each other and x lies on one of these lines. Let y be the point of intersection,
and z a point non-collinear with y. Let `0 be the line through z that meets the line m if
x ∈ `, or the line ` if x ∈ m. By (GQ1) there exists a line through x meeting `0 , giving us
the second line through x and thus proving (GQ2).

Many of the examples of generalized quadrangles that we saw were in fact rank 2 polar
spaces coming from polarities of PG(n, Fq ). We will see more soon.

6.2 Collinearity Graph of a Polar Space

We know that the collinearity graph of a generalized quadrangle with an order (s, t) is
always strongly regular. It turns out we can say the same for polar spaces that have an
order.

Theorem 6.12. If S be a polar space of order (s, t) on n points, s, t ≥ 1, which is not

a linear space, then the collinearity graph of S is an srg(n, k, λ, µ) with k = s(t + 1),
λ = k − 1 − (n − k − 1)/s and µ = t + 1.
Proof. Since the order is (s, t), the graph is clearly s(t + 1)-regular. Moreover, if x and y
are two non-collinear points, then every line through x contains a unique point collinear
with y by (PS1). Therefore, µ = t + 1.

Let x, y be two points on a line `. All the s − 1 points in ` \ {x, y} are common neighbours
of x and y. If z is any other common neighbour of x and y, then z must be collinear with
all points of ` by (PS1). Therefore, z is also a common neighbour of every other pair of
points on `. This implies that there exists a constant λ` ≥ s − 1 such that any two points
on ` have the same number of common neighbours.

Now fix a point p on ` and double count the pairs (q, r) where q ∈ ` \ {p} and r is a point
collinear with q but non-collinear with p. For each of the s choices of q, there are k −λ` −1
choices of r. For each of the n − k − 1 choices of r, there is a unique choice of q by (PS1)
and the fact that p, q are non-collinear. Therefore, s(k − λ` − 1) = n − k − 1 which gives us
λ` = k−1−(n−k−1)/s. In particular, λ` does not depend on the choice of `, and the graph
is strongly regular with parameters (n, s(t + 1), s(t + 1) − 1 − (n − s(t + 1) − 1)/s, t + 1).

Page 71
Polar Spaces

6.3 Quadrics and Symmetric Bilinear Forms

We now look at some of the standard examples of polar spaces, coming from polarities of
finite projective spaces.

Definition 6.13. Let F be a field and V a vector space over F > A function β : V ×V → F
is called a bilinear form if β(λu + µv, w) = λβ(u, w) + µβ(v, w) and β(u, λv + µw) for
all λ, µ ∈ F and u, v, w ∈ V . The form is called symmetric if β(u, v) = β(v, u) for all
u, v ∈ V . It is called non-degenerate if for every u ∈ V \ {0}, there exists a v such that
β(u, v) 6= 0.

Lemma 6.14. Let β be a non-degenerate symmetric bilinear form on F n+1 . Then S 7→

S ⊥ = {x ∈ PG(n, F ) : β(x, y) = 0 ∀y ∈ S}, is a polarity of PG(n, F ).
Proof. Say S ⊆ T , then T ⊥ is clearly contained in S ⊥ , since for every x ∈ V , β(x, y) = 0
for all y ∈ T implies that β(x, y) = 0 for all y ∈ S. Therefore, we have an inclusion
reversing map.

Let e1 , . . . , en+1 be a basis, and let bij = β(ei , ej ). Then by bi-linearity, we get that if
(u1 , . . . , un+1 ) and (v1 , . . . , vn+1 ) are the coordinates of u and v, then
X
β(u, v) = ui vj bij = uT Bv,

where B = [bij ](n+1)×(n+1) and we write u, v for the column vectors consisting of the
coordinates. The non-degeneracy implies that the matrix is non-singular, since if there
exists a non-zero v such that Bv = 0, then β(u, v) = 0 for all u. Now let S be a k-
dimensional projective subspace, and let A be a (k + 1) × (n + 1) matrix whose rows form
a basis of S. Then the coordinates of vectors in S ⊥ is equal to the set of solutions of
(AB)x = 0. Since B is non-singular, the matrix AB is also a (k + 1) × (n + 1) matrix of
rank k + 1, and S ⊥ is its null space, which by the rank-nullity theorem has vector space
dimension n − k, and hence projective dimension n − 1 − k. Therefore, S 7→ S ⊥ is a map
from k-dimensional subspaces to n − 1 − k dimensional subspaces of PG(n, F ).

Now (S ⊥ )⊥ = {x ∈ PG(n, F ) : β(x, y) = 0 for all y ∈ S ⊥ }}. Since β(x, y) = β(y, x), we
know from the definition of S ⊥ we know that β(x, y) = 0 for all x ∈ S, that is, S ⊆ (S ⊥ )⊥ .
But the dimension on the right hand side is n − 1 − (n − 1 − dim S) = dim S. This also
shows that ⊥ is a bijection, as it is its own inverse.

Therefore, non-degenerate symmetric bilinear forms give rise to possibly degenerate polar
spaces. In fact, degeneracy of the polar space can only occur in characteristic 2 fields.

Theorem 6.15. Let F be a field of characteristic not equal to 2, and β a non-degenerate

symmetric bilinear form on F n+1 . Then the polarity of PG(n, F ) arising from β gives a
non-degenerate polar space.
Proof. We only need to check (PS2). Let x be an isotropic point and let y 6∈ x⊥ . Then
the line ` joining them meets x⊥ in a unique point, x. An arbitrary point of ` \ {x} can be
written as y + λx for some λ ∈ F . We have β(y + λx, y + λx) = β(y, y) + 2λβ(x, y). We
know that β(x, y) 6= 0 since y 6∈ x⊥ . Pick λ = − 21 β(y, y)/β(x, y), to get a point z = x + λy
for which β(z, z) = 0, but β(x, z) 6= 0.

Page 72
 Polar Spaces

For obtaining finite polar spaces, we let F to be the finite field Fq , with q odd. The
symmetric bilinear forms in this case are completely classified, and thus we know what
the polar spaces look like. It’s convenient to write this classification in terms of a related
notion of quadratic forms, which also has the benefit that for q even quadratic forms also
give us a non-degenerate polar spaces.
Definition 6.16. A quadratic form on a vector space V , over a field F , is a function
Q : V → F such that
(i) Q(λv) = λ2 Q(v) for all v ∈ V , and
(ii) β(u, v) = Q(u + v) − Q(u) − Q(v) is a symmetric bilinear form.

We say that the bilinear form β above is obtained by polarising the quadratic form Q.
The standard examples of quadratic forms, which are in fact the only ones after choosing
a basis, come from evaluations of degree 2 homogeneous polynomials in F [x1 , . . . , xn ].
Definition 6.17. Let Q be a quadratic form over a vector space V . A non-zero vector u
is called singular if Q(u) = 0. A subspace S is called totally singular if Q(u) = 0 for all
u ∈ S. Q is called non-degenerate (or non-singular) if there is no non-zero vector u for
which both Q(u) = 0 and β(u, v) = 0 for all vectors v.

Note that the notion of non-degeneracy of a quadratic form is weaker than just saying
the its corresponding bilinear form β is non-degenerate.
Lemma 6.18. Let Q be a quadratic form and β its corresponding bilinear form. If a
non-zero vector is singular with respect to Q, then it is isotropic with respect to β.
Proof. Let u be a singular vector. Then Q(u + u) = Q(u) + Q(u) + β(u, u). Since
Q(u + u) = 22 Q(u) = 0, we have β(u, u) = 0.
Proposition 6.19. Let Q be a non-degenerate quadratic form over F n+1 . Then the totally
singular points and lines in PG(n, F ) form a non-degenerate polar space.
Proof. Let x be a singular point. Let y be any other singular point. Since Q(x) = Q(y) =
0, we have β(x, y) = Q(x + y). Therefore, if y is collinear with x, then the line joining
them is totally singular, and in particular Q(x + y) = 0, that is, β(x, y) = 0. Conversely,
if β(x, y) = 0, then for any λ, µ we have Q(λx + µy) = λ2 Q(x) + µ2 Q(y) + λµβ(x, y) = 0,
and hence x and y are collinear. This shows that the collinearity is the same as it would
be with respect to the symmetric bilinear form β. Moreover, for any singular point x, the
space x⊥ = {y ∈ PG(n, F ) : β(x, y) = 0} is a hyperplane, and not the whole space, by
non-degeneracy of Q. Hence, (PS1) is satisfied.
For (PS2), let x be a point of Q and let y 6∈ x⊥ be an arbitrary point outside the
hyperplane x⊥ in PG(n, q). Let x + λy be an arbitrary point of the line joining them
(excluding x). We have Q(x + λy) = Q(x) + λ2 Q(y) + λβ(x, y) = λ2 Q(y) + λβ(x, y). If
Q(y) = 0, then we have found singular point which is not collinear with x. If not, then
let λ = −β(x, y)/Q(y), which gives us a singular point not collinear with x. Therefore
(PS2) is satisfied.

The polar spaces associated to a quadratic form are known as orthogonal polar spaces.
The set of singular points of a quadratic form, in the projective space, is known as a
quadric. The rank of a polar space associated to a non-degenerate quadratic form Q is
the maximum vector space dimension of a totally singular subspace.

Page 73
Polar Spaces

Lemma 6.20. Let F be a field of characteristic not equal to 2. For every symmetric
bilinear form β on a vector space over F , the function Q(u) = β(u, u) is a quadratic
form. Moreover, a subspace S is totally singular with respect to Q if and only if it is
totally isotropic with respect to β.
Proof. We have Q(λu) = β(λu, λu) = λ2 β(u, u) = λ2 Q(u). We have Q(u + v) − Q(u) −
Q(v) = β(u + v, u + v) − β(u, u) − β(v, u) = 2β(u, v), which is a symmetric bilinear form
since 2 6= 0. All we need to show is that every totally isotropic space is also totally
singular. Let S be such that β(x, y) = 0 for all x, y ∈ S. Then Q(x) = β(x, x) = 0, for
any x ∈ S.

Therefore, as far as fields of characteristic not equal to 2 are concerned, symmetric bilinear
forms are equivalent to quadratic forms, and we don’t gain anything by studying either
of the two. On the other hand, in characteristic 2 fields we gain something since the
symmetric bilinear forms do not necessarily give a non-degenerate polar space, whereas
quadratic forms always do.

Lemma 6.21. Let Q be a quadratic form and ` a line in PG(n, F ). Then either ` is
totally singular, or it has at most 2 singular points on it.
Proof. Say ` is not totally singular, and let x be a non-singular point on it. Let y ∈ `\{x},
then an arbitrary point of ` \ {x} can be written as y + λx, with λ ∈ F . This point is
singular if Q(y + λx) = Q(y) + λ2 Q(x) + λβ(y, x) = 0, which is a non-trivial degree 2
equation in λ since Q(x) 6= 0 and hence has at most two solutions.

The lines with exactly 0, 1 and 2 singular points on them are sometimes called external,
tangent and hyperbolic lines.

Lemma 6.22. Let f be an irreducible degree 2 homogeneous polynomial in F [x, y]. Then
f has no zeros in PG(1, F ).
Proof. Let f = ax2 + bxy + cy 2 and let (x0 , y0 ) be a zero of f with (x0 , y0 ) 6= (0, 0).
WLOG say x0 6= 0.2 The ax20 + bx0 y0 + cy02 = x20 (a + by0 /x0 + cy02 /x20 ) = 0 and hence
a + by0 /x0 + cy02 /x20 = 0. Then y0 /x0 is a zero of the polynomial g(z) = a + bz + cz 2 ,
and hence g(z) = c(z − z1 )(z − z2 ) for some z1 , z2 ∈ F . Since f (x, y) = x2 g(y/x), This
implies that f (x, y) = cx2 (y/x − z1 )(y/x − z2 ) = c(y − z1 x)(y − z2 x), contradicting that
f is irreducible.

The following classification of quadratic forms over Fq follows from a classical result of
Witt, and we will not give a complete proof here.

Theorem 6.23. Let Q be a non-degenerate quadratic form on an n+1 dimensional vector

space over Fq . Let r be the maximum vector space dimension of a totally singular subspace
with respect to Q. Then there exists a basis in which Q has one of the following forms,
with n + 1 = 2r, 2r + 1 and 2r + 2, respectively.

• Q(x1 , . . . , xn+1 ) = x1 x2 + x3 x4 + · · · + x2r−1 x2r .

• Q(x1 , . . . , xn+1 ) = x1 x2 + x3 x4 + · · · + x2r−1 x2r + ξx22r+1 , where ξ = 1 if q is even,

and ξ = 1 or a chosen non-square if q is odd.
2
In fact, both x0 and y0 should be non-zero.

Page 74
 Polar Spaces

• Q(x1 , . . . , xn+1 ) = x1 x2 + x3 x4 + · · · + x2r−1 x2r + x22r+1 + ξx2r+1 x2r+2 + ηx22r+2 , where

η = 1 and T r(ξ −1 ) = 1 if q is even, and ξ = 0 and −η is a chosen non-square if q
is odd.
Proof. (Sketch) A projective line spanned by two linearly independent vectors u, v is
called a hyperbolic line if Q(u) = Q(v) = 0 and β(u, v) 6= 0. By Chevalley-Warning
theorem (or in fact a direct argument), there exists a non-zero vector u such that Q(u) = 0,
if n ≥ 2. Pick such a vector u, and then pick a v such that u, v span a hyperbolic line W
(this can always be done because of (PS2)). Write V as W ⊕W ⊥ , where ⊥ is defined using
β. The restriction of Q on W ⊥ is still non-degenerate, and then we can use induction. In
this basis consisting of hyperbolic lines the quadratic form looks as above, except until
the last step where we have a subspace X of dimension at most 2 for which Q has no
singular vectors. We then classify this case to get the above where the basis consists of
r hyperbolic lines, and then a basis for X. Note that Lemma 6.22 gives the irreducible
quadratic form on a 2-dimensional subspace, which can be taken to be the particular
forms mentioned in the third case. It can also be shown that this r does not depend on
which hyperbolic lines we use.

The three different quadratic forms, and the polar spaces associated with them, are called
hyperbolic, parabolic and elliptic, respectively.

As a convention, we also associated a parameter  with these spaces, which is equal to −1,
0 and +1, respectively. The role of  will be clear from the following proposition. Note
that the dimension of the underlying projective space of a quadric of rank r and parameter
 is 2r + . These three quadrics, or more precisely the polar spaces associated with them,
are also (somewhat confusingly) denoted as Q+ (2r − 1, q), Q(2r, q) and Q− (2r + 1, q). You
can think of the symbol here as the amount that needs to be added to the dimension n
so that we get twice the rank.

Thanks to this classification, we can compute various properties of quadrics/orthogonal

polar spaces over Fq by looking at these special forms. In particular, we can count the
total number of singular points.

Proposition 6.24. Let Q be a quadric of rank r and parameter . Then the number of
points in Q is
(q r − 1)(q r+ + 1)/(q − 1).
Proof. We can write Q(x1 , . . . , xn+1 ) = x1 x2 + x3 x4 + · · · + x2r−1 x2r + ϕ where, ϕ = 0,
ξx22r+1 or an irreducible degree 2 homogeneous polynomial in x2r+1 and x2r+2 , depending
on  = −1, 0 or 1, respectively, where n + 1 = 2r + . Let f (r) be the number of singular
points in with respect to Q in PG(n, q). Each such point corresponds to q − 1 solutions of
the equation corresponding to Q in Fn+1 q . Moreover, (0, . . . , 0) is also a solution in Fn+1
q .
Partition the set of solutions in Fq by whether Q0 (x3 , . . . , xn+1 ) = x3 x4 + · · · + ϕ is 0 or
n+1

not. If it is 0 then we have 2q − 1 choices for x1 and x2 , whereas if it is non-zero then we

have q − 1 choices. This gives us the following recurrence relation.

(q − 1)f (r) + 1 = (2q − 1)((q − 1)f (r − 1) + 1) + (q − 1)(q 2r−1+ − (q − 1)f (r − 1) − 1).

The recurrence relation simplifies to

f (r) = qf (r − 1) + q 2r−1+ + 1.

Page 75
Polar Spaces

Now f (0) = 0, since a rank 0 polar space has no points in it, and we can easily verify that
looking at these forms as well. This gives us the required value of f (r).

Corollary 6.25. There exists an ovoid in PG(3, q).

Proof. The rank 1 polar space of type +1 with equation Q(x1 , x2 , x3 , x4 ) = x1 x2 +
ϕ(x3 , x4 ), where ϕ ∈ Fq [x3 , x4 ] is an irreducible polynomial, has PG(3, q) as its underlying
space and it has q 2 + 1 points on it. No three points here can be collinear by Lemma
6.21.

To compute the number of lines in an orthogonal polar space, we compute the number
of lines through each point, which gives us the order (q, t) of these spaces. The number
of lines is then equal to the N (t + 1)/(q + 1) where N is the number of points. We get t
immediately from looking at quotient polar spaces.

Let U be a (k + 1)-dimensional subspace of an (n + 1)-dimensional vector space V . Then

the quotient V /U , defined naturally as the vector space on the equivalence classes of
the relation x ∼ y if x − y ∈ U , is an n − k dimensional vector space. Therefore, the
projective space we obtain from this quotient V /U is an n − k − 1 dimensional projective
space. For example, in PG(n, F ), the point-line geometry where points are all the lines
passing through a fixed point x and the lines are all the planes passing through x is in
itself a projective space isomorphism to PG(n − 1, F ). A similar notion of quotient spaces
can be defined for polar spaces as follows.

Let Q be a quadratic form giving rise to an orthogonal polar space of rank r in PG(n, F ).
Let U be a totally singular subspace. Define a quadratic form QU on U ⊥ /U by QU (x+U ) =
Q(x). Then this is well defined since for any u ∈ U , we have QU (x + u + U ) = Q(x + u) =
Q(x) + Q(u) + β(x, u) = Q(x), as Q(u) = β(x, u) = 0. The quotient polar space at U is
defined as the orthogonal polar space on U ⊥ /U with quadratic form QU .

Proposition 6.26. Let S be an orthogonal polar space of rank r. For a point x of S, the
point-line geometry S/x where the points are all the totally singular lines through x and
the lines are all the totally singular planes through x is a polar space of rank r − 1 of the
same type as S with x⊥ /x as its underlying projective space.

Remark 6.27. The recurrence relation f (r) = qf (r−1)+q 2r−1+ +1 can now be interpreted
as x being collinear to qf (r − 1) points and non-collinear to q 2r−1+ points.

Corollary 6.28. The orthogonal polar space of rank r and parameter  is of order

(q r−1 − 1)(q r+−1 + 1)

 
q, .
q−1

For example, the rank 2 orthogonal polar spaces give the generalized quadrangles of order
(q, 1), (q, q) and (q, q 2 ) (embedded inside PG(3, q), PG(4, q) and PG(5, q), respectively).

Corollary 6.29. The collinearity graph graph of an orthogonal polar space of rank r and
type , is strongly regular with parameters

q 2r+ − q r+ + q r − 1
n= ,
q−1

Page 76
 Polar Spaces

q 2r+−1 − q r+ + q r + q(q − 2)

k= ,
q−1
q 2r+−2 − q r+ + q r + q(q − 2)
λ= ,
q−1
q 2r+−2 − q r+−1 + q r−1 + (q − 2)
µ= .
q−1

One interesting thing to observe here is that because√λ − µ = −q r+−1 + q r−1 + q − 2,

the
√ second largest eigenvalue in absolute value is O( k) for  = 0 (and pretty close to
k for the other two values of ), which is in fact (asymptotically) the lowest possible
second largest eigenvalue value any k-regular graph on n > 2k vertices can have. Thus
we get something known as optimally pseudorandom graphs from the collinearity graphs
of parabolic polar spaces.
Another way of getting strongly regular graphs from quadrics is as follows.

Proposition 6.30. Let Q be a quadric of rank r and type  ∈ {−1, +1} in PG(2r + , q).
Then Q forms a two-intersection set.
Proof. Let H be a hyperplane. Then either H is equal to x⊥ for some x ∈ Q or H
intersects Q in a parabolic quadric of PG(2r − 1 + , q).

Another application of orthogonal polar spaces is in the construction of generalized

hexagons. One of the constructions of the only known family of generalized hexagons
of order (q, q), known as the split Cayley hexagons, takes the set of points as the points of
Q(6, q), and through each point x we pick a special set of q+1 totally singular lines through
x forming a totally singular plane. The way these lines are chosen can be described geo-
metrically, via something known as a triality, or algebraically using the standard equation
of the quadric. Quadrics, and polar spaces in general are therefore fundamental objects
in finite geometry. We now look at the classification of all polar spaces coming from
polarities.

6.4 Sesquilinear Forms and Classical Polar Spaces

Definition 6.31. Let V be a vector space over F and σ ∈ Aut(F ). A σ-sesquilinear form
on V is a map β : V × V → F such that

β(u1 + u2 , v) = β(u1 , v) + β(u2 , v),

β(u, v1 + v2 ) = β(u, v1 ) + β(u, v2 ),

β(λu, µv) = λσ(µ)β(u, v),
for all u, u1 , u2 , v, v1 , v2 ∈ V and all λ, µ ∈ F . If σ is the identity then this form is called
bilinear. The form β is called non-degenerate if β(u, v) = 0 for all v implies u = 0, or
equivalently, β(u, v) = 0 for all u implies v = 0.

Proposition 6.32. Every duality δ of PG(n, F ) is induced by a σ-sesquilinear form β on

F n+1 , with S 7→ S δ = {x : β(x, y) = 0 for all y ∈ S}

Page 77
Polar Spaces

Definition 6.33. A sesquilinear form β is called reflexive if β(u, v) = 0 if and only if

β(v, u) = 0 for all u, v.

Lemma 6.34. A duality is a polarity if and only if the sesquilinear form corresponding
to it is reflexive.

Theorem 6.35 (Birkhoff-von Neumann). Let V be a vector space of dimension at least

3 and ⊥ a polarity of PG(V ) Then the reflexive σ-sesquilinear form corresponding to ⊥
is one of the following:
(a) Alternating: σ = 1 and β(u, u) = 0 for all u ∈ V .
(b) Symmetric: σ = 1 and β(u, v) = β(v, u) for all u, v ∈ V .
(c) Hermitian: σ 2 = 1, σ 6= 1 and β(u, v) = σ(β(v, u)) for all u, v ∈ V .

The polar spaces corresponding to quadratic forms or σ-sesquilinear forms are known as
classical polar spaces. In general F need not be a field, and a division ring suffices, but
for our purposes of finite polar spaces we look at Fq . In any finite abstract polar space
one can define a notion of “rank” similar to the one that we have in these classical polar
spaces, and a deep result of Tits shows that every finite abstract polar space of rank ≥ 3
must be one of the classical polar spaces. Therefore, the only polar spaces that are not
completely understood are the rank 2 ones, a.k.a., generalized quadrangles.

6.5 Exercises

1. An ovoid O in PG(n, q) is a set of points with the following two properties: (i) no
three distinct points of O are collinear, and (ii) for all points x ∈ O there exists a
hyperplane Hx such that the set of lines through x that are tangent to O is equal
to the set of lines through x in Hx .
(a) Prove that an ovoid in PG(n, q) has q n−1 + 1 points.
(b) Prove that for every n ≥ 4, PG(n, q) does not contain any ovoids.
2. A bilinear form β : V ×V → F , for some vector space V over F , is called alternating
if β(u, u) = 0 for all u ∈ V . It is called non-degenerate if β(u, v) = 0 for all v ∈ V
implies that u = 0.
(a) Show that if β is an alternating bilinear form then β(u, v) = −β(v, u) for all
u, v.
(b) Prove that the maximum vector space dimension of a totally isotropic subspace
with respect to a non-degenerate alternating bilinear form over a vector space
of dimension n + 1 is equal to (n + 1)/2, and hence deduce that n must be odd.
(c) Prove that the totally isotropic points and lines of PG(n, F ) with respect to a
non-degenerate alternating bilinear form over the underlying vector space F n+1 ,
form a non-degenerate polar space. Also determine the order (s, t) of this polar
space if F = Fq .

Page 78
 Polar Spaces

3. Let S be an orthogonal polar space of rank r and type , over Fq . Prove that the
number of totally singular subspaces of vector space dimension r is equal to
r
Y
(q i+ + 1).
i=1

4. (a) Give an example of a non-degenerate symmetric bilinear form that contains

totally isotropic points and lines, such that these point and lines give rise to a
degenerate polar space.
(b) Prove that a quadratic form over a field of characteristic not equal to 2 is non-
degenerate if and only if the bilinear form associated with it is non-degenerate.
(c) Give an example of a non-degenerate quadratic form Q such that the bilinear
form β associated with it is degenerate.
5. Let β be a non-degenerate symmetric bilinear form over F5q , that gives a polarity ⊥ of
PG(4, q) by mapping a point x to the hyperplane x⊥ = {y ∈ PG(4, q) : β(x, y) = 0}.
Let z be a point of PG(4, q) such that z 6∈ z ⊥ , and let O be an ovoid in z ⊥ ∼
= PG(3, q).
Define a graph G with vertex set equal to the set of points x in PG(4, q) \ (z ⊥ ∪ {z})
such that x lies on a line joining z and a point of O, and making two vertices x, y
adjacent if x ∈ y ⊥ .
(a) Determine the number of vertices n in G and show that the number of edges is
at least Cn5/3 , for some constant C and large enough n.
(b) Prove that G does not contain any copies of the graph K3,3 .3

3
In fact, the graph G is asymptotically the densest possible graph on n vertices that does not contain
a K3,3 .

Page 79