UNIVERSITÀ DEGLI STUDI DI PERUGIA

Dipartimento di Matematica e Informatica

CORSO DI LAUREA TRIENNALE IN MATEMATICA

Codes and Projective Systems

Tesina di: Professore

Davide Turri Bruzzese Rita Vincenti

Anno Accademico 2015/2016

 Codes and Projective Systems

Abstract
In this talk we will shortly recall some basics of coding theory and, after
that, we will show some methods for build linear codes from geometrical
structures of projective spaces such as n-sets, arcs, caps, ovals, hyperovals
and then we will study parameters of that codes. Finally we will catch a
glimpse into a combinatorial generalization of finite projective planes, the
designs, again, introducing a method for build linear codes from it with
some properties of that codes
As a convention, GF (q) is the Galois field of order q, and, given a field K, the
elements of the vector space K n are column vectors, K m×n denotes the set of
matrices of m rows and n coloumns with entries in K; if we replace the field K
with any set A 6= ∅, Am×n again denotes the set of matrices with entries in A.
Definition 1 We call Hamming distance the function
h : GF (q)n × GF (q)n −→ N which assigns to every pair of vectors the number
of positions at which these two vectors are different. We also define a [n, k, d]q
linear code a vector subspace C ≤ GF (q)n of dimension k and minimum distance
d, where d = min{h(u, v)| ∀u, v ∈ C, u 6= v}
Observe that the hamming distance is a distance induced by the norm
w : GF (q)n → N which assigns to every vector the number of entries different
from 0; it’s easy to check that w is a norm and that the minimum distance of a
linear code d coincides with the minimum weight , i.e.
d = min{w(v)| ∀v ∈ C, v 6= 0}.

Definition 2 Given a [n, k, d]q linear code C ≤ GF (q)n we define a generator

matrix of C, A ∈ GF (q)k×n having as rows the (transposed) vectors of a basis
of C, we also define the dual code C ⊥ as the orthogonal complement of C and a
parity check matrix of C, a generator matrix of C ⊥ in GF (q)(n−k)×n (remember
that if C has dimension k then C ⊥ has dimension n − k )

It’s easy to see that if A is a generator matrix of C then C = {A> x|x ∈ GF (q)k }
that is the set of all the linear combinations of the coloumns of A> , or the
(transposed) linear combinations of rows of A. Notice that in general A is not
unique, but we can classify all the possible generator matrices of a code C: given
G generator matrix of C ≤ GF (q)n G0 ∈ GF (q)k×n is another generator matrix
of C iff there exists an automorphism N ∈ GF (q)k×k nonsingular such that
G0 = N G because given a basis of a vector space every automorphism of the

1
space maps that basis to another basis of the same space, and for every two
basis of a vector space there exists an automorphism sending the vector of the
first in the vector of the second in the same order.
For parity check matrices we argue the same: first, the parity check matrix of a
code is not unique in general, and, again, we can classify all the possible parity
check matrices of a code, namely, given a parity check matrix H of C ≤ GF (q)n
H 0 ∈ GF (q)(n−k)×n , H 0 is a parity check matrix of C iff there exists
N ∈ GF (q)(n−k)×(n−k) such that H 0 = N H, second we have that
C = {x ∈ GF (q n )| Hx = 0} where H is a parity check matrix of C

Proposition 3 Let C be a [n, k, d]q linear code and H a generator matrix of

C, the d is the least number of linearly dependent coloumns of H

Proof:
Let H = (H 1 . . . H n ), d0 be the least number of linearly dependent coloumns
of H and c ∈ C a word with minimum weight with nonzero entries ci1 , . . . , cid .
Hence 0 = Hc = ci1 H i1 + . . . + cid H id and so we have d ≥ d0 linearly dependent
coloumns, but if we had d0 < d we would have m < d linearly dependent
coloumns H j1 , . . . , H jm with kj1 H j1 + . . . + kjm H jm = 0 and so we would get
a word k ∈ C with entries kji at their places and the other entries 0, thus
w(c) = m < d which is not possible, and so d = d0 .

Theorem 4 (Singleton Bound)

Let C be a [n, k, d]q linear code, then d ≤ n − k + 1.
If d = n − k + 1 the code is said to be MDS, i.e Maximum Distance Separable.

Proof:
Let H = (H 1 . . . H n ) the parity check matrix of C, with rank n − k and w.l.o.g,
H 1 , . . . , H n−k are the linearly indipendent coloumns which give the rank, so
H 1 , . . . , H n−k+1 are linearly dependent; hence for the prevoius proposition d ≤
n − k + 1.

Definition 5 A matrix L = (lij )i,j ∈ GF (q)n×n is said to be monomial if there

exists a permutation σ ∈ Sn such that for all i, j lij 6= 0 iff j = σ(i)

Definition 6 An automorphism φ of GF (q)n is said to be monomial if w(φ(x)) =

w(x) for all x ∈ GF (q)n

Observe that given two matrices A1 , A2 ∈ GF (q)m×n A2 is obtained from A1

permutating coloumns and multiplying them by some nonzero scalar factors iff
there exists a monomial matrix L = (li j) ∈ GF (q)n×n such that A2 = A1 L; in-
deed obtaining A2 from A1 means that, taken appropriates σ ∈ Sn , λ1 , . . . , λn ∈
GF (q)∗ , we get A2 = (A2 1 . . . A2 n ) = (λ1 A1 σ(1) . . . λn A1 σ(n) ); it’s easy to check
that this operation is equivalent to multiplying A1 by the matrix L = (lij )ij
with entries liσ(i) = λi and the other entries zero, with a little computation we

2
get:
 
0
 ..   
 .  lσ(j)j a1σ(j)
ALj = A · 
   .. σ(j)j
lσ(j)j  =   = λσ(j)j A (1)

.

 . 
 ..  lσ(j)j a(r+1)σ(j)
0

Therefore AL = (λσ(1)1 A1 . . . λσ(n)n An ). Analogously permutating and mul-

tiplying by nonzero scalars the rows of a matrix is equivalent to do a left mul-
tiplication by a monomial matrix. Of course, for the same reason, for any two
vectors u, v ∈ GF (q)n u, v have the same weight iff the entries of v are ob-
tained from the entries of u by permutation and multiplication by some nonzero
scalars, again, iff there exists L ∈ GF (q)n monomial such that v = Lu. If we
put this latter result with a previous observation done on generator matrices we
obtain that two codes C, C 0 ≤ GF (q)n with generator matrices G, G0 respec-
tively are equivalent (by monomial isomorphisms) iff there exist N ∈ GF (q)k×k
nonsingular, M ∈ GF (q)n×n monomial such that G0 = N (GM ).
Proposition 7 Two codes C, C 0 ≤ GF (q)n are equivalent iff C ⊥ and C 0⊥ are
equivalent.
Proof:
(⇒) this simply follows from the adjunction formula which says that (M x) · y =
x · (M > y) for every x, y ∈ GF (q)n where, in this case, M ∈ GF (q)n is the
monomial matrix making C and C 0 equivalent, observing that M is monomial
iff M > is monomial.
(⇐) This follows from the fact that for every code (C ⊥ )⊥ = C

Then, again, two codes C, C 0 ≤ GF (q)n with parity check matrices H, H 0

respectively are equivalent (by monomial isomorphisms) iff there exists N ∈
GF (q)(n−k)×(n−k) nonsingular, M ∈ GF (q)n×n monomial such that H 0 =
(N H)M .

Definition 8 Let P G(r, q) be the r−dimensional projective space over the field
(q). We define an n-set of P G(r, q) the map I : P G(r, q) → Z+ 0 such that
GF P
I(P ) = n
P ∈P G(r,q)

An n-set is a structure that generalizes sets giving to every elements (of

P G(r, q) in that case) an “inclusion multiplicity”, i.e I(P ) = k means that
I contains k copies of P ; with abuse of notation I is representated as I =
{P1 , . . . , Pn } with repetitions allowed. Indeed we can view in a broad sense a
set as an n-set in which every element has inclusion multiplicity 1 or 0.

3
Definition 9 Given I = {P1 . . . , Pn } an n-set of P G(r, q) where for all i Pi =<
pi >, with pi column vector of the homogeneous coordinates, we define a matrix
of I, A = (p1 p2 . . . pn ) ∈ GF (q)(r+1)×n , I is proper if rank(A) = r + 1.

In general the matrix of an n-set is not unique because homogeneous coordinates

of the points are not unique; however we can give all the possible matrices of an
n − set: A, A0 ∈ GF (q)(r+1)×n are matrices of the same n-set I = {P1 , . . . , Pn }
iff A0 is obtained changing the order and the homogeneous coordinates of every
point of I, equivalently permutating the coloumns of A and multiplying them by
some nonzero scalar factors, this computation is equivalent to obtaining A0 from
a multiplication of A by monomial matrix L = (lij ) ∈ GF (q)n×n , i.e A0 = AL

Definition 10 Given I = {P1 , . . . , Pn }, J = {Q1 , . . . , Qn } two n-sets of P G(r, q)

they are equivalent (by projectivity) if there exists a linear projectivity φ : P G(r, q) →
P G(r, q) such that φ(Pi ) = Qi for all i

Clearly this is an equivalence relation; notice that I, J are equivalent iff there
exists M ∈ GF (q)n×n monomial, N ∈ GF (q)(r+1)×(r+1) nonsingular such that
B = (N A)M , where A and B are matrices of I and J respectively (the same
happens between linear codes), because we have a projectivity φ such that
φ(Pi ) = Qi for all i iff we have a matrix N ∈ GF (q)(r+1)×(r+1) such that,
given Pi =< pi >, N pi are homogeneous coordinates of the Qi ’s not in the
same order up to proportionality, this happens iff B 0 = (N p1 . . . N pn ) = N A,
but B, B 0 are matrices of the same n-set iff there exists a monomial matrix
M ∈ GF (q)n×n such that B 0 = BM .
Now we are interested in constructing linear codes from projective system: we
can reach this task in essentially two ways:

Definition 11 Let I = {P1 , . . . , Pn } be an n-set of P G(r, q), Pi =< pi > where

pi are fixed homogeneous coordinates of the points, we define the code generated
by I as the linear code C with generator matrix
G = (p1 . . . pn ) ∈ GF (q)(r+1)×n

Definition 12 Let I = {P1 , . . . , Pn } be an n-set of P G(r, q), Pi =< pi > where

pi are fixed homogeneous coordinates of the points, we define the code generated
by I as the linear code C with parity check matrix
H = (p1 . . . pn ) ∈ GF (q)(r+1)×n

Notice that in general we cannot build an n-set from a generator (or parity
check) matrix of a linear code because even if the rank is maximum, we can have
a zero column vector; neverthless the parity check matrix of a code contains a
zero column iff it has one linearly dependent column and so iff it has minimum
distance 1, thus a code is generated from an n-set iff it’s parity check matrix
does not have zero column, iff it’s minimum distance it at least 2. As we see
from previous results on equivalence between codes and n-sets respectively, even
if we do not have a bijection between the two objects, from equivalent n-sets we
obtain equivalent codes because the equivalence condition between matrices of

4
n-sets and matrices of codes, respectively, is the same.
Now let’s study for both methods the parameters of our codes:
Let I = {P1 , . . . , Pn } be an n-set of P G(r, q), where Pi =< pi > and C the code
built from I, with G generator matrix: C has clearly lenght n, and dimension
k = rank(G) ≤ r + 1 which is the maximum number of linearly indipendent
distincts points of I, and the minimum distance in d = n − l, where l is the
maximum number of (not necessarily distinct) points of I don’t lying on the
same hyperplane each of them counted with the inclusion multiplicity of I:
Let c ∈ C be a nonzero word of C, so c> = a> G where a ∈ GF (q)(r+1) , i.e.
c> = (a> · p1 , . . . , a> · pn ), but, for every i, a> · pi = 0 iff Pi satisfies the equa-
tion a1 x1 + . . . + ar+1 xr+1 = 0 (where a> = (a1 , . . . , ar+1 )), the equation of an
hyperplane, and so the number of nonzero entries of c is the number of points
Pi don’t lying on the hyperplane with equation a1 x1 + . . . + ar+1 xr+1 = 0 again
counted with inclusion multiplicity of I, and so C is an [n, k, d]q linear code.
Now from I = {P1 , . . . , Pn }, n-set of P G(r, q), with Pi =< pi > we take C
the code built from I, with H matrix of I taken as parity check matrix: C has
again length n, but dimension n − k ≥ n − r − 1, where k = rank(H), and for
the proposition 3 d is the least number of linearly dependent coloumns of H
or equivalently the least number of linearly dependent points, not necessarily
distincts, of I.

Let us see, as an example, the following result, which is a part of a bigger

lemma affirming that, although we have an upper bound on the Singleton de-
fect, i.e. the quantity n − k + 1 − d, we do not have a nontrivial lower bound,
because given a code with some parameters, we can build from this a code with
a bigger singleton defect.

Proposition 13 Let C be an [n, k, d]q code with d ≥ 2 (i.e. made from a

projective n-set ), so there exists an [n + 1, k, d]q code

Proof:
Let I = {P1 , . . . , Pn } be an n-set in P G(k − 1, q) with Pi =< pi > defining
the generator matrix of C G = (p1 . . . pn ) and for d = n − l, where the maxi-
mum l is realized by the hyperplane π0 (which is not unique); in π0 we have,
counted with inclusion multiplicity if I, the points Pi1 , . . . , Pil , so we pick a
point P ∈ π0 not in I. Let I 0 = I ∪ {P } which is an n + 1-set of P G(k − 1, q)
and let G0 = (p1 . . . pn p), obviously rank(G0 ) = rank(G) = k, and even if π0
could not be the unique hyperplane realizing the maximum, π0 cointains one
more point of I 0 and so and so the maximum number of points in I 0 lying on an
hyperplane is l + 1, so G0 is the generator matrix of a [n + 1, k, d0 ]q code with
minimum distance d0 = (n + 1) − (l + 1) = n − l = d.

Yet another example: let I = {P1 , . . . , Pn } be the n-set of all the points taken
r+1
one time of P G(r, q), where Pi =< pi > and n = q q−1−1 ; let H = (p1 | . . . |pn ) be
the matrix of I as the parity check matrix of the code C: this code has length

5
n and dimension n − r − 1 because in H occurs the coloumns
   
1 0
0  .. 
e1 =  .  , . . . , e n =  .  (2)
   
 ..  0
0 1

which are the canonical basis of GF (q)r+1 ;

we have that d > 2 because every two distinct points of P G(r, q) have indipen-
dent homogeneous coordinates, as in the previous example, and we obviously
have the linearly dependent coloumns e1 , e2 , e1 + e2 somewhere in H and so
d = 3. In conclusion we get a [n, n − r − 1, 3]q code. If q = 2 this code is the
Hamming code of parameters [2r+1 − 1, 2r+1 − r, 3]2 .

From now on we will consider classical sets of P G(r, q) instead of n-sets, paying
attention to the cases r = 2, 3; indeed, as we said before, we identify sets with
n-sets in which every point of the space is taken 1 time or is not taken at all
and so we can apply the former code-building methods.

Definition 14 A ⊆ P G(r, q) is an (n, s)−set if |A| = n,and s is the maximum

number sucht that every s points of A are indipendent. If s = 3 A is said to be
an n-cap, if s = r + 1 A is an n-arc, or simply an arc.

In P G(2, q) every arc is a cap and viceversa for r + 1 = 3.

Let’s now states some immediate results:

Proposition 15 Let I = {P1 , . . . , Pn } be an arc of P G(r, q) with Pi =< pi >

and C the code with parity check matrix H = (p1 | . . . |pn ), so C is an mds code.

Proof:
H ∈ GF (q)(r+1)×n has rank r + 1, and the minimum distance of C is d = r + 2
because r + 2 coloumns cannot be indipendent, but every r + 1 coloumns are
indipendent by definition of arc and so C is an [n, n − r − 1, r + 2]q code which
is mds.

Of course an n-cap in the projective plane P G(2, q) generates in that way an

[n, n − 3, 4]q code, but in general we cannot see if a cap in P G(r, q) with r > 2
generates an mds code because we cannot calculate exactly the rank of the par-
ity check matrix: the matrix given by a cap has rank 3 ≤ ρ ≤ r + 1, because
i > 3 coloumns linearly dependent are allowed anyway.

Definition 16 A q + 1-arc in P G(2, q) is said to be an oval, a q + 2-arc in

P G(2, q) an hyperoval

Indeed an oval generates a [q + 1, q − 2, 4]q code, an hyperoval a [q + 2, q − 1, 4]q

code.

6
Definition 17 Given O ⊆ P G(r, q) an oval (or an hyperoval) and a line l, l
is:
• a secant if exactly two points of O lie on l
• a tangent if exactly one point of O lies on l
• an external line if no points of O lie on l
Observe that through every point of an oval O passes exactly one tangent:
we already know that for every point of P G(2, q) passes exactly q + 1 lines, but
in the pencil of lines passing through a point P ∈ O, we have the set of lines
{P Q| ∀Q ∈ O \ {P }} which has q elements because every three points of O are
not collinear and so two different Q’s define two different lines, hence we have in
the pencil of lines through P one line left which is a tangent. As an immediate
consequence an oval O has exactly q tangent.
But if we pick any point P in an hyperovalO, as before, we have that the set
of q + 1 lines {P Q| ∀Q ∈ O \ {P }} describes the pencil of lines through P ,
which are all secants, so an hyperoval doesn’t have tangents. We give now some
fundamental theorems about ovals and hyperovals:
Theorem 18 (Qvist’s theorem 1)
Given O ⊆ P G(2, q) oval with q odd, every P ∈
/ O lies on 0 or 2 tangents of O
Proof:
Let P ∈/ O: if P doesn’t lie on any tangent of O we are done; otherwise, if there
exists τ = P R a tangent of O in R, then we have an extra tangent of O passing
through P : If it were not true, i.e. if P had a unique tangent passing through
him, then all the other lines passing through P different from τ would have been
secants or externals, but that’s not true because for the axioms of projective
plane two distinct secants passing through P identify two disjoint couples of
points in O \ {R}. And so all the secants passing through P partition O \ {R}
in couples and so |O \ {R}| = q is even, contrary to the hypothesis that q is
odd. Again, for the same reason every Q ∈ τ \ {R} has an extra tangent passing
through him, and so we have q tangents of O against q points in τ \ {R}: for
the pidgeonhole principle every point in τ \ {R} has only one extra tangent.
Theorem 19 (Qvist’s theorem 2)
Given O ⊆ P G(2, q) oval with q even, every P ∈ / O lies on 1 or q + 1 tangents
of O, and so all the tangents of O are concurrent in one point called the nucleus
of O
Proof:
Let P ∈/ O, we have two cases:
• If P doesn’t lie on any secant of O, then in the pencil of lines passing
through P we have the set of lines {P Q| ∀Q ∈ O} which are, for the
case hypothesis, all tangents of O and the P Q lines are q + 1 (so they
individuate all the lines through P and all the tangents of O) then we are
done.

7
 • If P lies on the secant QR for Q, R ∈ O, then there must be a tangent of
O passing through P , if it were not so, the we would have any line passing
through P to be secant or external, hence all the secant throguh P would
have partititioned O in disjoint couples as we claimed in the proof of the
previous theorem, but we assumed |O| = q + 1 is odd, and so this is not
possible. For the same reason every point on the line QR there has a
tangent passing through him and so we get again q + 1 tangents of O
against q + 1 points of QR, and so for the pidgeonhole principle we must
have for every point of QR exactly one tangent passing through him.

An absolutely irreducible conic C : f = 0 does not have linear components

because f ∈ GF (q)[x0 , x1 , x2 ] homogeneous of degree 2 does not have non-
invertible proper divisor which, in this case, are homogeneous polynomial of
degree 1, giving us us equation of lines; and so for Bezout’s theorem for every
two points P1 , P2 ∈ C the line l = P1 P2 is not contained on the curve C, and so
C is a cap. We know by properties of algebraic curves in the projective planes
that every absolutely irreducibile conic is equivalent by projectivity to the curve
C0 : x1 2 − x0 x2 = 0 with support C0 = {(1 : t : t2 )| ∀t ∈ GF (q)} ∪ {(0 : 0 : 1)}.
Moreover this curve is nonsingular: we can easily verify that for every point
P = (p0 : p1 : p2 ) ∈ C0 the gradient of the equation evaluated in P is
(−p2 , 2p1 , −p0 ) 6= (0, 0, 0) (if q is odd this gradient is never zero in P G(2, q),
if q is even it is zero only in (0 : 1 : 0) ∈ / C0 ) and so every point is simple,
with the unique tangent lP : p2 x0 − 2p1 x1 + p0 x2 = 0. The curve C0 has
projective parametrization as which has clearly q + 1 elements; so, as a projec-
tivity is a bijection, every irreducible conic can is an oval. Indeed we conclude
that in any projective plane P G(2, q) there exists an oval, namely the conic
C0 : x1 2 − x0 x2 = 0. In general, in the case q even, if we add to an oval O
his nucleus X we have an hyperoval by the second Qvist’s theorem, so, if q is
even, P G(2, q) has an hyperoval; in particular, if we pick P1 = (1 : 0 : 0) and
P2 = (0 : 0 : 1), having tangents l1 : x0 = 0 and l2 : x2 = 0 respectively, their
tangents intersect in X = (0 : 1 : 0), hence by Qvist’s theorem 2) this point is
the nucleus of C0 and C0 ∪ {X} is an hyperoval.
However, as the next theorem states, we cannot extend any oval or hyperoval,
depending on the case q even or q odd, to a greater cap; hence we are given the
maximum cardinality of an arc in P G(2, q), and so the maximum cardinality of
a code built from an arc.

Theorem 20 (Bose’s theorem)

Let maxs (r, q)(be the maximum size of an (n, s)-set in P G(r − 1, q).Then
q + 1 if q odd
max3 (3, q) =
q + 2 if q even

Proof:
Let A ⊆ P G(2, q) be an (n, 3)-set, so for any point P ∈ A in every line passing
through P there is at most one more point of A, so counting the occurrence of
the points of A \ {P } in the pencil of lines through P we have n − 1 ≤ q + 1,

8
then n ≤ q + 1, and so maxs (r, q) ≤ q + 2. We have now to prove that if
maxs (r, q) = q + 2 then q is even:
If A0 ⊆ P G(2, q) is an (q + 2, 3)-set, then for a point P ∈
/ A0 , every line l passing
through P l passes exactly through two or zero points of A0 , remembering that
an hyperoval doesn’t have tangents, as we stated in the proofs of Qvist’s theo-
rems, the secants through P partition A0 in disjoint couples, and so |A0 | = q + 2
is even and so q is even.

These methods shown above are not the unique methods for generating codes
from geometries: we can generate a code directly from the inner structure of a
finite projective plane, jumping to a wider combinatorial object that generalizes
projective planes as incidence strctures to generate linear codes.

Definition 21 Let X 6= ∅ 6= B be a finite set and a finite n-set respectively

called set of points and blocks, we call design the triple D = (X, B, I) where
I ⊆ X × B is a binary relation called incidence. D is called simple if B does not
contain repeated blocks.

Definition 22 Let D = (X, B, I) be a design, D is called a (v, k, λ)-BIBD with

v > k ≥ 2, λ > 0 if
• |X| = v
• |B| = k i.e. every block is incident with exactly k elements of X
• for every x, y ∈ X with x 6= y {x, y} is incident with exactly λ blocks

BIBD stands for Balanced Incomplete Block Design, which is symmetric if v = b.

Indeed we are allowed to identify the binary relation of incidence with the set
inclusion and the blocks with subsets of X, so we write x ∈ B saying that the
point x is incident with the block B.
Observe that a finite projective plane P = (P, L) of order n is a (n2 + n +
1, n + 1, 1)-BIBD because every line is incident with exactly n + 1 points, for the
axioms of projective plane |P| = n2 + n + 1, and for every two distinct points
passes one line, and so, for every q, P G(2, q) is a (q 2 + q + 1, q + 1, 1)-BIBD.
Now let us states some useful properties of BIBD’s:

Proposition 23 Let D = (X, B, I) be a (v, k, λ)-BIBD, so every point of D is

contained in r = λ(v−1)
k−1 blocks

Proof:
Given a point x, let rx be the number of blocks incident with x, we define
I = {(y, B)| y ∈ X, y 6= x, B ∈ B, {x, y} ⊆ A}, so we count the elements of this
set in two different ways:
• every fixed block B contains k − 1 y’s, thus |I| = rx (k − 1)
• every fixed point y 6= x is contained is λ blocks, hence |I| = λ(v − 1)

9
 λ(v−1)
Compairing these two results we get λ(v − 1) = rx (k − 1) and so rx = k−1 ,
λ(v−1)
but rx doesn not depend on x and so we get r = k−1

vr
Proposition 24 Let D = (X, B, I) be a (v, k, λ)-BIBD, so D has b = k =
2
λ(v −v)
k2 −k blocks

Proof:
Given b = |B|, we define I = {(x, B)| x ∈ X, B ∈ B, x ∈ B}; again we count the
elements of I in two ways:
• every (fixed) x is contained in r blocks B, so |I| = rv
• every (fixed) B contains k points x, hence |I| = bk
vr
again, compairing these two results, we get bk = |I| = rv, and so b = k .

An immediate consequence is that for every (v, k, λ)-BIBD we have that

λ(v − 1) ≡ 0 mod (k − 1) and v(v − 1)λ ≡ 0 mod k(k − 1), that is a necessary
condition on parameters.

Definition 25 Given D = (X, B, I) a (v, k, λ)-BIBD, with a labelling of points

and blocks X = {x1 , . . . , xv }, B = {B1 , . . . , Bb }, we(define the incidence matrix
1 if xi ∈ Bj
of D, M = (mij )ij ∈ {0, 1}v×b with entries mij =
0 otherwise

Of course we can see the incidence matrix to have her entries in the field
we need to use, and a general design is uniquely determined by its incidence
matrix. Observe that for every (v, k, λ)-BIBD, named M its incidence matrix
we have that:
1. every column contains k ones
2. every rows contains r ones
3. two distinct rows have apiece λ ones in the same column

Theorem 26 (Ryser)
Given a design D = (X, B, I) with 2 ≤ k < v, and M ∈ GF (q)v×b its incidence
matrix, M M > = λJv + (r − λ)Iv iff D is a (v, k, λ)-BIBD and 1v > M = k1b > ,
where Jn is a square matrix of order n, and 1n is a vector of length n both having
all entries 1 ∈ GF (q)

Proof:
(⇒) If D is a (v, k, λ)-BIBD then, because of the observation made
( on the in-
b r if i = j
cidence matrix, L = M M > has ij-th entry bij =
P
mil mlj =
l=1 λ otherwise
because in the first case we count the elements incident with a block, in the
second the number blocks incident two distinct elements; thus we can write

10
L = λJv + (r − λ)Iv . The multiplication of 1v > by a column of M is equivalent
to counting the number of ones in that column, but every column of M has
exactly k ones, so 1v > M = k1b > .
(⇐) Clearly from the size of M we get |X| = v and |B| = b. As before, the
equality 1v > M = k1b > gives us the number of ones in every column of M , which
is the number of blocks incident with a point, equal to k for hypothesis; from
M M > = λJv + (r − λ)Iv we have that every entry outside the main diagonal is
λ, but again, this entry count the number of lined-up ones in two distinct rows,
that is, the number of blocks incident with a pair of distinct points. Hence D
is a (v, k, λ)-BIBD.

Proposition 27 If D is a (v, k, λ)-BIBD det(L) = (r − λ)v−1 (r + (v − 1)λ) =

(r − λ)v−1 rk

Proof:  
r λ ... λ
λ r . . . λ
If M is the incidence matrix of D, then L =  . ..  so we make
 
 .. ..
. .
λ ... λ r
two steps: first we subtract the first row from every other rows, Li 7→ Li − L1
for every i > 1 and we add the first column to every other column Lj 7→
j 1
L
 + L for every j > 2 and we obtain  an upper diagonal matrix of the form
r + (v − 1)λ λ ... λ

 0 (r − λ) . . . λ  
 .. .. .. 
 . ... . . 
0 ... 0 (r − λ)
Hence to compute determinant we multiply all the entries on the diagonal of
the transformed matrix.

Now we are ready to construct codes from designs and study some of their
basic properties:
Definition 28 Given D = (X, B, I) a design with incidence matrix M we define
the linear code C = {M > x| x ∈ GF (q)v } ≤ GF (q)b
Observe that if D is symmetric and p - det(L), then of course rank(M ) = v, but
if p - v and p|det(L), so det(L) = 0 in GF (q), then rank(M ) = v −1: in this case
we can diagonalize L giving its v eigenvectors which form a basis of GF (q)v :
indeed L is diagonalizable if it is a linear combination of a diagonalizable matrix
and the identity matrix, so here we have to check that Jv is diagonalizable. Jv
has one eigenvector 1v with eigenvalue v and for the eigenvalue 0 we have the

11
eigenvectors:
     
1 1 1
−1 0  0
     
0 −1  0
 ;  ... (3)


 ..   ..   ..

 .   .   .

0 0 −1

All this vectors (1v included) are linearly indipendent, because if we put them
together in a matrix and we get an upper triangular matrix summing to the first
column all the other columns we get the determinant (−1)v−1 v 6= 0 and so Jv
is diagonalizable, hence with linear algebra considerations we obtain that L is
diagonalizable with eigenvalues vλ + r − v together with the eigenvector 1v and
r−λ with multiplicity v−1 together with all the other eigenvectors shown above.

Furthermore, we can give a bound on the dimension of the code in some cases:
for example if p|r − λ then L is singular, and if p|r, then, for the proof of Ryser’s
theorem, and for the congruence r ≡ λ ≡ 0 mod p, two any rows are orthogonal,
so the columns of M > are in C ⊥ , and so their span which is C; for C ⊆ C ⊥ we
get dimC ≤ dimC ⊥ , then dimC ≤ v − dimC, so dimC ≤ v2 .

References:
• Albrecht Beutelspacher, Uta Rosenbaum: Projective Geometry: From
Foundations to Applications. Cambridge University Press, 1998
• Massimo Giulietti Appunti del Corso di Codici e Crittografia: Campi
finiti Codici lineari Codici algebrico-geometrici Crittografia su Curve El-
litticheA.A. 2013/2014
• Douglas R. Stinson Combinatorial designs : constructions and analysis,
Springer 2004
• Thomas Beth, Dieter Jungnickel, Hanfried Lenz Design theory Vol. 1
Second Edition, Cambridge University Press 1999
• Tom Edgar, Finite Projective Geometries and Linear Codes (In partial ful-
fillment of the requirements for the degree of Master Science Fort Collins,
Colorado Spring 2004)
• Matthew Brown (Hyper)ovals and ovoids in projective spaces (Socrates
Intensive Course Finite Geometry and its Applications Ghent, 3−14 April
2000)
• Simeon Ball, Zsuzsa Weiner An Introduction to Finite Geometry 5 Septem-
ber 2011
• William G. Bridges, Marshall Hall Jr. , John L. Hayden Codes and De-
signs, Journal of Combinatorial Theory, Series A 31, 155-174 (1981)

12
• Katarzyna Kiczka Codes and Projective Systems Combinatorial Geometry
AA 2011-2012

13