Manuscript submitted to Website: http://AIMsciences.

org
AIMS’ Journals
Volume 4, Number 3, August 2010 pp. 405–417

LDPC CODES ASSOCIATED WITH LINEAR

REPRESENTATIONS OF GEOMETRIES

Peter Vandendriessche
Boudewijn Hapkenstraat 5
8820 Torhout, Belgium

(Communicated by Marcus Greferath)

Abstract. We look at low density parity check codes over a finite field K
associated with finite geometries T2∗ (K), where K is any subset of PG(2, q), with
q = ph , p 6= char K. This includes the geometry LU (3, q)D , the generalized
quadrangle T2∗ (K) with K a hyperoval, the affine space AG(3, q) and several
partial and semi-partial geometries. In some cases the dimension and/or the
code words of minimum weight are known. We prove an expression for the
dimension and the minimum weight of the code. We classify the code words of
minimum weight. We show that the code is generated completely by its words
of minimum weight. We end with some practical considerations on the choice
of K.

1. Introduction. Originally introduced by Gallager [5], low density parity check

(LDPC) codes are used frequently today due to their excellent empirical perfor-
mance under belief-propagation/sum-product decoding [19]. In some cases, their
performance is even near to the Shannon limit [19]. In general, a binary LDPC
code C is a linear block code defined by a sparse parity check matrix H, this is a
matrix that contains a lot more 0s than 1s.
To exploit structural properties, one usually wants an explicit construction rather
than just random matrices. Lately there have been many different constructions:
based on permutation matrices [4],[31], Ramanujan graphs [20],[24], expander graphs
[28], q-regular bipartite graphs [14] or other incidence structures in discrete mathe-
matics. In particular, one can take the incidence matrix of a finite geometry as the
parity check matrix of a code.
Examples of such codes can be found in [9, 10, 11, 12], [18], [34]. Later, simula-
tion results of Liu and Pados [16] showed that several generalized polygon LDPC
codes have powerful bit-error-rate performance when decoding is carried out via
low-complexity variants of belief propagation. It would be interesting to perform
the same simulations for the incidence geometries studied in this article, since all
handled structures have a girth of at least 6 in their associated Tanner graph. If K
is an arc, then the Tanner graph even has girth at least 8.
One class of geometries studied for this purpose are linear representations of
geometries. One case that received a lot of attention lately is T2∗ (K), with K a
hyperoval [23, 33]. Here the minimum weight is known, the dimension is known

2000 Mathematics Subject Classification. Primary: 51E15, 51E20, 94B05.

Key words and phrases. LDPC codes, linear codes, finite geometry, linear representation, min-
imum distance, dimension.

405
406 PETER VANDENDRIESSCHE

when the characteristic of the code field char K 6= 2 and then we also know that
the code is generated by its code words of minimum weight. Other structures
studied in less detail in [23] include T2∗ (B) with B a Baer subplane, T2∗ (U) with
U a unital, and T2∗ (L) with L the pointwise union of two intersecting lines. One
linear representation that has received a lot of attention is LU (3, q)D (and its dual
LU (3, q), which is not a linear representation) [13],[14],[27],[33]. In [14], the authors
conjecture the binary dimension of the associated code to be
q 3 − 2q 2 + 3q − 2
2
h
if q is odd. Here q = p with p prime denotes the order of the finite field of the
geometry. This conjecture was proven in [27]. Over all code fields with char K 6= p,
this was proven in [33]. In this paper we present a uniform approach, having both
of these as an immediate corollary.
In this article we study the general problem of codes associated with linear rep-
resentations of geometries when char K 6= p. We generalize several of the above
results, including the main theorems in [27] and [33]. When char K 6= p, we com-
pute the minimum weight, the rank and the code rate of the code, we classify
the code words of minimum weight and we prove that every code word is a linear
combination of the code words of minimum weight. We end with some practical
considerations on the choice of K.

2. Preliminaries. Let us begin by introducing some basic notations and defini-

tions.
Notation 2.1. A finite field Fq of order q has q = ph elements, where p is a prime
number and h is a positive integer. The characteristic of that field is p. We denote
by char F the characteristic of the field F .
Notation 2.2. A linear [n, k, d]-code C over the field Fq can be defined by a parity
check matrix H, this is an (n − k) × n matrix with the property that a word
c = (c1 , . . . , cn ) is a code word of C if and only if HcT = ~0 over Fq , i.e. if and only
if HcT ≡ 0 (mod p), where q = ph with p prime and h a positive integer. In this
paper we let H be the incidence matrix of a geometrical structure, hence its entries
are only the elements 0 and 1 of the field Fq .
Notation 2.3. We denote by PG(n, q) the n-dimensional projective space over
the finite field Fq . For n = 2, we call this a projective plane and write PG(2, q).
We denote by AG(n, q) the n-dimensional affine space over the finite field Fq . A
hyperoval is a set of q +2 points in PG(2, q) such that no three of them are collinear.
Hyperovals exist if and only if q is even. More background on (substructures of)
projective and affine spaces can be found in [7].
Definition 2.4. Let PG(3, q) be the 3-dimensional projective space over the field
Fq . Let Π0 := PG(2, q) be a (hyper)plane in it and let K be an arbitrary subset of
the points of that hyperplane. We define the geometry T2∗ (K) as follows:
• the points of T2∗ (K) are the affine points, being the points of PG(3, q)\PG(2, q),
• the lines are the affine lines of PG(3, q) which go through a point of K,
• the incidence relation is inherited from PG(3, q).
Remark 2.5. Note that through every (affine) point we have |K| lines, one through
each point of K, while every line contains q points. In total there are q 3 points and
|K|q 2 lines: q 2 through each point of K.
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 407

Remark 2.6. Let N = |K| and let H be the q 3 × N q 2 incidence matrix of T2∗ (K),
where points correspond to rows of H and lines correspond to columns of H and to
the positions in the code. Let C be the linear code with H as its parity check matrix,
over an arbitrary finite field K. One can associate a coefficient to each line in a code
word w, being its value at the corresponding position. A word c = (c1 , . . . , cN q2 ) ∈
2
KN q is in C if and only if w · H T = ~0, hence (since Hji = 1 ⇔ `i 3 pj ) if and only
if X X
ci Hji = ci = 0
`i `i 3pj
as an element of K for every point pj . Alternatively formulated: a word is a code
word of C if and only if the sum of the coefficients of the lines through every point
equals 0 over K.
Definition 2.7. Let ri , rj ∈ K with i < j and let π be a projective two-dimensional
plane through ri , rj different from Π0 . The plane word through ri and rj in π is the
code word with
• +1 in the positions corresponding to the lines of π through ri ,
• −1 in the positions corresponding to the lines of π through rj ,
• 0 in the positions corresponding to all other lines.
Notation 2.8. We denote by C 0 the code generated by all plane words. Given a
plane word w through pi and pj , define T (w) to be the plane π in Definition 2.7
and L(w) to be the line pi pj in Definition 2.7.
Definition 2.9. Let L be a line in Π0 containing at least two points of K. Let
π be a projective (two-dimensional) plane through L different from Π0 , and let
p0 , . . . , pk−1 be the points of L ∩ K. We define a generalized plane word in π to be
a code word with
• a0 in the positions corresponding to the lines of π through p0 ,
• a1 in the positions corresponding to the lines of π through p1 ,
• ...
• ak−2 in the positions corresponding to the lines of π through pk−2 ,
• −a0 − a1 − . . . − ak−2 in the positions corresponding to the lines of π through
pk−1 ,
• 0 in the positions corresponding to all other lines,
for some scalars a0 , a1 , a2 , . . . , ak−2 ∈ K.
Remark 2.10. Note that if a line would contain two points of K, then it is not a
line of T2∗ (K), because is contained in the plane at infinty and hence not an affine
line.
Remark 2.11. Note that a sum of plane words in a fixed plane π is a generalized
plane word in π, and a sum of generalized plane words in π is still a generalized plane
word in π. Moreover, if π ∩ K = {p0 , . . . , pk−1 }, then the set of generalized plane
words in π is spanned by the plane words through (p0 , p1 ), (p0 , p2 ), . . . , (p0 , pk−1 ).
Hence, C 0 is also the code spanned by all generalized plane words and we need at
most one generalized plane word per plane to obtain any word of C 0 .
Remark 2.12. It is known that T2∗ (K) is a partial geometry if and only if K is a
(maximal) {qn − q + n; n}-arc for some n ≥ 1 (see [32]) and T2∗ (K) is a semipartial
geometry if and only if K is a Baer subplane or a unital (see [3]). A good general
reference on T2∗ (K) is [2].
408 PETER VANDENDRIESSCHE

3. Dimension of C 0 .
Notation 3.1. Denote by L the set of projective lines at infinity that contain at
least one point of K. Denote by LN the size of L, i.e.
L = {`1 , `2 , . . . , `LN }
and by LS the summed size of L, i.e.
X
LS = |` ∩ K|.
`∈L
Let K be an arbitrary field with char K 6= p.
Lemma 3.2. Let ` ∈ L be a line in the plane at infinity, containing exactly k points
of K. Then there are exactly q(k − 1) linearly independent plane words among all
plane words w with L(w) = `.
Proof. Number the k points p0 , p1 , . . . , pk−1 , then the pairs
(p0 , p1 ), (p0 , p2 ), . . . , (p0 , pk−1 )
each yield q different plane words, and these are linearly independent. Now for all
other pairs, the plane words through (pi , pj ), with i < j, can be written as the
difference of the corresponding plane words through (p0 , pi ) and (p0 , pj ). Hence the
result follows.
Pn
Lemma 3.3. Fix an arbitrary point p0 ∈ K. Let1 i=1 λi vi = ~0 be a linear com-
bination of generalized plane words yielding the zero word, at most one generalized
plane word per plane. If L(vi ) = L(vj ) and this line contains p0 , then λi = λj .
Proof. The q affine lines through p0 in T (vi ) each get a contribution of λi from vi .
The q affine lines through p0 in T (vj ) each get a contribution of λj from vj . All
other generalized plane words vm contribute equally much to the sum of both sets
of q lines (namely λm if T (vm ) contains p0 and 0 otherwise). Denote by R the total
summed contribution to both sets of q lines.
Since the total sum of all contributions is 0 for every line (since we assumed that
this linear combination yields the zero word) we have qλi + R = 0 = qλj + R, hence
qλi = −R = qλj . Since q 6= 0 as an element of K, it follows that λi = λj .
Corollary 3.4. We did not assume λi 6= 0, hence if one of the generalized plane
words vi appears in the linear combination with a nonzero λi , then all generalized
plane words vj through the same line at infinity should appear with λj = λi . Hence
if we start from an empty code (considered as vector space), and we consider one by
one all lines ` at infinity and we add the q generalized plane words through ` to this
vector space, then each line increases the dimension with at least (|` ∩ K| − 1)(q − 1),
since the codimension can be at most |` ∩ K| − 1.
Theorem 3.5. The dimension of C 0 is (N − 1) + (q − 1)(LS − LN ).
Proof. Take any point p0 and look at the lines L0 , . . . , Lq through p0 in the plane
at infinity. The plane words through (p0 , p), for p ∈ K \ {p0 }, form a basis for the
linear combinations of plane words on the lines L0 , . . . , Lq . Starting from an empty
vector space V as described in Corollary 3.4, these plane words contribute (N − 1)q
1 By convention, we choose all generalized plane words containing lines through p in their
0
support, to have coefficient +1 on the lines through p0 (this can always be accomplished by
scaling the λi ).
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 409

to the dimension, because of Lemma 3.2. Now adding every other line L ∈ L at
infinity, not through p0 , contributes
• at least (|L ∩ K| − 1)(q − 1) to dim V , since Lemma 3.3 states that in any
linear combination of generalized plane words yielding the zero word, all planes
through L appear with the same coefficient, hence we miss at most |L ∩ K| − 1
degrees of freedom,
• and at most (|L ∩ K| − 1)(q − 1) to dim V , since one can write the zero word
as a linear combination of plane words through (p0 , p), (p0 , p0 ), (p, p0 ) for any
two points p, p0 ∈ L ∩ K (note that all plane words through lines through p0
are already in the code at this point).
Therefore, the dimension is exactly
X
(N − 1)q + (|L ∩ K| − 1)(q − 1).
p0 6∈L∈L
P
Note that p0 ∈L∈L (|L ∩ K| − 1) = N − 1 since both represent all points of K except
for p0 . Hence
dim(C 0 ) = (N − 1)q + p0 6∈L∈L (|L ∩ K| − 1)(q − 1)
P
P
= (N − 1) + (q − 1) P L∈L (|L ∩ K| −
1)

= (N − 1) + (q − 1) L∈L |L ∩ K| − |L|
= (N − 1) + (q − 1)(LS − LN ).

Remark 3.6. We now know that C 0 is a linear [N q 2 , N −1+(q−1)(LS −LN )]-code.

There is no general expression for LS − LN in terms of q and N . However, there is
an easy algorithm to compute LS − LN for an arbitrary set K:
Let K be an arbitrary subset of PG(2, q). Fix any point p0 (inside K or outside
K). Call a line through p0
• a secant if it contains two or more points of K \ {p0 },
• a tangent if it contains exactly one point of K \ {p0 }, and
• a passant if it contains no points of K \ {p0 }.
When adding/removing a point p0 ,
• LS increases/decreases by q + 1, while
• LN increases/decreases by the number of passants through p0 .
Hence, LS − LN increases/decreases by the number of non-passant lines through
p0 .
Some examples:
• If K is a k-arc, then adding the ith point increases LS − LN by i − 1. Hence,
k
X k(k − 1)
LS − LN = (i − 1) = .
i=1
2

• If K is the pointwise union of two intersecting lines, then adding the points
on the first line increases LS − LN by 1 each time (except for the first point),
while adding the other q points increases it by q + 1 each time. Hence, in this
case
LS − LN = q + q(q + 1) = q 2 + 2q.
410 PETER VANDENDRIESSCHE

Remark 3.7. Since one has in general that LS = N (q + 1), the dimension formula
can be rewritten as q 2 N −q 3 +(q −1)(q 2 +q +1−LN ). Since the parity check matrix
is a q 3 × q 2 N matrix, this means that the rank deficiency of the parity check matrix
is q − 1 times the number of lines at infinity, skew to K. It may be interesting to
find out if there exists a more direct way to obtain this formula.

4. Dimension of C. We will compute the dimension of T2∗ (K) as follows. First

we will compute dim C in the case K = PG(2, q), and find that it equals dim C 0 in
that case, hence C = C 0 . Then we will present a technique to keep this property
valid while removing arbitrary points from K. Every subset K of PG(2, q) can be
obtained by removing a finite number of points from PG(2, q), so the conclusion
will follow. As in the previous section, we will always assume that q 6= 0 over K,
i.e. char K 6= p.

4.1. The case K = PG(2, q). In this case, we simply have T2∗ (K) = AG(3, q).
Remark that AG(3, q) is a 2-(q 3 , q, 1) block design, using lines as blocks. Denote by
An the incidence matrix of AG(n, q), with n ≥ 2, where points correspond to rows
and lines correspond to columns.
It is a classical result in design theory (see [1] for a proof and more general
background on designs) that An ATn = (q n−1 + q n−2 + . . . + q 2 + q)I + J, where J
denotes the q n × q n matrix with all entries equal to 1. This has determinant
n
−1
(q n + q n−1 + . . . + q 2 + q)(q n−1 + q n−2 + . . . + q 2 + q)q .
In fact, An ATn has q n − 1 eigenvalues equal to q n−1 + q n−2 + . . . + q 2 + q and one
eigenvalue equal to q n + q n−1 + . . . + q 2 + q. The determinant is non-zero when
q n−1 + q n−2 + . . . + q 2 + q 6= 0 and q n + q n−1 + . . . + q 2 + q 6= 0. For most choices of
the characteristic, this already shows that An has full rank, however in some cases
further study is required:
• The case q = 0 has been excluded by our assumptions. In fact, An does not
have full rank in this case. We will further assume q 6= 0.
• The case q n−1 + q n−2 + . . . + q + 1 = 0 is easily solved. Note that this implies
q 6= 0 and q n−2 + q n−3 + . . . + q + 1 6= 0, hence in this case only one of
the eigenvalues of An ATn equals zero over K, and we see that (1, . . . , 1)T is
an eigenvalue corresponding to this eigenvector. Hence (1, . . . , 1)T is (up to
scalar multiples) the only eigenvector corresponding to this eigenvalue, but
one can verify that
(1, . . . , 1)ATn = q n (1, . . . , 1) 6= ~0
since q 6= 0. Hence, this is not a code word, and hence there is no v 6= ~0 such
that An v = ~0, i.e. An has full rank in this case.
• The case q n−2 + q n−3 + . . . + q + 1 = 0 is more difficult. Purely combinatorial
approaches seem to fail, but a geometric trick works. We will now develop this
technique to find the rank of An over K for the case q n−2 +q n−3 +. . .+q+1 = 0.
Lemma 4.1. Let k ≥ 2. If the incidence matrix of AG(k + 1, q) is rank deficient
over K, then the incidence matrix of AG(k, q) is also rank deficient over K.

Proof. Assume that the incidence matrix of AG(k + 1, q) is rank deficient. Since
k ≥ 2, there are more lines than points. Hence, rank deficiency means that there
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 411

exists a linear combination of points whose corresponding line-incidence vectors

yield the zero vector (zero for each line):
X
ci pi = ~0
pi ∈AG(k+1,q)

with not all ci = 0 over K.

Now consider any hyperplane Π ∼ = AG(k, q) in our AG(k + 1, q) which contains
at least one point with non-zero coefficient in the linear combination. For all lines
contained in Π, the linear combination
P of the point-line incidence vectors has to be
zero as well. I.e. in Π we also have pi ∈Π ci pi = ~0. Since Π contains at least one
point with non-zero coefficient in the linear combination, this linear combination is
nontrivial. Hence, the incidence matrix of AG(k, q) is rank deficient as well.

Theorem 4.2. The incidence matrix of AG(n, q) (with n ≥ 2), q = ph with p

prime, has full rank over all fields with char K 6= p.

Proof. We will prove this by induction on n. For n = 2 this is clear from the remarks
in the beginning of this section, since the problematic case above (q n−2 +. . .+q+1 =
0) reduces to 1 = 0, hence can be excluded immediately. For each n ≥ 2 it follows
by contraposition of Lemma 4.1 that if the statement is true for n, then it is also
true for n + 1. Hence, by induction, it is true for all n ≥ 2.

We have proven thatf An has full rank for all n ≥ 2 when char K 6= p. Explicitly,
for the general setting of Tn∗ (K) with K = PG(n, q) (which yields exactly AG(n +
1, q)) the geometry has q n (q n + q n−1 + . . . + q + 1) lines and q n+1 points, hence
Theorem 4.2 yields
dim C = q n (q n + q n−1 + . . . + q 3 + q 2 + 1).
Now compare this to the result of Section 3. In Section 3 we worked with n = 2,
hence N = q 2 + q + 1 and T2∗ (K) = AG(3, q). This gives us
dim(C 0 ) = (N − 1) + (q − 1)(LS − LN )
= q 2 + q + (q − 1)((q 2 + q + 1)(q + 1) − (q 2 + q + 1))
= q 2 (q 2 + 1)
= dim C.
and hence dim C = dim C 0 . Hence, the code associated with AG(3, q) is spanned
completely by its plane words. However, for n > 3, Theorem 3.5 is no longer valid.
We finish with a conjecture for the higher dimensions:
Conjecture 4.3. If K = PG(n, q) with n > 2 and char K 6= p then the code
associated with Tn∗ (K) = AG(n + 1, q) over K also has dim C 0 = dim C.
4.2. The general case. The main idea here is the following: if we remove a point
from K, we claim that the property that the code is spanned by its plane words
0
remains valid. To distinguish between different point sets, we denote by CK , CK
∗
respectively the full code and the plane words code associated with Tn (K). Similarly,
we denote by LN,K , LS,K the respective values of LN and LS for the set K.
Theorem 4.4. Let K be a nonempty subset of PG(2, q) and let char K 6= p. We
0 0
have dim CK = dim CK (and hence CK = CK ).
412 PETER VANDENDRIESSCHE

Proof. According to Theorem 3.5, if T ⊆ PG(2, q), then removing a point p0 from
T decreases dim C 0 by
dim CT0 − dim CT0 \{p0 } = 1 + (q − 1)((LS,T − LN,T ) − (LS,T \{p0 } − LN,T \{p0 } ))
and hence decreases dim C by at least this amount.
Fix one point p0 ∈ K. Now remove all other points of PG(2, q) one by one, first
the points outside K then the points inside K, except for p0 . Denote by Ti the set
in the intermediary step with i points:
T| PG(2,q)| = PG(2, q), TN = K, T1 = {p0 },
and define Q = | PG(2, q)| − 1. Here, | PG(2, q)| denotes the number of points in
PG(2, q), which is q 2 + q + 1. Note that in any code word, every affine point lies
on either 0 or at least 2 lines of the support of that code word, hence C{p0 } = {~0}.
Note that C{p0 } is simply CK with K = {p0 }. Hence, we have
dim CPG(2,q) = dim CPG(2,q) − dim C{p0 }
XQ


= dim CTi+1 − dim CTi
i=1
Q
X
≥ dim CT0 2 − dim CT0 1 +


dim CTi+1 − dim CTi
i=2
≥ ···
XQ  
≥ dim CT0 i+1 − dim CT0 i
i=1
0 0
= dim CPG(2,q) − dim C{p 0}
0
= dim CPG(2,q) .
0
It was proven in the previous subsection that dim CPG(2,q) = dim CPG(2,q) , hence
we must have equality in each inequality “≥”. This means that
dim CTi+1 − dim CTi = 1 + (q − 1)((LS,Ti+1 − LN,Ti+1 ) − (LS,Ti − LN,Ti ))
for each i. A simple induction gives dim CTi = dim CT0 i for all i, in particular for
0
i = N we have dim CK = dim CK .

Hence for T2∗ (K) in general it is now proven that dim C = dim C 0 and the code
C is generated completely by its plane words.
Remark 4.5. If Conjecture 4.3 is true, then Theorem 4.4 can be extended to
arbitrary subsets of PG(n, q): then we have dim C = dim C 0 for the code associated
with Tn∗ (K) with arbitrary K ⊆ PG(n, q).
Remark 4.6. As an immediate consequence, we get that in the binary code asso-
ciated with T2∗ (K) for q odd, all code words have even weight.

5. The minimum distance of C. Now that the dimension and structure of C

are known, we can attack another one of its key properties: the minimum distance.
In some sporadic cases the minimum distance is known [13, 23], however in most
cases one only has lower bounds from the tree bound [29], the bit-oriented bound
and the parity-oriented bound [30].
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 413

Theorem 5.1. For any finite field K, all code words c with w(c) < 2q must be
contained in a single plane. If w(c) = 2q, then either c is a plane word, supp(c) is
the set of lines of a hyperbolic quadric with two intersecting lines contained in K,
or char K = p = 2.
Proof. This follows from [23], Proposition 4.
Now, we will use the structure of C to sharpen this result and classify the mini-
mum weight code words.
Theorem 5.2. If char K 6= p, there are no code words c ∈ C with w(c) < 2q. If
w(c) = 2q, then either c is a plane word or supp(c) is the set of lines of a hyperbolic
quadric with two intersecting lines contained in K.
Proof. The second part follows immediately from Theorem 5.1. For the first part,
assume that there exists a code word with w(c) < 2q and let U be its support.
By Theorem 5.1, the support of this code word is contained in a plane T . Define
m = |K ∩ T | and write K ∩ T = {p1 , . . . , pm }. Since supp(c) ⊂ T , we have that c is
a generalized plane word in T . Hence, either each of the q lines through pi appear
in U , or none of them do. Since one needs either 0 or at least 2 lines through a
point, it follows that w(c) ≥ 2q, a contradiction.
Remark 5.3. Note that it may actually happen that there are minimum weight
code words other than plane words – Theorem 5.1 classfies them. However, it
follows from Theorem 4.4 that these other minimum weight code words are also a
linear combination of plane words. Hence, even in this case, the statements ‘C is
generated by its (generalized) plane words’ and ‘C is generated by its code words
of minimum weight’ are equivalent.
So far we have proven that C is a linear [N q 2 , N − 1 + (q − 1)(LS − LN ), 2q]-code
and it is completely generated by its minimum weight code words.
Now, let us see what happens for Tn∗ (K) with n > 2, assuming Conjecture 4.3 is
true.
Theorem 5.4. If Conjecture 4.3 is true, then d(C) = 2q is true for Tn∗ (K) with
arbitrary n ≥ 2 and arbitrary K ⊆ PG(n, q) (still assuming char K 6= p).
Proof. Assume there exists a non-zero code word c with w(c) < 2q and let U be its
support. Let T be any plane and define t = |U ∩T |. Since the sum of the coefficients
has to be 0 in each point, any point on a line of the support lies on at least one
other line of the support. Hence we have
t(q − t + 1) < 2q − t,
meaning t > q or t < 2. Since t is an integer, this means t ≥ q + 1 or t ≤ 1.
Now, if there are at least two planes for which t ≥ q + 1, then w(c) ≥ 2q + 1,
contradiction. Hence there is at most one such plane and U is completely contained
in this plane. The rest of the proof can be copied from Theorem 5.2.
The following theorem summarizes the results obtained so far:
Theorem 5.5. The code associated with T2∗ (K) over any field K, with char K 6= p,
is a linear [N q 2 , N − 1 + (q − 1)(LS − LN ), 2q]-code and it is completely generated by
its minimum weight code words. If Conjecture 4.3 is true, then for char K 6= p the
code associated with Tn∗ (K) also has d(C) = 2q and it is also generated completely
by its minimum weight code words.
414 PETER VANDENDRIESSCHE

6. Some practical considerations and further work. A commonly used ap-

proach when constructing good LDPC codes is the maximization of the girth of its
Tanner graph [8],[20],[35]. It is known that high girth decreases the dependence
between passing messages in the belief-propagation sum-product algorithm. The
Tanner graph of T2∗ (K) always has a girth of at least 6. If K is an arc, then the
girth is 8, as T2∗ (K) contains no triangles.
Liu and Pados [16] mention an opposing objective: the minimization of the
diameter of the Tanner graph, which brings them to generalized polygons. If K is
not contained within a line, the diameter of T2∗ (K) is at most 6. If K contains no
tangents at infinity, the diameter is as low as 4. Examples of such choices of K
include K a hyperoval and K a double blocking set.
There is a unique choice for K that combines both of the preceding objectives:
the case where K is a hyperoval, which only exists for q even. Then T2∗ (K) is a
generalized quadrangle with girth 8 and diameter 4. From the above points of
view, this is probably the most appealing case, however, the restriction char K 6= p
excludes the most important field for practical applications: F2 .

Hyperoval q = ph dimF2 C dimR C = dimR C 0 dimF2 C 0

Regular hyperoval q = 2 9 9 8
Regular hyperoval q = 4 50 50 37
Regular hyperoval q = 8 341 324 194
Regular hyperoval q = 16 2670 2312 1105
Lunelli-Sce hyperoval q = 16 2550 2312 1107
Regular hyperoval q = 32 22248 17424 6578
Translation hyperoval q = 32 21258 17424 6608
Cherowitzo hyperoval q = 32 20358 17424 6613
Payne hyperoval q = 32 20388 17424 6613
Segre hyperoval q = 32 20553 17424 6613
O’Keefe-Penttila hyperoval q = 32 20343 17424 6613
Regular hyperoval q = 64 188665 135200 39937
Adelaide hyperoval q = 64 169772 135200 40312
Subiaco I hyperoval q = 64 169254 135200 40312
Subiaco II hyperoval q = 64 169388 135200 40309
Table 1. Simulation results for the binary codes associated with
T2∗ (K) where K is a hyperoval.

For the binary code associated with T2∗ (K) when K is a hyperoval, the results in
this paper are no longer valid. Lemma 3.3 no longer guarantees the lower bound on
dim C 0 and since the 2-rank is at most the real rank, the dimension of C could be
larger than what we have derived in this paper. In Table 6, we have calculated the
dimension (and hence the code rate, which is dimension over length) of the F2 -code
and R-codes associated with T2∗ (K) with K a hyperoval, by computer simulations, for
multiple types of hyperovals in PG(2, 2h ). For h ≤ 5, these are the only hyperovals
(for proofs of these facts, see [6, 21, 22, 26]). For h = 6, it is commomly believed
that these are the only hyperovals, but a proof has not been found yet. For h > 6, a
classification is not even conjectured and the computations also become unfeasible.
We see that the results indeed deviate from the numbers in Theorem 5.5. In
this case we get even better parameters: the dimension increases while no other
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 415

visible parameters change. However, we lose the structural property that the code
is spanned by its code words of minimum weight. For other choices of K when q is
even, even the minimum distance may decrease. This has been investigated more
closely in [23]. In general, only a minimum weight of q + 1 can be guaranteed.
If one wants to maintain the structure property and still use a binary code, then
q must be odd. Some examples that are ‘near-optimal’ choices for K in these cases
include:
• K is a (q + 1)-arc (and hence a conic by [25]). Compared to the case where K
is a hyperoval, the dimension and rate are slightly smaller, the code is a bit
shorter in length and the girth remains 8. The diameter is 6 now, but there
are only very few pairs of vertices where this distance is actually reached. It
would be interesting to perform practical simulations to find out if this case
still guarantees a fast average decoding speed.
• K has no tangents. For example, K is a dual double blocking set. Here the
Tanner graph of T2∗ (K) has girth 6 and the diameter is still 4, but the length
of the code is necessarily longer compared to the case where K is a hyperoval,
without an increase in minimum distance.
To finish, we take a look ahead on possible further work on this subject. Each of
the following could be a significant contribution to the understanding of this class
of codes.
• For Tn∗ (K), with n > 2, little is known. If Conjecture 4.3 is true, it could
be interesting to find a generalization of the formula in Theorem 3.5 and to
analyze its geometric interpretation. Regarding the optimal choice of K, there
is no n-dimensional equivalent of the hyperoval, so it would be interesting to
know which choices for K yield interesting geometries (if any).
• Practical simulations on encoding/decoding speed or other performance mea-
surements of T2∗ (K) for some good choices for K would be very helpful. Espe-
cially the case K is a hyperoval, a conic or some of the examples studied in [23]
or in this article, would be interesting. It would also be interesting to know if
the case char K = p generally performs better or worse than char K 6= p.
• If char K = p, few results in this paper remain valid. The only trick that works
completely when char K = p is that if the incidence matrix of AG(n + 1, q)
is rank deficient, then the incidence matrix of AG(n, q) is rank deficient. A
suited structural property could potentially be preserved in a way similar to
Section 4.2. This suggests that it may be a good help to first find out the
structure of the base case AG(2, q), especially which code words remain valid
if we remove certain classes of parallel lines. Another indication that this may
be an interesting topic is the minimum distance. From Theorem 5.1, code
words c with w(c) < 2q are necessarily contained within a plane. If one knows
the structure of the LDPC code associated with AG(2, q), one is likely to find
a general result on the minimum weight. Until now, the only known lower
bounds are the bounds in [23].
• If char K = p, it would be useful to find a structure or dimension result even
just for special cases. Even for T2∗ (K) with K a hyperoval this seems a lot
harder. From the simulation results in Table 6, one can see that the dimensions
are different between different types of hyperovals. This may be related to the
approach in [23]: dim C may depend on how many points of a conic are con-
tained in K, while the difference in dim C 0 between regular/translation/other
416 PETER VANDENDRIESSCHE

hyperovals may be related to Remark 7 in [23], since other hyperovals are not
known to have such special points.

Acknowledgements. The author wants to thank Leo Storme for his detailed
proofreading and useful suggestions. The author also wants to thank the anonymous
referees for their useful comments and suggestions.
REFERENCES

[1] P. J. Cameron and J. H. Van Lint, “Designs, Graphs, Codes and their Links,” Cambridge
University Press, 1991.
[2] F. De Clerck and H. Van Maldeghem, On linear representations of (α, β)-geometries, Euro-
pean J. Combin., 15 (1994), 3–11.
[3] I. Debroey and J. A. Thas, Semi partial geometries in AG(2, q) and AG(3, q), Simon Stevin,
51 (1978), 195–209.
[4] M. P. C. Fossorier, Quasicyclic low-density parity check codes from circulant permutation
matrices, IEEE Trans. Inform. Theory, 50 (2004), 1788–1793.
[5] R. G. Gallager RG, Low density parity check codes, IRE Trans. Inform. Theory, 8 (1962),
21–28.
[6] M. Hall, Ovals in the Desarguesian plane of order 16, Ann. Mat. Pura Appl., 102 (1975),
159–176.
[7] J. W. P. Hirschfeld, “Projective Geometries over Finite Fields,” 2nd edition, Oxford University
Press, 1998.
[8] X.-Y. Hu, E. Eleftheriou and D. M. Arnold, Regular and irregular progressive edge-growth
Tanner graphs, IEEE Trans. Inform. Theory, 51 (2005), 386–398.
[9] S. J. Johnson and S. R. Weller, Construction of low-density parity-check codes from combina-
torial designs, in “Proceedings of the IEEE Information Theory Workshop,” Cairns, (2001),
90–92.
[10] S. J. Johnson and S. R. Weller, Construction of low-density parity-check codes from Kirkman
triple systems, in “Proceedings of the IEEE Globecom Conference,” San Antonio, 2001.
[11] S. J. Johnson and S. R. Weller, Codes for iterative decoding from partial geometries, in
“Proceedings of the IEEE International Symposium on Information Theory,” Switzerland,
2002.
[12] S. J. Johnson and S. R. Weller, Regular low-density parity-check codes from oval designs,
Eur. Trans. Telecommun., 14 (2003), 399–409.
[13] J. L. Kim, K. Mellinger and L. Storme, Small weight codewords in LDPC codes defined by
(dual) classical generalized quadrangles, Des. Codes Cryptogr., 42 (2007), 73–92.
[14] J. L. Kim, U. Peled, I. Perepelitsa, V. Pless and S. Friedland, Explicit construction of families
of LDPC codes with no 4-cycles, IEEE Trans. Inform. Theory, 50 (2004), 2378–2388.
[15] Y. Kou, Y, S. Lin and M. P. C. Fossorier, Low-density parity-check codes based on finite
geometries: a rediscovery and new results, IEEE Trans. Inform. Theory, 47 (2001), 2711–
2736.
[16] Z. Liu and D. A. Pados, LDPC codes from generalized polygons, IEEE Trans. Inform. Theory,
51, 3890–3898.
[17] D. J. C. MacKay, Good error correcting codes based on very sparse matrices, IEEE Trans.
Inform. Theory, 45 (1999), 399–431.
[18] D. J. C. MacKay and M. C. Davey, Evaluation of Gallager codes for short block length and
high rate applications, in “Codes, Systems and Graphical Models” (eds. B. Marcus and J.
Rosenthal), Springer-Verlag, New York, (2000), 113–130.
[19] D. J. C. MacKay and R. M. Neal, Near Shannon limit performance of low density parity
check codes, Electron. Lett., 32 (1996), 1645–1646.
[20] G. A. Margulis, Explicit constructions of graphs without short cycles and low density codes,
Combinatorica, 2 (1982), 71–78.
[21] C. M. O’Keefe and T. Penttila, Hyperovals in PG(2, 16), European J. Combin., 12 (1991),
51–59.
[22] T. Penttila and G. F. Royle, Classification of hyperovals in PG(2, 32), J. Geom., 50 (1994),
151–158.
[23] V. Pepe, L. Storme and G. Van de Voorde, Small weight codewords in the LDPC codes arising
from linear representations of geometries, J. Combin. Des., 17 (2009), 1–24.
 LDPC CODES ASSOCIATED WITH LINEAR REPRESENTATIONS OF GEOMETRIES 417

[24] J. Rosenthal and P. O. Vontobel, Construction of LDPC codes using Ramanujan graphs and
ideas from Margulis, in “Proceedings of the 38th Allerton Conference on Communications,
Control and Computing” (eds. P.G. Voulgaris and R. Srikant), Monticello, (2000), 248–257.
[25] B. Segre, Ovals in a finite projective plane, Canad. J. Math., 7 (1955), 414–416.
[26] B. Segre, Sui k-archi nei piani finiti di caratteristica due (in Italian), Rev. Math. Pures Appl.,
2 (1957), 289–300.
[27] P. Sin and Q. Xiang, On the dimension of certain LDPC codes based on q-regular bipartite
graphs, IEEE Trans. Inform. Theory, 52 (2006), 3735–3737.
[28] M. Sipser and D. A. Spielman, Expander codes, IEEE Trans. Inform. Theory, 42 (1996),
1710–1722.
[29] R. M. Tanner, A recursive approach to low complexity codes, IEEE Trans. Inform. Theory,
27 (1981), 533–547.
[30] R. M. Tanner, Minimum distance bounds by graph analysis, IEEE Trans. Inform. Theory, 47
(2001), 808–821.
[31] R. M. Tanner, D. Sridhara, A. Sridharan, T. E. Fuja and J. D. Costello Jr, LDPC block codes
and convolutional codes based on circulant matrices, IEEE Trans. Inform. Theory, 50 (2004),
2966–2984.
[32] J. A. Thas, Partial geometries in finite affine spaces, Math. Z., 158 (1978), 1–13.
[33] P. Vandendriessche, Some low-density parity-check codes derived from finite geometries, Des.
Codes Cryptogr., 54 (2010), 287–297.
[34] P. O. Vontobel and R. M. Tanner, Construction of codes based on finite generalized quadran-
gles for iterative decoding, in “Proceedings of 2001 IEEE International Symposium Informa-
tion Theory,” Washington, (2001), 233–233.
[35] H. Xiao and A. H. Banihashemi, Improved progressive-edge-growth (PEG) construction of
irregular LDPC codes, IEEE Commun. Lett., 8, 715–717.

Received December 2009; revised May 2010.

E-mail address: [email protected]