Des. Codes Cryptogr.

(2008) 46:45–56
DOI 10.1007/s10623-007-9133-y

On perfect p-ary codes of length p + 1

Olof Heden

Received: 8 February 2007 / Revised: 1 September 2007 / Accepted: 5 September 2007 /

Published online: 13 November 2007
© Springer Science+Business Media, LLC 2007

Abstract Let p be a prime number and assume p ≥ 5. We will use a result of L. Redéi
to prove, that every perfect 1-error correcting code C of length p + 1 over an alphabet of
cardinality p, such that C has a rank equal to p and a kernel of dimension p − 2, will be
equivalent to some Hamming code H . Further, C can be obtained from H , by the permutation
of the symbols, in just one coordinate position.

Keywords Perfect codes · Redei Theorem

AMS Classification 94B25

1 Introduction

I was introduced to perfect codes in 1965 by my father (G. Heden, private communication),
who showed me the following example:

1 1 1 x x x 2 2 2
1 x 2 1 x 2 1 x 2
1 x 2 x 2 1 2 1 x
1 x 2 2 1 x x 2 1

The columns in the above example have the property, that any possible column in the three
symbols 1, x, 2 differs, in at most one position, from a unique column in the set above. And
thus, it will be an example of a perfect 1-error correcting ternary code of length four.
A general definition of a perfect code is the following.

Communicated by V.A. Zinoviev.

O. Heden (B)
Department of Mathematics, KTH, Stockholm 100 44, Sweden
e-mail: [email protected]

123
46 O. Heden

Let A be a finite set of q symbols. A perfect e-error correcting q-ary code of length n is
a subset C of words of the direct product An
C ⊆ A × A × · · · × A = An ,
such that any word x̄ of length n, differs in at most e positions from a unique code word c̄ of
C. A code C is a perfect code if C is a perfect e-error correcting q-ary code of length n for
some parameters e, q and n.
In the example above, A = {1, x, 2}, and so q = 3, and further n = 4 and e = 1. It is also
rather easy to check, that if you allow permutations of the coordinate set and permutations
of the set of symbols in some of the positions, then any perfect code with these parameters
will be unique.
More precisely and more general, we define C and C  to be equivalent, if we can obtain

C from C, by a sequence of the following two operations:
(i) a permutation π of the set of n coordinate positions.
(ii) permutations τi of the set of symbols in position i, for i = 1, 2, . . . , n.
The first perfect codes were found by Hamming [5]. He found a rather big class of per-
fect 1-error correcting binary codes. In 1948 Golay [4] found one perfect 2-error correcting
3-ary code of length 11 and one perfect 3-error correcting binary code of length 23. In 1973,
Tietäväinen [18] and independently Leontiev and Zinoviev [10], showed that for no other
parameters, in the prime power case, there will be perfect e-error correcting codes with e ≥ 2.
Let us also mention, that the case of perfect q-ary codes, where q is not a power of a prime
number, was considered by Best, who showed that there are no unknown perfect e-error
correcting codes over arbitrary alphabets for e distinct from 1, 2, 6, 8, see [1]. The case when
e = 1 and q not a prime power is known to be extremely difficult.
Perfect 1-error correcting binary codes have been extensively studied. There are many
constructions leading to distinct equivalence classes of such codes. Important contributors
are here Vasil’ev, Zinoviev, Solov’eva, Phelps, Hergert and Avgustinovich, see for example
the book On Perfect Codes and Related Topics [17] by Solov’eva for references, as well as
for elementary definitions and facts on perfect codes.
Perfect 1-error correcting q-ary codes, where q  = 2 is a prime number or a power of a
prime number, have been less studied. Constructions are given e.g. by Schönheim [16], Lind-
ström [11], Phelps [12] and Etzion [3]. Further, there are some recent classification results
by Phelps, Villanueva and Rifà, see [13] and [14].
It can by rather easy combinatorial arguments be shown, that the shortest possibly length,
for which there exists a non trivial perfect 1-error correcting q-ary code, is q + 1. This par-
ticular case is not at all covered by the results of Phelps, Rifà and Villanueva [14,15]. The
purpose of this study, is to give a small result in this case, showing that a particular class of
these codes are equivalent to the linear perfect codes, that we now describe.
We will need a description of a linear perfect 1-error correcting q-ary code of length q +1.
Let Fq be a finite field with the q elements
Fq = {0, α1 , α2 , . . . , αq−1 }.

Proposition 1 The null space of the matrix


0 α1 α1 α1 . . . α1
,
α1 0 α1 α2 . . . αq−1
will be a linear perfect 1-error correcting q-ary code H of length q + 1.

123
On perfect p-ary codes of length p + 1 47

The code H , so constructed in the above proposition, will here be called a Hamming code,
although we are aware of the fact that by a Hamming code in general, is meant any linear
perfect code.
To describe the results of this paper, we need the concepts of rank and kernel.
We consider Fqn as a n-dimensional vector space over the finite field Fq . The linear span
of the code words of a code C, will be a subspace C of Fqn . The dimension of this subspace
will be the rank of C:
rank(C) = dim(C).
A period of a q-ary code C of length n, is a word p̄ ∈ Fqn satisfying
p̄ + c̄ ∈ C for all words c̄ ∈ C.
If q = p a prime number, then the set of all periods of a perfect code will be a subspace of
Fpn . This subspace will be the kernel of C, denoted by ker(C). In case q = p t , with t > 1,
the definition will necessarily be different, see e.g., [14]. If the all zero word 0̄ belongs to C,
as we will always assume below, then ker(C) is a subset of C.
We now give an example of a perfect 1-error correcting 5-ary code of length 6, which has
rank 5 a kernel of dimension 3 and which is equivalent to a Hamming code.

Example Let F5 = Z5 and consider the null space H of the matrix


011111
.
101234
In the first position we then perform the permutation τ1 (x) = x for x = 0, 1, 2, τ1 (4) = 3
and τ1 (3) = 4. The perfect code H is then transformed into a perfect code C, equivalent to
H . The words (1, 1, 4, 0, 0, 0) and (3, 4, 1, 0, 0, 0) belong both to C, but
(1, 1, 4, 0, 0, 0) + (3, 4, 1, 0, 0, 0) = (4, 0, 0, 0, 0, 0)  ∈ C,
as C contains the all zero word and the minimum distance of C equals three. Hence C is not
linear. Let K be the subspace of H consisting of those words (c1 , c2 , . . . , c6 ) of H such that
c1 = 0. It is easy to see that K is a subset of the kernel of C and that dim(K) = 3. It follows
that K actually is the kernel of C.
Trivially C is a subset of the null space of the matrix
 
011111 .

Hence, dim(C) ≤ 5, and as C is non linear and contains 54 elements, the only possibility
will be that the rank of C equals 5.

It must be remarked that in the binary case, equivalent perfect codes have kernels of the
same dimension and they have also the same rank, if both codes contain the all zero word.
The purpose of this paper is to show the following theorem:

Theorem 1 Let p be a prime number and assume p ≥ 5. Any perfect 1-error correcting
p-ary code C of length p + 1, rank p and with a kernel of dimension p − 2 is equivalent to
a linear perfect code.

It is easy to see, and must be remarked, that every perfect 1-error correcting ternary code
of length four and containing the all zero word will be linear and equivalent to the code in our
first example. Let us also remark that in case q = p t is a prime power, with t ≥ 2 and q ≥ 9,
perfect 1-error correcting codes of length q + 1, that are not equivalent to any linear codes,

123
48 O. Heden

but group codes, were found by Lindström in 1969, see [11]. Such codes, including the case
q = 8, were also constructed by Phelps et al. [14]. Indeed, it would be very interesting to
find an example of a perfect code of length q + 1 that is not equivalent to any group code.
We have not yet found any direct proof of Theorem 1. In our proof we will use the tech-
nique with super duals of codes. This method has been shown to be useful to find and prove
results concerning perfect 1-error correcting binary codes, see e.g., [7] and [8]. When using
the super dual method, it turned out that the following result of L. Redéi was useful:

Theorem 2 (Redéi [12, Satz 24’, p. 237]) If p is a prime number, then for any non linear
function f : Fp → Fp , the number of distinct differential quotients
f (x) − f (y)
, x, y ∈ Fq and x  = y,
x−y
will be at least (p + 3)/2.

In fact, by using the super dual property, and the above theorem, the proof of Theorem 1
is close to a triviality. So one further purpose of this paper is to demonstrate how the super
dual method can be used to, rather easy, obtain results also on perfect q-ary codes.

2 A group algebra

The results in this section concern codes in Fpn where p is a prime number and n is arbitrarily
chosen. We will use the fact that Fp = Z/pZ, and the notation
Fp = {0, 1, 2, 3, . . . , p − 1}.
Let C [x1 , . . . , xn ] denote the set of all linear combinations, over the complex number field
C , of all monomials

x1s1 x2s2 . . . xnsn where (s1 , s2 , . . . , sn ) ∈ Fpn .

The multiplication of monomials is defined by
x1s1 x2s2 . . . xnsn · x1t1 x2t2 . . . xntn = x1s1 +t1 x2s2 +t2 . . . xnsn +tn ,
where the addition in the exponents is performed in the field Fp . This multiplication is
extended to all elements of C [x1 , . . . , xn ] by the use of the distributive law.
To every subset C of Fpn we may associate an element C(x̄) of C [x1 , . . . , xn ] by
 s
C(x̄) = x11 . . . xnsn where s̄ = (s1 , s2 , . . . , sn ). (1)
s̄∈C

Let S1 (0̄) denote the set of words at distance at most one from the all zero word. For example
we thus get that

n
p−1
S1 (0̄)(x̄) = 1 − n + (1 + xi + xi2 + · · · + xi ). (2)
i=1

The following lemma follows immediately from the definition of multiplication in C [x1 ,
. . . , xn ]. It must be remarked that this lemma and the following two lemmas were already
proved in [6].

123
On perfect p-ary codes of length p + 1 49

Lemma 1 For any two subsets A and B of Fpn ,

(A + B)(x̄) = A(x̄)B(x̄),
where the elements of A + B are counted with their multiplicities.
Let ϑ  = 1 be a p:th root of unity. (As p is a prime, ϑ will then automatically be a primitve
p:th root of unity.) Define, for any t¯ ∈ Fpn ,

1 
n
(1 + ϑ −ti xi + ϑ −2ti xi2 + ϑ −3ti xi3 + · · · + ϑ −(p−1)ti xi
(p−1)
yt¯(x̄) = ).
pn
i=1
The following lemma is easy to verify, see e.g., [6]:
Lemma 2 The set C [x1 , . . . , xn ] is a vector space of dimension p n over the complex number
field and the set of vectors yt¯(x̄), for t¯ ∈ Fpn , constitutes a set of base vectors for this vector
space.
Hence any element C(x̄) can in a unique way be expanded

C(x̄) = At¯(C)yt¯(x̄), (3)
t¯∈Fpn

for some coefficients At¯(C) ∈ C . In accordance with the following lemma, these coefficients
will be called the Fourier coefficients of C.
Lemma 3 For any two words t¯ and t¯ of Fpn ,

yt¯(x̄) if t¯ = t¯
yt¯(x̄)yt¯ (x̄) =
0 else
Proof Straightforward verifications, see e.g. [6].
For any subset C of Fpn , the Fourier coefficients At¯(C) of C may be calculated by a
substitution:
At¯(C) = C(ϑ t1 , ϑ t2 , . . . , ϑ tn ). (4)
We will need the following proposition. Note that in this proposition, n will be equal to p + 1.
p+1
Proposition 2 For any perfect 1-error correcting code C in Fp
At¯(C)  = 0 and t¯  = 0̄ ⇒ w(t¯) = p.
Proof We first note that

p if si = ti
1 + ϑ −ti ϑ si + ϑ −2ti ϑ 2si + · · · + ϑ −(p−1)ti ϑ (p−1)si =
0 else
From Lemma 1 and the definition of a perfect code, we get that
S1 (0̄)(x̄)C(x̄) = p n y0̄ (x̄).
Hence, from Lemma 3 and Eq. 3 we thus get that
At¯(C)  = 0 and t¯  = 0̄ ⇒ At¯(S1 (0̄)) = 0.
By using the relation (4) and Eq. 2, we deduce in this case that
At¯(S1 (0̄)) = 1 − (p + 1) + (p + 1 − w(t¯))p = (p − w(t¯))p,
and the proposition is proved.

123
50 O. Heden

In the next proposition, we need the definition of the dual space D ⊥ of a subspace D of
the vector space Fpn :

D ⊥ = {c̄ ∈ Fpn | c̄ · d̄ = 0 for all d̄ ∈ D},

where c̄ · d̄ denotes the dot product

c̄ · d̄ = c1 · d1 + · · · + cn · dn (mod p)
of the vectors c̄ = (c1 , . . . , cn ) and d̄ = (d1 , . . . , dn ) of Fpn .
Proposition 3 The dual space of the kernel of C will be the linear span of those t¯ ∈ Fpn such
that At¯(C)  = 0:
ker(C)⊥ = { t¯ ∈ Fpn | At¯(C)  = 0 }.
Proof The kernel of C consists of exactly those words d̄ = (d1 , . . . , dn ) such that
d̄ + C = C
or equivalently
x d1 x d2 . . . xndn C(x̄) = C(x̄).
From the relation (4) we get that

At¯(d̄ + C) = ϑ d1 ·t1 ϑ d2 ·t2 · · · ϑ dn ·tn At¯(C) = ϑ d̄·t¯At¯(C).

As the Fourier expansion (3) is unique for every code C, we thus may conclude that

d̄ ∈ ker(C) ⇐⇒ ϑ d̄·t¯ = 1 for all t¯ such that At¯(C)  = 0.

As ϑ is a p:th root of unity, where p is a prime number and ϑ  = 1, the proposition
follows.
It must be remarked that what will be used below, and what is a direct consequence of the
above proposition, is especially the fact that
t¯  ∈ ker(C)⊥ ⇒ At¯(C) = 0.
The converse statement will not be true.
Now we discuss the calculation of the Fourier coefficients of a perfect code C in Fpn . We
will make use of the following lemma.
Lemma 4 For any t¯ ∈ Fpn the following is true:
(i) At¯(D ∪ E) = At¯(D) + At¯(E) for any two sets D and E.
(ii) At¯(D + E) = At¯(D) · At¯(E) for any two sets D and E.
(iii) At¯(D) = |D| if D is a linear subspace and t¯ ∈ D ⊥ .
Proof The lemma follows immediately from the definition of Fourier coefficients.
The kernel of C is a subspace of Fpn and any word of C either belongs to ker(C) or to a
coset of ker(C). Hence
C = (ḡ0 + ker(C)) ∪ (ḡ1 + ker(C)) ∪ (ḡ2 + ker(C)) ∪ . . . ∪ (ḡt + ker(C)), (5)
where t + 1 = |C|/|ker(C)|.
Thus, from Eq. 4 and Proposition 3 follows that
 
|ker(C)| ti=0 ϑ ḡi ·t¯ if t¯ ∈ ker(C)⊥ ,
At¯(C) = (6)
0 if t¯  ∈ ker(C)⊥ .

123
On perfect p-ary codes of length p + 1 51

For any code C and any t¯ ∈ ker(C)⊥ we let f¯t¯(C) denote the following vector:

f¯t¯(C) = (ḡ0 · t¯, ḡ1 · t¯, . . . , ḡt · t¯).

We observe that for any word t¯ ∈ ker(C)⊥ , the vector f¯t¯(C) = (f0 , f1 , . . . , ft ) is indepen-
dent of the choice of coset representatives of the kernel of C.
Further, the dot product is linear, and hence

f¯λt¯+µs̄ (C) = λf¯t¯(C) + µf¯s̄ (C).

Hence, from Eq. 6 we may thus deduce that we get all non-zero fourier coefficients At¯(C)
of C and thus also all information of C, by considering the vectors f¯t¯(C) for a set of base
vectors t¯ for the dual space of the kernel of C.

We will need the following lemma.

Lemma 5 Let p be a prime number, ϑ  = 1 be a p:th root of unity and fi , for i =

1, 2, . . . , p − 1, integers such that 0 ≤ fi ≤ p − 1. Then

1 + ϑ f1 + ϑ f2 + · · · + ϑ fp−1 = 0 ⇐⇒ {0, f1 , f2 , . . . , fp−1 } = {0, 1, 2, . . . , p − 1}.

The proof follows immediately from a result about field extensions:

Proof As ϑ is a p:th root of unity distinct from 1, we get that ϑ is a zero of the polynomial

x p−1 + x p−2 + · · · + x + 1.

As p is a prime number, it will follow from [19, p. 177] (or [9, p. 204]), that this polynomial
is irreducible in the polynomial ring Q[x], where Q is the field of quotients.

The next proposition concerns perfect p-ary codes of length p + 1 and with a kernel of
dimension p−2. The number of words of C will in this case be p p−1 and hence the number of
cosets in C of the kernel will be exactly p. Thus, if 0̄ belongs to C, we may conclude that for
any t¯ ∈ ker(C)⊥ , f¯t¯(C) = (0, f1 , . . . , fp−1 ), for some sequence of integers f1 , . . . , fp−1 ,
such that 0 ≤ fi ≤ p − 1, i = 1, 2, . . . , p − 1.

p+1
Proposition 4 Assume that C is a perfect code in Fp containing the all zero word and
that the dimension of the kernel of C equals p − 2.
For any f¯t¯(C) = (0, f1 , . . . , fp−1 ) the following is true:

t¯ ∈ ker(C)⊥ \ {0̄} and w(t¯)  = p ⇒ {0, f1 , . . . , fp−1 } = {0, 1, 2, . . . , p − 1}. (7)

Proof If w(t¯)  = p and t¯  = 0̄ then, by Proposition 2, the Fourier coefficient At¯(C) must be
equal to zero. Hence, if further t¯ ∈ ker(C)⊥ we will get by Eq. 6, that


t
ϑ ḡi ·t¯ = 0.
i=0

The proposition now follows from Lemma 5, as we may assume that ḡ0 = 0̄.

123
52 O. Heden

3 The super dual

It will be convenient to consider a concatenated matrix associated to every perfect code. This
concatenated matrix will here be called the super dual matrix of the code. For the binary case,
this concept was introduced in [7].
Let t¯1 , t¯2 , . . . , t¯κ be any set of base vectors for the dual space of the kernel of C and let
f¯t¯(C) have the same meaning, as in the previous section. A generator matrix for the super
dual of the perfect code C or a super dual matrix of C will be the matrix:
⎡ ⎤
t¯1 f¯t¯1 (C)
⎢ t¯2 f¯t¯ (C) ⎥
⎢ 2 ⎥
(T |F ) = ⎢ . .. ⎥ .
⎣ .. . ⎦
¯tκ f¯t¯κ (C)
The super dual of C is defined to be the row space of this matrix.
Let col(F ) denote the set of columns of the matrix F .
Lemma 6 If C is a perfect code in Fpn , with a super dual matrix (T |F ), then

C = {c̄ ∈ Fpn | T c̄T ∈ col(F )}.

Proof From the definition of the super dual follows, that the null space N (T ) of the matrix
T equals the kernel of C. Further, it follows from that definition, by using Eq. 5 and the
definition of the vectors ft¯(C), that
C = ∪ti=0 (N (T ) + ḡi ), where T ḡiT = column no, i + 1 of F . (8)
The lemma is proved.
Let T denote the set of columns, that we get by multiplying each of the columns of the
matrix T , with the elements of the field Fp . By [2] and the previous lemma, it is rather easy
to prove the following proposition:
Proposition 5 For any perfect code C in Fpn , the couple (T , col(F )) constitutes a tiling of
Fpκ , where κ = n − dim(ker(C)).
This proposition will not be used in this study.

4 Proof of Theorem 1
p+1
In this section, we will assume that C is a perfect code in Fp of rank p and with a kernel
of dimension p − 2. We also assume that the all zero word belongs to C.
p+1
As the vector space Fp has dimension p + 1 over the finite field Fp , we get that
dim(ker(C)⊥ ) = 3.
By Proposition 3, there are words t¯, s̄ and ū, that constitute a set of base vectors for ker(C)⊥
and such that
At¯(C)  = 0, As̄ (C)  = 0, Aū (C)  = 0.
Trivially, any vector in the dual space of C will belong to the dual space of ker(C). We
may thus also assume that e.g., t¯ ∈ C⊥ . This implies that
f¯t¯(C) = (0, 0, 0, . . . , 0).

123
On perfect p-ary codes of length p + 1 53

From Proposition 2, we may conclude that w(t¯) = p.

We get equivalent perfect codes by permutations of the set of coordinate positions and
also by permutations of the alphabet in each coordinate position, e.g., by multiplying with
a non zero element of the field Fp . Multiplying with non zero elements will transform any
non zero Fourier coefficient At¯(C) into a non zero Fourier coefficient Aπ(t¯) (C  ) of the code
C  so equivalent to C. Hence, we may without loss of any generality, and for the purpose of
simplifying notation, assume that
t¯ = (0, 1, 1, . . . , 1).
Consider a super dual matrix (T |F ) of the perfect code C. From elementary linear algebra,
we get from Lemma 6 and Eq. 8, that the rank of C will be equal to
rank(C) = dim(ker(C)) + rank(F ) = p − 2 + dim({f¯s̄ (C), f¯ū (C)}).
Hence, as the rank of C equals p, we may conclude that the vectors f¯s̄ (C) and f¯ū (C) must
be linearly independent. We will use the following notation:
f¯s̄ (C) = (0, f (1), f (2), . . . , f (p − 1)), f¯ū (C) = (0, g(1), g(2), . . . , g(p − 1)). (9)
We may conclude, that there will be a generator matrix for the super dual of C of the following
shape: ⎡ ⎤
0 1 1 ··· 1 0 0 0 ··· 0
⎣ t21 t22 t23 · · · t2,p+1 0 f (1) f (2) · · · f (p − 1) ⎦ .
t31 t32 t33 · · · t3,p+1 0 g(1) g(2) · · · g(p − 1)
We will, beside Theorem 2, also use the following variant of that theorem.

Theorem 3 (Redéi [12, Satz 24", p. 241]) If p is a prime number, then for any two maps f
and g from Fp to Fp such that the following three conditions are true

(i) neither f nor g are constant functions,

(ii) neither ξ → f (ξ ) nor η → g(η) are permutations of Fp
(iii) [ ξ, η ∈ Fp and ξ  = η ] ⇒ [ f (ξ )  = f (η) or g(ξ )  = g(η) ],

then the number of different differential quotients

g(ξ ) − g(η)
ξ, η ∈ Fp and f (ξ )  = f (η)
f (ξ ) − f (η)
will be at least (p + 1)/2.

We now prove that the functions f and g, defined in Eq. (9) and with f (0) = 0 and
g(0) = 0, will satisfy the conditions of the above theorem.
As the rank of the code C is greater than p − 1, we may conclude, that for some elements i
and j , f (i)  = 0 and g(j )  = 0. Else, the vectors f¯s̄ and f¯ū would not be linearly independent.
Hence the condition (i) of the theorem is satisfied.
From Eq. 6 and Lemma 5, as both As̄ (C) and Aū (C) are non zero, follows that the
condition (ii) of the theorem is satisfied.
Finally also the condition (iii) must be satisfied, as for any two distinct coset representa-
tives ḡi and ḡj of ker(C) in C in Eq. 5,

T ḡiT  = T ḡjT ,
compare Eq. 8.

123
54 O. Heden

As the condition (ii) of the Redéi Theorem is satisfied, it follows from Proposition 4, that
w(s̄) = p and w(ū) = p, and further, that the same must be true if we add multiples of the
vector t¯. Hence

w(s̄ + λt¯) = p and w(ū + λt¯) = p for every λ ∈ Fp .

This implies that the generator matrix of the super dual of C, after a suitable permutation of
the set of coordinate positions, can be written
⎡ ⎤
0 1 1 1 ··· 1 0 0 0 ··· 0
⎣1 0 1 2 · · · p − 1 0 f (1) f (2) · · · f (p − 1) ⎦ .
α π(0) π(1) π(2) · · · π(p − 1) 0 g(1) g(2) · · · g(p − 1)

where α  = 0 and where π is a permutation of the set of elements in Fp .

We now consider the number of values of µ, for which the function g + µf is not a per-
mutation of the set of elements of Fp . We get an estimate of that number, from the theorem
of Redéi.
The functions f and g satisfy the conditions of the above theorem of Redéi. Let denote
the set of elements µ such that
g(ξ ) − g(η)
−µ = , where f (ξ )  = f (η),
f (ξ ) − f (η)
is achieved for two distinct elements ξ and η of Fp . By the above theorem of Redéi we get
that
| | ≥ (p + 1)/2.

Further we may conclude that

µ∈ ⇒ {0}  = {(g + µf )(ξ ) | ξ ∈ Fp }  = Fp (10)

We now are able to prove that π must be a linear function. Assume that the function π is
not a linear function. Let denote the set of elements µ such that
π(ξ ) − π(η)
−µ = , for some ξ, η ∈ Fp , such that ξ  = η.
ξ −η
Then, as p is a prime number, by the other theorem of Redéi, Theorem 2,

| | ≥ (p + 3)/2.

We also note that

µ∈ ⇒ {π(ξ ) + µξ | ξ ∈ Fp }  = Fp .

As | ∪ | ≤ p, there will be at least one element µ0  = 0 such that

1
∈ ∩ .
µ0
Let γ0 be defined by using two distinct elements ξ and η with the property

γ0 = ξ + µ0 π(ξ ) = η + µ0 π(η), (11)

and let
v̄ = s̄ + µ0 ū − γ0 t¯.

123
On perfect p-ary codes of length p + 1 55

From Eq. 11, we get that w(v̄) ≤ p − 1. From the fact that π is assumed to be non linear,
we get that v̄  = 0̄. Further, from Eq. 10 we get that

{0} ∪ {f (ξ ) + µ0 g(ξ ) − γ0 · 0 | ξ ∈ Fp }  = Fp ,

and consequently, by Lemma 5, we may conclude that Av̄ (C)  = 0. We have thus proved that
if π is non linear, then there is an element v̄ such that

Av̄ (C)  = 0 and 1 ≤ w(v̄) ≤ p − 1.

This contradicts Proposition 2 and we may thus conclude that π has to be linear.
As the function π now is proved to be linear, we can, by using a sequence of suitable ele-
mentary row operations, achieve a generator matrix for the super dual of C of the following
shape: ⎡ ⎤
0 1 1 1 1 ... 1 0 0 0 0 0 0
(T |F ) = ⎣ 0 0 1 2 3 . . . p − 1 0 α1 α2 α3 . . . αp−1 ⎦ . (12)
1 0 0 0 0 ... 0 0 β1 β2 β3 . . . βp−1
It follows from Proposition 4, as the weights of the second and third row, of the matrix T in
the above Eq. 12, are not equal to p, that

{α1 , α2 , α3 , . . . , αp−1 , 0} = Fp and {β1 , β2 , β3 , . . . , βp−1 , 0} = Fp . (13)

If C is nonlinear, then we get from the super dual property that the kernel of C equals the
null space N (T ) of the matrix T . Further we get that

(x1 , x2 , . . . , xp+1 ) ∈ ker(C) ⇒ x1 = 0. (14)

Let

0 1 1 1 1 ··· 1
H = .
1 0 1 2 3 ··· p − 1
We then get that the null space of T satisfies


p+1 0
N (T ) = x̄ = (0, x2 , x3 , . . . , xp+1 ) ∈ Fp | H x̄ T = .
0

By Lemma 6 and especially Eq. 8, we then get that C is the union of the null space N (T )
of the matrix T and cosets of N (T ) with coset representatives ḡi such that
⎡ ⎤
0
T ḡi = ⎣ αi ⎦ ,
βi

or equivalently ḡi = (βi , −αi , αi , 0, 0, . . . , 0), for i = 1, 2, . . . , p − 1.

We now perform the following permutation of the symbols in the first position:

τ1 (βi ) = −αi for i = 1, 2, 3, . . . , p − 1, and τ1 (0) = 0.

By this operation, the perfect code C is transformed into a perfect code χ(C) equivalent to
C and the set of coset representatives ḡi , i = 1, 2, . . . , p − 1, is mapped onto the set of non
p+1
zero words of a subspace of Zp . This implies that χ(C) is the Hamming code, as defined
in Proposition 1.
Theorem 1 is proved.

123
56 O. Heden

5 Remarks

The corresponding problem for binary codes is almost trivial. In F2n , if a set C ⊆ F2n is a
union of a linear subspace L of F2n and a coset of L, then also C will be a linear subspace.
We have used rather deep results, the Redéi Theorem and a result about irreducible poly-
nomials over Q[x], to prove a rather weak result, that all perfect codes of length p + 1, rank
p and with a kernel of dimension p − 2 are equivalent to linear perfect codes. Indeed, it
would be interesting to find a direct proof, avoiding these results.
Another more interesting, and actually a most interesting problem, would be to find and
analyze a perfect code of length p, rank p and with a kernel of dimension p − 2.

Acknowledgments I am grateful to my father Göran Heden, who introduced me to the beautiful subject of
perfect codes, and to my thesis supervisor Bernt Lindström, who convinced me in 1973, that perfect codes is
a serious mathematical field, by letting me know about the result of Tietäväinen [18].

References

1. Best M.R.: Perfect codes hardly exist. IEEE Trans. Inform. Theory 29(3), 349–351 (1983).
2. Blokhuis A., Lam C.W.H.: More coverings by rook domains. J. Combin. Theory Ser. A 36, 240–244
(1984).
3. Etzion T.: Nonequivalent q-ary perfect codes. SIAM J. Discrete Math. 9(3), 413–423 (1996).
4. Golay M.J.E.: Notes on digital coding. Proc. IRE 37 (1949), Correspondence 657.
5. Hamming, Error detecting and error correcting codes. Bell Syst. Tech. J. 29, 147–160 (1950).
6. Heden O.: A generalised Lloyd Theorem and mixed perfect codes. Math. Scand. 37, 13–26 (1975).
7. Heden O.: Perfect codes from the dual point of view I. submitted to Discrete Math.
8. Heden O.: A full rank perfect code of length 31. Des. Codes Cryptogr. 38, 125–129 (2006).
9. Lang S.: Algebra. Addison-Wesley Publishing Company, Reading (1965).
10. Leontiev V.K., Zinoviev V.A.: Nonexistence of perfect codes over galois fields. Problems Inform. Theory
2(2), 123–132 (1973).
11. Lindström B.: On group and nongroup perfect codes in q symbols. Math. Scand. 25, 149–158 (1969).
12. Phelps K.T.: A general product construction of perfect codes. SIAM J. Algebra Discrete Method. 4,
224–228 (1983).
13. Phelps K.T., Villanueva M.: Ranks of q-ary 1-perfect codes. Des. Codes Cryptogr. 27(1–2), 139–144
(2002).
14. Phelps K.T., Rifà J., Villanueva M.: Kernels and p-kernels of p r -ary 1-perfect codes. Des. Codes Cryptogr.
37(2), 243–261 (2005).
15. Rédei L.: Lückenhafte Polynome über endlischen Körpern. Birkhäuser verlag Basel und Stuttgart (1970).
16. Schönheim J.: On linear and nonlinear single-error-correcting q-ary perfect codes. Inform. Control 12,
23–26 (1968).
17. Solov’eva F.I.: On Perfect Codes and Related Topics. Com2 Mac Lecture Note Series 13, Pohang (2004).
18. Tietäväinen A.: On the nonexistence of perfect codes over finite fields. SIAM J. Appl. Math. 24, 88–96
(1973).
19. van der Waerden B.L.: Algebra, Erster Teil, Siebte Auflage der Modernen Algebra. Springer-Verlag,
Berlin (1966).

123