Professional Documents
Culture Documents
6, JUNE 2003
I. INTRODUCTION
A quaternary linear [n; k] code C is a k -dimensional vector sub-
space of n 4 , where 4 is the finite field of four elements. The elements
of 4 are taken to be f0; 1; ; 2 g, where is denoted by 2 and 2 is
denoted by 3. The rate of a linear [n; k] code C is defined as k=n. The
elements of C are called codewords. The weight wt(x) of a codeword Fig. 1. The hyperoval of PG(2; 4).
x is the number of nonzero coordinates. The minimum weight of C is
the smallest weight among all nonzero codewords of C . An [n; k; d] and global conjugation.) This [6; 3; 4] code corresponds to the unique
code is an [n; k] code with minimum weight d. hyperoval of PG(2; 4) (or the 6-arc, six points no three on a line), as
The weight enumerator of C is shown in Fig. 1. The columns of the generator matrix correspond to
n points in PG(2; 4), which are denoted by (x; y; z ) with x; y; z 2 4 .
WC (y) = Ai yi The lines are denoted by [a; b; c] with a; b; c 2 4 . The notation and
i=0 terminology for geometry follows that in [3]. The order of the auto-
morphism group is 2160 = 24 3 33 3 5 (cf. [10]). Recently, Olsson has
where Ai is the number of codewords of weight i in C . shown that the [18; 9; 8] code over 4 is unique [8].
The dual code C ? of C is defined as A natural question that now arises is how many inequivalent optimal
linear rate–1=2 codes are there for other small lengths? In the next sec-
C ? = fx 2 4n jx 1 y = 0 for all y 2 C g tion, we present the classification of all optimal linear [n; n=2] codes
up to length 18.
where the Hermitian inner product is defined as x 1 y = xy with the
conjugacy operation y = y 2 . A code C is (Hermitian) self-dual if
C = C ? . This corresponds to the family 4H in [10]. Note that there are II. CLASSIFICATION OF CODES UP TO LENGTH 18
several families of self-dual codes over 4 , such as Euclidean (4E ), but
A. Classification of Codes
we consider only Hermitian self-dual codes here. A code C is formally
self-dual if C and C ? have identical weight enumerators. Self-dual In the classification of codes, we are only interested in one repre-
codes are by definition also formally self-dual. sentative from each code equivalence class. There is no unique defini-
A linear [n; k] code C is optimal if C has the highest minimum tion of code equivalence; we here use the definition that two codes are
weight among all linear [n; k] codes (see [1] for current bounds on equivalent if one can be obtained from the other by using a distance-pre-
the highest minimum weight). It is trivial that there is a unique quater- serving transformation that also preserves any special characteristics of
nary linear code with parameters [2; 1; 2]. It is well known that there the code (here, linearity).
are unique optimal quaternary [4; 2; 3] and [6; 3; 4] linear codes, up The set of all such transformations for linear codes are exactly those
to equivalence, the latter being the hexacode [10]. (The transforma- obtained by a composition of a monomial transformation and global
tions which give equivalent codes for the family 4H are monomials action of a field automorphism. This result seems to be well known,
but the only place where we have seen a proof of it is [7, Theorem
A.2]. This definition differs somewhat from the definition given in [10],
Manuscript received December 15, 1999; revised May 8, 2001. This work was
supported in part by the Natural Sciences and Engineering Research Council of
where the special case of dual linear codes are considered (but the def-
Canada and the Academy of Finland under Grants 44517 and 100500. initions coincide for small fields). See also [2, p. 27].
T. A. Gulliver was with the Department of Electrical and Electronic Engi- The optimal codes are determined in the following way. Every
neering, University of Canterbury, Christchurch, New Zealand. He is now with [n; n=2; d] code is equivalent to a code with generator matrix of the
the Department of Electrical and Computer Engineering, University of Victoria, form (I ; A) where A is an n=2 2 n=2 matrix (or with a parity-check
matrix of the same size). Thus, we need only consider n=2 2 n=2
STN CSC, Victoria, BC V8W 3P6, Canada (e-mail: agullive@ece.uvic.ca).
among all matrices obtained from A by permuting its rows. This or- TABLE I
dering is defined by regarding an 4 -vector as an integral vector by AUTOMORPHISM ORDERS FOR THE [8; 4; 4] CODES
mapping 0 to 0, 1 to 1, to 2, and 2 to 3.
The final generator (or parity-check) matrices are constructed by
adding one row (column) at a time and after each such step removing
copies of equivalent codes and codes that have minimum distance less
than d. The equivalence check was carried out using an approach that is
described in detail in [9]. The approach is based on a transformation of
the code into a graph. The graph is then fed into the program nauty [6],
which returns a certificate. Two codes have the same certificate exactly
when they are equivalent.
In the worst case, it takes exponential time for the program nauty
to compute a certificate. For the small instances considered here,
however, the time consumed is not prohibitive, and the classification
was completed using less than a week of CPU time. Most of the
time was used to classify the [14; 7; 6] codes. As those codes have
the same minimum distance as the optimal [12; 6; 6] code, there
is a multitude of such codes and, in particular, of intermediate
codes in the search.
Many of the computations were carried out independently in Finland
and New Zealand. The algebra software Magma was utilized by
the first author for determining code equivalence and automorphism
groups under monomial transformations (see also [5] for an algorithm
for determining monomial equivalence of codes), however, the results
presented in this correspondence also include global conjugation. 1000 0111 1000 0111
G8 9 1 =
0100 1011 ; G8 9 2 = 0100 1011
; ;
0010 1102 ; ;
0010 1102
B. Length 8
0001 1210 0001 1213
The maximum minimum distance of a quaternary linear [8; 4] code 1000 0111 1000 0111
is d = 4 [1]. The results of this search show that an optimal quaternary
G8 9 3 =
0100 1011 ; G8 10 = 0100 1011
linear [8; 4; 4] code is equivalent to one of C8; 1 , C8; 2 , C8; 3 , C8; 4 , 0010 1203 0010 1123
; ; ;
C8; 5 , C8; 6 , C8; 7; 1 , C8; 7; 2 , C8; 7; 3 , C8; 8; 1 , C8; 8; 2 , C8; 8; 3 , C8; 9; 1 , 0001 1313 0001 1132
C8; 9; 2 , C8; 9; 3 , and C8; 10 with generator matrices respectively.
The weight enumerators of C8; 1 , C8; 2 , C8 3 , C8 4 , C8 5 , C8 6 ,
1000 0111 1000 0111 C8; 7; 1 , C8; 7; 2 , C8; 7; 3 , C8; 8; 1 , C8; 8; 2 ,
; ; ;
C8 8 3 , C8 9 1 , C8 9 2 ,
;
1000 0111 1000 0111 where Ci; j; k has weight enumerator Wi; j , and k is omitted if only one
G8 = 0100 1011 ; G8 = 0100 1011 code has this weight enumerator.
; 7; 3
0010 1102 ; 8; 1
0010 1102 There are 16 inequivalent optimal quaternary linear [8; 4; 4] codes.
0001 1130 0001 1121 It is interesting that three of these—C8; 4 , C8; 7; 1 , and C8; 9; 1 —do not
have global conjugation in their automorphism group. By applying
1000 0111 1000 0111 global conjugation to these three codes, we get a total of 19 distinct
G8 = 0100 1011 ; G8 = 0100 1012 codes if only monomial permutations are used to test equivalence.
; 8; 2
0010 1123 ; 8; 3
0010 1120 Table I gives the orders of the automorphism groups for the [8; 4; 4]
0001 1203 0001 1203 codes.
1542 IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 49, NO. 6, JUNE 2003
TABLE II
AUTOMORPHISM ORDERS FOR THE [10; 5; 5] CODES
TABLE III
N (n), N (n) AND N (n) FOR 2 18
n
its automorphism group is jAut(q14 )j = 6552 = 23 3 32 3 7 3 13. same minimum distance as the optimal [18; 9; 8] codes, which leads to
Indeed, this is the only weight enumerator with gaps. an extensive number of (final and intermediate) codes, which, in turn,
affects the size of memory needed and the computation time (cf. our
F. Length 16 earlier comment on [14; 7; 6] codes).
The classification of optimal quaternary linear [16; 8; 7] codes Let NT (n), NS (n), and NF (n) denote the number of optimal codes,
shows that there are three codes, C16; 1 ; C16; 2 , and C16; 3 , with the number of optimal Hermitian self-dual codes and the number of
generator matrices optimal formally self-dual codes which are not self-dual. In Table III,
we list the values of NT (n), NS (n), and NF (n) for 2 n 18.
10000000 11111100
01000000 32211010 ACKNOWLEDGMENT
00100000 23311001
The authors would like to thank David Glynn who provided the geo-
00010000 13020211
G16; 1 = metric interpretation of the [6; 3; 4] and [8; 4; 4] codes.
00001000 30230311
00000100 20131121
REFERENCES
00000010 31332221
00000001 23003131
[1] A. E. Brouwer, “Bounds on the size of linear codes,” in Handbook of
Coding Theory, V. S. Pless and W. C. Huffman, Eds. Amsterdam, The
10000000 11111100 Netherlands: Elsevier, 1998, pp. 295–461.
[2] R. L. Griess, Jr., Twelve Sporadic Groups. Berlin, Germany: Springer-
01000000 32211010
Verlag, 1998.
00100000 23311001 [3] J. W. P. Hirschfeld, Projective Geometries Over Finite Fields, 2nd
00010000 13020211 ed. Oxford, U.K.: Oxford Univ. Press, 1998.
G16; 2 = [4] W. C. Huffman, “Characterization of quaternary extremal codes
00001000 22012311
of lengths 18 and 20,” IEEE Trans. Inform. Theory, vol. 43, pp.
00000100 00223221 1613–1616, Sept. 1997.
00000010 31221131 [5] J. S. Leon, “Computing automorphism groups of error-correcting
codes,” IEEE Trans. Inform. Theory, vol. IT–28, pp. 496–511, May
00000001 03210331
1982.
10000000 11111100 [6] B. D. McKay, “nauty User’s Guide (version 1.5),” Comp. Sci. Dept.,
Australian Nat. Univ., Canberra, Tech. Rep. TR-CS-90-02, 1990.
01000000 32211010 [7] J. Olsson, “Linear codes with performance close to the Singleton
00100000 23121001 bound,” Ph.D. dissertation, Linköping, Sweden, 1999.
00010000 23210211 [8] , “On the quaternary [18,9,8] code,” in Proc. WCC’99 Workshop
G16; 3 = : on Coding and Cryptography, Paris, France, Jan. 1999, pp. 65–73.
00001000 10302211 [9] P. R. J. Östergård, “Classifying subspaces of Hamming spaces,” Des.
00000100 33311121 Codes Cryptogr., vol. 27, pp. 297–305, 2002.
00000010 12003121 [10] E. M. Rains and N. J. A. Sloane, “Self-dual codes,” in Handbook of
Coding Theory, V. S. Pless and W. C. Huffman, Eds. Amsterdam, The
00000001 11010331 Netherlands: Elsevier, 1998, pp. 177–294.
All three codes have weight enumerator
7 8 9 10 11
W16 = 1 + 720y + 810y + 4800y + 3360y + 18144y
12 13 14 15 16
+7560y + 20160y + 4320y + 5328y + 333y :
jAut(C16; 1 )j = 90 = 2 3 32 3 5:
G. Length 18
The classification of optimal quaternary linear [18; 9; 8] codes
shows that there is a unique code C18 , which confirms the result given
by Olsson [8]. Note that this is the unique extremal self-dual [18; 9; 8]
code [4].
III. SUMMARY
The problem of classifying optimal linear codes is more difficult than
classifying extremal quaternary self-dual codes as the significant struc-
ture of the latter codes can be utilized in that process. So, whereas ex-
tremal quaternary codes have been classified up to length 20, we had
to stop the current classification at length 18. One reason for the diffi-
culty of classifying the optimal [20; 10; 8] codes is that they have the