Journal of Combinatorial Theory, Series A 94, 152155 (2001)

doi:10.1006jcta.2000.3127, available online at http:www.idealibrary.com on

NOTE
Binary B 2 -Sequences : A New Upper Bound
Gerard Cohen
Ecole Nationale Superieure des Telecommunications, 75 634 Paris 13, France
E-mail: coheninfres.enst.fr

Simon Litsyn
Department of Electrical Engineering-Systems, Tel-Aviv University, 69978 Ramat Aviv, Israel
E-mail: litsyneng.tau.ac.il

and
Gilles Zemor
Ecole Nationale Superieure des Telecommunications, 75 634 Paris 13, France
E-mail: zemorinfres.enst.fr
Communicated by the Managing Editors
Received February 23, 2000

We show that the maximum size of a B 2 -sequence of binary n-vectors for large
enough n is at most 2 0.5753n, thus improving on the previous bound 2 0.6n due to
B. Lindstrom.  2001 Academic Press

1. INTRODUCTION
A B 2 -sequence of vectors of dimension n over the alphabet A/N is a
set C/A n such that the sums x+ y are different for all distinct pairs
[x, y] of elements of C. When n=1 and A=N, we have a classical Sidon
set [7]. The binary case, i.e. A=[0, 1], was studied by B. Lindstrom in
[3, 4]. The associated asymptotic problem is to compute the quantity
r=lim sup
n

1
log 2 M n
n

where M n denotes the maximum size of a B 2 -sequence of dimension n.

152
0097-316501 35.00
Copyright  2001 by Academic Press
All rights of reproduction in any form reserved.

153

NOTE

The best bounds on r were provided by B. Lindstrom [4],

0.5r0.6.

(1)

Z. Furedi (unpublished) obtained later an =-improvement on the upper

bound 0.6.
In this note we shall call upon techniques from coding theory to improve
the upper bound to
r0.5753.
Our attention was drawn to this old problem by a relatively new cryptographic application concerned with fingerprinting binary data : this consists of identifying different copies of a binary document by codewords of
some code C. The problem of collusion occurs when two users compare
their fingerprinted copies. In certain situations (see [1]) pirate users will
erase the coordinates where their codewords differ. This means that they
replace one of the two fingerprints x and y by a ternary vector `(x, y) #
[0, 1, =] n defined by ` i =x i = y i when x i = y i and by ` i == when x i { y i .
We therefore need codes C with the property that `(x, y) always identifies [x, y] for any pair of codewords [x, y]. Notice that `(x, y) is
obtained from the real sum x+ y by changing every 2 coordinate into a 1,
every 1 coordinate into = and leaving 0's unchanged. What we are looking
for is thus exactly a B 2 -sequence.
For other relations between coding and additive number theory, see, for
example, [2].

2. A NEW UPPER BOUND ON r

We first need a lemma (similar to a result by van Tilborg [8]). Let
d( } , } ) stand for the Hamming distance.
Lemma 1. Let C be a B 2 -sequence. Then
|[(c 1 , c 2 ): c 1 , c 2 # C; d(c 1 , c 2 )=i ]| 

\i+ 2 .
i

Proof. Clearly the problem is equivalent to considering a code C with

all pairwise differences of codewords being distinct. Let us fix some i positions out of n, and pick the set of all pairs (c 1 , c 2 ) with d(c 1 , c 2 )=i and
such that they differ in the chosen i positions. The difference c 1 &c 2 has

154

NOTE

(n&i ) 0's in the positions outside the chosen ones, and \1 at the chosen
i positions. Since all differences must be distinct there can be at most 2 i
such pairs: furthermore there are exactly ( ni ) choices for i positions. K
Let (B 0 =1, B 1 , ..., B n ) stand for the distance distribution of the code C,
Bi=

1
|[(c 1 , c 2 ): c 1 , c 2 # C; d(c 1 , c 2 )=i ]|.
|C|

Thus the claim of the lemma is that

Bi 

( ni ) 2 i
.
|C|

(2)

If n is big enough this reduces to

1
log 2 B i H(&)+&&R(C),
n

(3)

where H(x)=&x log 2 x&(1&x) log 2 (1&x) is the binary entropy function, &=in, and R(C)= 1n log 2 |C| is the rate of C.
The smallest non-zero i such that B i {0 is called the minimum distance
of the code C and denoted by d. The following quantity, $ LP (R), is known
to coding theorists as the ``linear programming bound'' [6] and is the best
asymptotic upper bound to date on the normalized minimum distance
$=dn of a code as a function of its rate. We have
$ LP (R)=

min
0;:12, H(:)&H( ;)=1&R

:(1&:)&;(1&;)
12+- ;(1&;)

We shall use :
Lemma 2. For every code of big enough length n and rate R there exists
a component B &n of the distance distribution,
0<&$ LP (R),
such that
1
log 2 B i R&1+H(&).
n

(4)

NOTE

155

The proof of this lemma uses the non-negativity of the dual Hahn transform of a function of the distance distribution : it can be found in [5.
Theorem 4].
Lemmas 1 and 2 together will give our new upper bound.
Theorem 1.

An upper bound on the asymptotic rate of a B 2 -sequence is

rz,

where z<0.5753 is the root of the equation

2z&1=$ LP (z).
Proof. Comparing (3) and (4) we see that the only way both of them
are satisfied is that the solution & of
H(&)+&&R=R&1+H(&),
which is
&=2R&1,
satisfies &$ LP (R). Solving the equation
2R&1=$ LP (R),
we find its root, z=0.57525..., and we are done. K

REFERENCES
1. D. Boneh and J. Shaw, Collusion-secure fingerprinting for digital data, IEEE Trans. Inform.
Theory 44 (1998), 18971905.
2. G. Cohen and G. Zemor, Subset sums and coding theory, Asterisque 258 (1999), 327339.
3. B. Lindstrom, Determination of two vectors from the sum, J. Combin. Theory 6 (1969),
402407.
4. B. Lindstrom, On B 2 -sequences of vectors, J. Number Theory 4 (1972), 261265.
5. S. Litsyn, New upper bounds on error exponents, IEEE Trans. Inform. Theory 45 (1999),
385398.
6. R. J. McEliece, E. R. Rodemich, H. C. Rumsey, Jr., and L. R. Welch, New upper bounds
on the rate of a code via the DelsarteMacWilliams inequalities, IEEE Trans. Inform.
Theory 23 (1977), 157166.
7. S. Sidon, Ein Satz uber trigonometrische Polynome und seine Anwendung in der Theorie
der FourierReihen, Math. Ann. 106 (1932), 536539.
8. H. C. A. van Tilborg, Upper bounds on |C 2 | for a uniquely decodable pair (C 1 , C 2 ) for a
two-access binary adder channel, IEEE Trans. Inform. Theory 29 (1983), 386389.