Information Processing Letters 81 (2002) 277–282

Improved lower bound on the Shannon capacity of C7

Aleksander Vesel a,∗,1 , Janez Žerovnik b,c,2
a PEF, University of Maribor, Koroška c. 160, si-2000 Maribor, Slovenia
b FME, University of Maribor, Smetanova 17, si-2000 Maribor, Slovenia
c IMFM, Jadranska 19, si-1111 Ljubljana, Slovenia

Received 6 March 2000

Communicated by F. Dehne

Abstract
An independent set with 108 vertices in the strong product of four 7-cycles (C7 ! C7 ! C7 ! C7 ) is given. This improves
the best known lower bound for the Shannon capacity of the graph C7 which is the zero-error capacity of the corresponding
noisy channel. The search was done by a computer program using the “simulated annealing” algorithm with a constant time
temperature schedule.  2002 Elsevier Science B.V. All rights reserved.
Keywords: Shannon capacity; Combinatorial problems; Design of algorithms

1. Introduction and definitions
The study of Shannon capacity was introduced by
Shannon in [12], and motivated by the determination
of the zero-error capacity of a noisy channel. Shannon
formulated the problem in the form of graph theory
and supplied some partial results. It turns out that the
solution of the problem requires the determination of
the independence number of product of graphs which
contain odd cycles.
The Shannon capacity of the 7-cycle is one of the
unsolved problems given in [5]. It may be interesting
to note that the capacity of C5 was studied already
* Corresponding author.
E-mail addresses: vesel@uni-mb.si (A. Vesel),
janez.zerovnik@imfm.uni-lj.si (J. Žerovnik).
1 Partially supported by the Ministry of Science and Technology
of Slovenia, grant no. 0101-P-504.
2 Partially supported by the Ministry of Science and Technology
of Slovenia, grant no. J2-1015-0101.
0020-0190/02/$ – see front matter  2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 2 0 - 0 1 9 0 ( 0 1 ) 0 0 2 2 9 - 0
by Shannon [12] in 1956, and was determined only
in 1979 by Lovász [9].
All graphs considered in this paper are connected fi-
nite undirected graphs without loops or multiple edges.
If G is a graph, we shall write V (G) or V for its ver-
tex set and E(G) or E for its edge set. E(G) is a set of
unordered pairs xy = {x, y} of distinct vertices of G.
The strong product of graphs G and H is the graph
G ! H with vertex set G × H and (x1 , x2 )(y1 , y2 ) ∈
E(G ! H ) whenever x1 y1 ∈ E(G) and x2 = y2 ,
or x2 y2 ∈ E(H ) and x1 = y1 , or x1 y1 ∈ E(G) and
x2 y2 ∈ E(H ). The strong product is commutative and
associative in an obvious way, having the trivial graph
as a unit.
A set S ⊆ V (G) is independent if xy ∈ / E(G) for
any pair of vertices x, y ∈ S. Cardinality of a largest
independent set S of G is the independence number of
G, α(G).
The Shannon capacity Θ(G), of a graph G, is
defined by:
278 A. Vesel, J. Žerovnik / Information Processing Letters 81 (2002) 277–282

"1/n
Θ(G) := sup α(Gn )
!
, Table 1
n Independence numbers of C7n , n " 5
where Gn = G ! G ! · · · ! G (n times). n 1 2 3 4 5
This paper is structured as follows. In the next sec-
α(C7n ) 3 10 33 100–115 343–402
tion we recall an upper bound on the Shannon capac-
ity and some known results on the independence num-
ber of C7n . In Section 3 we consider the allocation of Even for the product of three cycles there are several
potential independent vertices in C74 . In Section 4 we products with unknown independence numbers [15].
introduce the simulated “annealing” algorithm with a The general lower bound on the independence
constant time temperature schedule and with heuris- number follows from the well-known inequality
tically restricted solutions. An independent vertex set
with 108 vertices is obtained with the algorithm. This Lemma 3. For all graphs G and H ,
improves the best known lower bound for the Shan- α(G ! H ) # α(G)α(H ).
non capacity of C7 . We conclude with two questions
presented in Section 5. The next lemma can be used to bound independent
size growth in the strong products of cycles.
2. Preliminaries Lemma 4. For any graph G and k # 2,
# $
Lovász [9] obtained general upper bounds on the α(G)
α(G ! C2k+1 ) " kα(G) + .
Shannon capacity which yield the exact value of the 2
capacity for many graphs, including the long-standing
open case of the pentagon. The independence numbers of C7n are known only
for n " 3. For n = 2 the result is obtained using
Theorem 1 [9]. For any d-regular graph G on n Theorem 2 and for n = 3 the result was obtained by
vertices a computer search [2,15].
−n · λmin For n > 3 this numbers are bounded by Lemmas 4
Θ(G) " , and 3. The situation is showed in Table 1 for n " 5.
d − λmin
Note that there is an exception: the lower bound for
where λmin denotes the least eigenvalue of G.
α(C75 ) is obtained by a construction of a set of 343
independent vertices in [2].
However, the Shannon capacity of the 7-cycle is still
One can obtain the following lower bound and
an unsolved problem. In order to improve the lover
the upper bound on the Shannon capacity of C7
bound, the independence number of C74 is considered
applying the lower bound for α(C75 ) and Theorem 1,
in this paper.
respectively.
The independence numbers of the strong product
of even cycles are well known. The problem is √
Proposition 5. 5
343 = 3.21409 " Θ(C7 ) " 3.3176.
incomparably more difficult for more cycles with
odd length. Hales [7], as well as Sonnemann and
In order to improve the lower bound the computer
Krafft [14], discovered the explicit formula for the
search for larger independent sets in C74 has been
independence numbers of the strong product of two
applied. We used a randomized heuristics based on
odd cycles.
the simulated annealing, one of the most popular
heuristics.
Theorem 2 [7,14]. For j, k ∈ N, j # k,
# $
k
α(C2j +1 ! C2k+1 ) = j k + .
2 3. The structure of a solution

The independence numbers of products of more At the beginning of our computer search, the allo-
then two odd cycles are still unknown in most cases. cation of the independent vertices in the initial solu-
 A. Vesel, J. Žerovnik / Information Processing Letters 81 (2002) 277–282 279

tion as well as in the intermediary solutions was not
restricted. Since the search was not particularly suc-
cessful (only independent sets with at most 100 ver-
tices were found), we constrained the distribution of
the independent vertices in the solutions as follows.
It is a straightforward exercise to obtain the follow-
ing well-known result.
Lemma 6. For a graph G and n ∈ N,
α(G ! Kn ) = α(G).
There are 10 cells with three independent vertices in
the table. Without the numbers in the cells, Table 2 can
also be seen as a representation of the graph C7 ! C7 ,
where a cell corresponds to a simple vertex. Note
that the positions of the cells with three independent
vertices correspond to the positions of the vertices in
the largest independent vertex set of C7 ! C7 . (α(C7 !
C7 ) = 10 by Theorem 2.) Thus, the allocation of 108
potentially independent vertices seems to be very well
balanced. It is also obvious that larger independent sets
cannot be balanced in this way.

The following obvious corollary will help us to

establish the structure of a possible solution. 4. Simulated “annealing” with a constant
temperature schedule
Corollary 7.
(i) α(C72 ! K4 ) = α(C72 ) = 10. Simulated annealing (see, for example, [8]) can be
(ii) α(C73 ! K2 ) = α(C73 ) = 33. understood as a random relaxation of the iterative im-
provement algorithm [1]. The moves to higher cost so-
This result suggests that vertices in a large indepen- lutions are accepted only with a certain probability.
dent set should be allocated quite evenly in the graph. This acceptance probability depends on the parameter,
The C74 can be visualized as a a square array of which is usually called the temperature. Sometimes
49 cells where a cell corresponds to a copy of C7 ! very good solutions are found by SA, but it is also
C7 . For i ∈ {0, 1, . . ., 48}, a copy of C7 ! C7 will well known that SA can be extremely time consuming.
be denoted (C7 ! C7 )i . Consider Table 2, where the Furthermore, the best known convergence results hold
number in a cell denotes the number of potentially only if the so-called cooling schedule is very slow [10,
independent vertices in the corresponding product of 6]. On the other hand, it is known for some time that at
two 7-cycles (C7 ! C7 )i . least asymptotically and with a usual implementation,
From Corollary 7 it follows that (i) four mutually where rejected moves are explicitly counted, the prob-
adjacent cells can have at most ten independent ver- ability of success of simulated annealing is worse than
tices together and (ii) two adjacent rows (columns) can as simple a procedure as multistarts of the iterative im-
have at most 33 independent vertices together. One can provement algorithm [4]. Recently, a detailed analysis
easily verify that both conditions are fulfilled in the of various temperature schedules was given [3].
represented distribution. Among various temperature schedules, a constant
temperature schedule for simulated “annealing” was
found competitive or at least worth consideration [3].
Table 2 Similar experience with a SA-like algorithm for graph
A distribution of 108 potentially coloring [11], namely very fast convergence of the
independent vertices
algorithm on an interval of “good temperatures” [13],
3 2 3 2 2 2 2 motivated us to try constant time schedule.
2 2 2 2 3 2 2 While all attempts with standard cooling schedules
failed, we succeeded using a constant time tempera-
2 3 2 2 2 2 3
ture schedule.
2 2 2 3 2 2 2 Based on the discussion described above we present
3 2 2 2 2 3 2 the following procedure (see Fig. 1). The cost function
2 2 3 2 2 2 2 C(S) is given by C(S) = |S| − mS , where mS is the
number of edges in the subgraph induced by the set of
2 2 2 2 2 3 2
vertices S.
280 A. Vesel, J. Žerovnik / Information Processing Letters 81 (2002) 277–282

Procedure CONSTANT ANNEALING(T );

1. Randomly select a subset of vertices S such that the number of vertices
in (C7 ! C7 )i equals the number in the corresponding cell of Table 2;
3. for j := 1 to iteration length do begin
4. Randomly select i in {0, 1, . . . , 48}.
5. Randomly select vertices v and u in (C7 ! C7 )i , such that v ∈ S
and u ∈
/ S;
6. Sj := S ∪ {u} \ {v};
7. If C(S) # C(Sj ) then S := Sj
8. else if e(C(S)−C(Sj ))/T > rnd[0, 1) then S := Sj ;
end;
end.

Fig. 1.

Table 3 Theorem 8.
Influence of the parameter T on the performance of CONSTANT √
√ 2 arctan( 93 )+π
ANNEALING √
4 7(2 7 cos( ) + 1)
108 " Θ(C7 ) " 6
√ .
T 0.05 0.06 0.07 0.08 0.09 √ 2 arctan( 93 )+π
7 + 2 7 cos( 6 )
Smax 104 105 104 105 106
S̄ 102.70 103.15 103.60 104.40 104.20 Corollary 9. 3.2237 " Θ(C7 ) " 3.3176.
T 0.10 0.11 0.12 0.13 0.14
Smax 108 106 105 103 104
5. Conclusions
S̄ 104.35 103.85 103.30 102.60 102.70
In this paper we have reported on searching for
independent vertices in the strong product of four 7-
cycles C7 ! C7 ! C7 ! C7 . For computer search we
The algorithm was implemented in Delphi (Object used simulated annealing, one of the most popular
Pascal) and tested on the personal computer with randomized heuristics.
Pentium II processor. Experiments with CONSTANT In order to build an effective algorithm we also
ANNEALING for various ranges of the parameter heuristically restricted the structure and the cardinality
T have been performed. Table 3 shows the results of independent vertex sets. The algorithm was efficient
produced after 20 runs of the procedure with iteration enough to find an independent set with 108 vertices in
length = 109 and with T ranging from 0.05 to 0.14. the graph. This improves the best known lower bound
S̄ and Smax are the average size and the largest size of for the Shannon capacity of the graph C7 .
the independent set found, respectively. We think that the following two questions regarding
It appears that the procedure yields best results for the perspective of computer search in order to improve
T between 0.08 and 0.11. However, we cannot deduce the Shannon capacity of C7 are considerable.
that the optimal value of the parameter T is achieved
at T = 0.1 from Table 3. Indeed, the independent set Question 1. Is α(C74 ) > 108?
with 108 vertices depicted in Fig. 2 was found in one
of the preliminary runs of the procedure at T = 0.11. Although many experiments were done on search-
This improves the best known lower bound for the ing an independent set with more than 108 vertices, all
Shannon capacity of the graph C7 . The new lower attempts failed. Therefore we conjecture that α(C74 ) =
bound together with the upper bound is stated in the 108. Furthermore, we showed in the paper that the in-
next theorem. dependent set with more than 108 vertices cannot be
 A. Vesel, J. Žerovnik / Information Processing Letters 81 (2002) 277–282
Fig. 2. 108 independent vertices of C7 ! C7 ! C7 ! C7 .
well balanced in a way we did it for the smaller one.
This in some sort supports the conjecture.
Question 2. Is α(C75 ) > 343?
The graph C75 is a huge one. It has 16 807 vertices
and it seems that the problem exceeds the capability of
todays personal computers. Perhaps the search could
be performed on a supercomputer or as the alternative
one could consider a parallel implementation.
References
[1] E.H.L. Aarts, J.K. Lenstra, Local Search Algorithms, John
Wiley & Sons, Chichester, 1997.
[2] L.D. Baumert, R.J. McEliece, E. Rodemich, H.C. Rumsey,
R. Stanley, H. Taylor, A combinatorial packing problem, in:
Computers in Algebra and Number Theory, Proc. SIAM-
AMS Sympos. Appl. Math., Amer. Math. Soc., Providence,
RI, 1971, pp. 97–108.
281
[3] H. Cohn, M. Fielding, Simulated annealing: searching for an
optimal temperature schedule, SIAM J. Optim. 9 (1999) 779–
802.
[4] A. Ferreira, J. Žerovnik, Bounding the probability of success
of stochastic methods for global optimization, Comput. Math.
Appl. 25 (1993) 1–8.
[5] C.D. Godsil, Problems in algebraic combinatorics, Electron. J.
Combin. 2 (1995) 1–20.
[6] B. Hajek, G. Sasaki, Simulated annealing — to cool or not,
System Control Lett. 12 (1989) 443–447.
[7] R.S. Hales, Numerical invariants and the strong product of
graphs, J. Combin. Theory Ser. B 15 (1973) 146–155.
[8] P.J.M. Laarhoven, E.H.L. Aarts, Simulated Annealing, Theory
and Applications, D. Reidel Publishing Company, Dordrecht,
1987.
[9] L. Lovasz, On the Shannon capacity of a graph, IEEE Trans.
Inform. Theory 25 (1979) 1–7.
[10] D. Mitra, F. Romeo, A. Sangiovanni-Vincentelli, Convergence
and finite time behavior of simulated annealing, Adv. in Appl.
Probab. 18 (1986) 747–771.
[11] A. Petford, D. Welsh, A randomised 3-colouring algorithm,
Discrete Math. 74 (1989) 253–261.
282 A. Vesel, J. Žerovnik / Information Processing Letters 81 (2002) 277–282

[12] C.E. Shannon, The zero-error capacity of a noisy channel, IRE
Trans. Inform. Theory 2 (1956) 8–19.
[13] J. Shawe-Taylor, J. Žerovnik, Adapting temperature for some
randomized local search algorithms, Preprint Ser. Department
of Math. 36 (614) (1998).
[14] E. Sonnemann, O. Krafft, Independence numbers of product
graphs, J. Combin. Theory Ser. B 17 (1974) 133–142.
[15] A. Vesel, The independence number of the strong product of
cycles, Comput. Math. Appl. 36 (1998) 9–21.