Professional Documents
Culture Documents
Biclique Helly Graphs
Biclique Helly Graphs
net/publication/220397983
Biclique-Helly Graphs
CITATIONS READS
11 44
2 authors:
Some of the authors of this publication are also working on these related projects:
Recognizing vertex intersection graphs of paths on bounded degree trees View project
All content following this page was uploaded by Jayme Luiz Szwarcfiter on 03 February 2015.
Biclique-Helly Graphs
Marina Groshaus1, , Jayme L. Szwarcfiter2,
1
Universidad de Buenos Aires, Facultad de Ciencias Exáctas y Naturales, Departamento
de Computación, Argentina. e-mail: groshaus@dc.uba.ar
2
Universidade Federal do Rio de Janeiro, Instituto de Matemática, NCE and COPPE,
Brasil. e-mail: jayme@nce.ufrj.br
Abstract. A graph is biclique-Helly when its family of (maximal) bicliques is a Helly family.
We describe characterizations for biclique-Helly graphs, leading to polynomial time recog-
nition algorithms. In addition, we relate biclique-Helly graphs to the classes of clique-Helly,
disk-Helly and neighborhood-Helly graphs.
Key words. Bicliques, bichromatic cliques, biclique-Helly graphs, clique-Helly graphs, disk-
Helly graphs, neighborhood-Helly graphs.
1. Introduction
Helly families of subsets have been studied in different contexts. In the scope of
graph theory, this study has motivated the introduction of some classes of graphs,
as clique-Helly graphs [6, 9, 10, 12, 19], disk-Helly graphs [2, 3], and neighborhood-
Helly graphs [4]. These classes correspond to the cases where the families subject to
the Helly property are cliques, disks and neighborhoods, respectively.
Bicliques, in general, have been considered in some different contexts, e.g. [13,
14, 16, 18].
In this work, we consider the graphs whose bicliques form a Helly family, the
biclique-Helly graphs. Besides the interest of examining bicliques in the scope of
the Helly property, these graphs could be of interest in the study of retracts [11].
In fact, retracts of bipartite graphs are related to neighborhood-Helly graphs [1]
and the latter are related to biclique-Helly graphs in some different aspects. We also
mention that some optimization problems, as the edge modification problem, have
been already studied for the class of biclique-Helly graphs [7].
We describe two characterizations for biclique-Helly graphs. Both of them lead
to algorithms for recognizing biclique-Helly graphs, in polynomial-time complexity.
Recall that a graph might have an exponential number of bicliques [15]. Therefore,
Partially supported by UBACyT Grants X184, X127, X212, PICT ANPCyT Grant
11-09112 and PID Conicet Grant, Argentina.
Partially supported by the Conselho Nacional de Desenvolvimento Cientı́fico e
Tecnológico, CNPq, and Fundação de Amparo à Pesquisa do Estado do Rio de Janeiro,
FAPERJ, Brasil.
634 M. Groshaus, J.L. Szwarcfiter
the algorithm by Berge (see [4], [5], [19]) for recognizing Helly families would not
recognize biclique-Helly graphs within polynomial time. Finally, we relate the class
of biclique-Helly graphs to those of clique-Helly graphs, disk-Helly graphs and
neighborhood-Helly graphs.
Denote by G an undirected graph, V (G) and E(G) its vertex and edge sets,
respectively. All graphs here considered are finite. A vertex v ∈ V (G) is universal
when it is adjacent to every other vertex of G. Denote by N (v) the set of neighbors of
v ∈ V (G). Denote by P3 an induced path with three vertices, and by P3 its comple-
ment. Write Ck for an induced cycle having k vertices. Say that a vertex v dominates (incidente)
an edge e, when one of the extremes of e either coincides or is adjacent to v. When v
dominates every edge of G then v is an edge dominator of G. For S ⊆ V (G), denote
by G[S] the subgraph induced in G by S. Say that a subgraph H of G is isometric if
for any pair of vertices x, y of H , the distance between x, y in H is the same as in G.
A clique is a maximal subset of vertices inducing a complete subgraph. The disk
Dl (v) of v ∈ V (G) is the subset of vertices of G whose distance to v is less or equal to
l. A biclique is a maximal subset B ⊆ V (G) inducing a complete bipartite subgraph
in G. Write B = X ∪ Y for the corresponding bipartition, restricting to X, Y = ∅.
Let A be a family of subgraphs of a graph G. Denote by GA = ∪H ∈A H the
subgraph formed by the vertices and edges that are contained in the graphs of A.
On the other hand, denote by G[A] the subgraph of G induced by ∪H ∈A V (H ).
We employ the following generalization of the concept of extended triangles [8,
17]. Let S ⊆ V (G), |S| = 3. Denote by BS the family of bicliques of G that contain
at least two vertices of S.
The induced subgraph G[BS ] is called the extension of S. Clearly, GBS is a span-
ning subgraph of G[BS ].
Let F be a family of subsets of some set. Say that F is intersecting when the sub-
sets of F pairwise intersect. On the other hand, when every intersecting subfamily
of F has a common element then F is a Helly family. A graph G is biclique-Helly
(clique-Helly, neighborhood-Helly, disk-Helly) when its family of bicliques (cliques,
neighborhoods, disks) is Helly.
The two characterizations of biclique-Helly graphs are formulated in Sections
2 and 3, together with the corresponding recognition algorithms. The relations
between biclique-Helly graphs and the other mentioned classes are described in
Section 3. In particular, in order to relate biclique-Helly and (open) neighborhood-
Helly graphs, we employ a characterization of the latter class, which is also proved
in Section 3. All the proposed characterizations are based on extensions.
Lemma 1. Let G be a graph. Then G has neither triangles nor C5 ’s if and only if each
of the extensions in G is a bipartite graph.
Biclique-Helly Graphs 635
Fig. 1. Lemma 1
Proof. Let us suppose that S = {v0 , v1 , v2 } is a subset of V (G) and examine G[BS ].
If G[BS ] is empty, there is nothing to prove. Otherwise G[BS ] contains at least two
vertices of S. First, discuss the case when there are only two vertices v1 , v2 ∈ S in
G[BS ]. Suppose v1 , v2 are adjacent. Let X, Y be the subsets of neighbors and non
neighbors of v1 in G[BS ], respectively. By definition of G[BS ], any vertex of Y must
be adjacent to v2 . Because G has no triangles, both X, Y must be independent sets,
meaning that G[BS ] is a bipartite graph. The second alternative is to assume that
v1 , v2 are non adjacent. In this case, define X as to be the set of common neighbors
of v1 and v2 in G[BS ], while Y is the subset formed by the remaining vertices of this
graph. By a similar argument as above, we conclude that X ∪ Y is a bipartition of
G[BS ], as required.
We examine the situation where all vertices of S are in G[BS ]. The proof is by
way of contradiction. Let us suppose that C is an odd cycle of length ≥ 7 of G[BS ],
formed by the vertices x0 , x1 , . . . , x2k . Moreover, suppose that there is no odd cycle
of length < |V (C)|. This implies that C is isometric. For each xi , let Bi be the bic-
lique of BS that contains xi . Note that | Bi ∩ S |≥ 2 for any i. Thus, for each pair
xi , xi of vertices of C there exists at least one vertex of S, denoted by vi,i , that
belongs to both Bi and Bi . Moreover, if xi xi ∈ E(C) then vi,i is adjacent precisely
to one of the vertices xi or xi , otherwise Bi ∪ Bi (and hence G itself) contains a
triangle or a C5 . Thus, each edge of C is dominated by some vertex of S. Without
loss of generality, suppose that v0 dominates more than two edges of C. Since C
is isometric, this implies that v0 is adjacent to exactly two vertices of C, those two
vertices being at distance 2 in C. Hence there exists a cycle C of the same size as C,
that contains v0 . Without loss of generality, suppose C = C and v0 = x0 . It then
follows from the isometry of C, that v3,4 = v0 . Denote v3,4 = v1 .
From what precedes, it is easy to see that v2 is adjacent either to x3 or to x4 . In
the first case v2 and v1 are not adjacent and in the second case they are. In both
cases, v0 is therefore at distance ≥ 2of v2 and of v3 . Let us now consider those two
cases.
Case 1. v2 is adjacent to x3 :
Hence v2 is not adjacent to v1 . Consider B2k and note that x2k is at distance ≥ 2
of both v1 and v2 and at distance 1 of v0 . Since in this case S is an independent
set, the vertices of S ∩ B2k must belong to the same part of the bipartition of B2k .
This implies that v0 ∈ / V (B2k ) and that both v1 and v2 are at distance 2 from x2k .
Now, without loss of generality, suppose v1 ∈ B1 . Observe that if x1 v1 ∈ E(G) then
636 M. Groshaus, J.L. Szwarcfiter
the vertices x2k , x0 , x1 , v1 , together with some vertex of B2k adjacent to x2k and v1 ,
either contain a triangle or form an induced C5 in G. Thus x1 v1 ∈ E(G). But in this
case, x1 , x2 , x3 , v1 , together with a vertex of B2k adjacent to v1 and x1 , leads to a
similar contradiction, as before.
Case 2. v2 is adjacent to x4 :
Hence v2 is also adjacent to v1 . Again consider B2k . We claim that v2 is adjacent
to x2k . Otherwise, both v1 and v2 are at distance ≥ 2 from x2k . It therefore follows
from v1 v2 ∈ E(G), that they can not both belong to B2k . Thus v0 ∈ V (B2k ), which in
turn implies that the other vertex vj of S, that belongs to B2k , being not adjacent to
v0 , is adjacent to x2k . Since we suppose that the claim is false, we must have vj = v1 .
This implies that the vertices x0 , x1 , x2 , x3 , v1 either form a C5 or contain a triangle,
a contradiction. Thus, v2 is adjacent to x2k . Now consider B1 . By a symmetrical
argument, as in the preceding claim, we can show that v1 is adjacent to x1 . Thus,
v0 , x1 , v1 , v2 , x2k induce a C5 , a contradiction. The proof is complete.
Theorem 1 [4, 5]. Let F be a family of subsets Si of some set U. Then F is Helly if and
only if for every triple T of vertices of U, the intersection of the subsets Si satisfying
|Si ∩ T | ≥ 2 is non empty.
Given S = {va , vb , vc } ⊆ V (G), the algorithm below constructs the vertex set of
the extension G[BS ]. For simplicity, we write V ∗ with the meaning of V (G[BS ]).
V ∗ := V ∗ ∪ [Ni ⊕ Nj ],
otherwise compute
V ∗ := V ∗ ∪ Nij ∪ [N (Nij ) ∩ N ij ]
The correctness of the algorithm follows from the above observations (i) and (ii).
All the computations of the algorithm can be easily performed in O(|V (G)|) time,
except that of finding N (Nij ), which requires O|(E(G)|) time. Consequently, each
extension of G[BS ] can be constructed in O(|E(G)|) time, i.e. the overall complexity
of the algorithm is O(|(V (G)|3 |(E(G)|).
The reason why the above assertion is true is simple. Suppose uj ∈ U is adjacent
to ui ∈ U in H (G). Then vi , vj are not adjacent in G. The latter implies wi , uj not
adjacent in H (G). Similarly, wj ∈ W adjacent to ui implies wj not adjacent to wi .
Consequently, Observation 1 is true.
Denote by C a family of intersecting bichromatic cliques of H (G). Let B be the
family of bicliques of G, corresponding to C. Observe that this is a 1-1 correspon-
dence, because the pair of cliques of H (G) which correspond to the same biclique
of G are disjoint. Since G is biclique-Helly, the bicliques of G contain a common
vertex vi ∈ V (G). Let Bj be the biclique of B corresponding to the bichromatic
clique Cj ∈ C. Because vi ∈ B1 , it follows ui ∈ C1 or wi ∈ C1 . Without loss of
generality, assume ui ∈ C1 . We show that ui ∈ Cj , also for clique Cj ∈ C, j = 1.
Suppose the contrary and let ui ∈ Cj . Because vi ∈ Bj , it follows wi ∈ Cj . If
wi ∈ C1 then ui and wi must be adjacent, which can not occur. Then ui ∈ C1 \ Cj
and wi ∈ Cj \ C1 . Because C1 and Cj intersect, there must be some vertex of H (G)
belonging to C1 ∩ Cj . Such a vertex would have to be simultaneously adjacent to ui
and wi , which contradicts Observation 1. Consequently, Cj contains ui , implying
that H (G) is indeed bichromatic-Helly.
Conversely, by hypothesis Conditions 1 and 2 are true for some graph G. We
show that G is biclique-Helly. Let B be an intersecting family of bicliques of G. We
construct a family C of bichromatic cliques of H (G). Each Bi ∈ B corresponds in
H (G) to a pair of disjoint bichromatic cliques Ci and Ci . We choose Ci ∈ C as to
be Ci or Ci , according to the following rule. Arbitrarily, choose C1 = C1 . Note that
since B1 and Bi intersect, C1 must intersect Ci or Ci . Then, for i > 1, if C1 and
Ci intersect then choose Ci = Ci , and otherwise Ci = Ci . First, we show that C,
as above obtained, is an intersecting family. Let Ci , Cj ∈ C. If i = 1 then C1 inter-
sects any Cj , by construction. Let i, j = 1. Since Bi and Bj intersect there exists
vs ∈ Bi ∩ Bj . Then Ci contains us or ws . Without loss of generality, let us ∈ Ci .
We will show that us ∈ Cj , meaning that Ci and Cj intersect. Suppose the contrary,
us ∈ Cj . Then ws ∈ Cj . Consider the following alternatives for locating C1 ∩ Ci
and C1 ∩ Cj .
Case 1. C1 ∩ Ci ∩ U = ∅ and C1 ∩ Cj ∩ U = ∅
Let ui ∈ C1 ∩ Ci ∩ U and uj ∈ C1 ∩ Cj ∩ U . If ui = us then C1 also contains
us , implying that uj is adjacent to us , because us , uj ∈ C1 . Since ws ∈ Cj , uj is
also adjacent to ws , contradicting Observation 1. Consequently, ui = us . Similarly,
uj = us . Then ui , uj , us are distinct. Since us , ui ∈ Ci , vs and vi are not adjacent
in G. On the other hand, because ws , uj ∈ Cj , vs and vj are adjacent. Finally, vi
and vj are not adjacent, because ui , uj ∈ C1 . Since vi , vj ∈ B1 , there exists vp ∈ B1
adjacent to both vi , vj . If vp , vs are adjacent then vp , vj , vs form a triangle, which
contradicts Condition 1. Then vp , vs are not adjacent. Because vi , vs ∈ Bi , there
exists a vertex vq ∈ Bi , simultaneously adjacent to vi and vs . We conclude that vq is
not adjacent neither to vp nor vj , otherwise G would contain a triangle. However
in this situation, the vertices vi , vj , vs , vp , vq induce a C5 in G, not possible.
Case 2. C1 ∩ Ci ∩ U = ∅ and C1 ∩ Cj ∩ W = ∅
Let ui ∈ C1 ∩ Ci ∩ U and wj ∈ C1 ∩ Cj ∩ W . Similarly as in Case 1, we know that
vi and vs are not adjacent. Because ws , wj ∈ Cj , vs and vj are also not adjacent, and
Biclique-Helly Graphs 641
Proof. Biclique-Helly graphs do not contain triangles and so they are clique-Helly.
Proof. First, we will prove that if G is disk-Helly and has no triangles, then it is
C5 − f ree. By contrary assume that the vertices v1 , . . . , v5 form a C5 . Consider
the disks D1 (v2 ), D1 (v3 ) and D1 (v5 ). Since these disks are a Helly family, there is a
vertex v adjacent to v2 , v3 , v5 . Then v forms a triangle with v2 and v3 , contradicting
the hypothesis. Now we will prove that the family of bicliques of G is Helly. Let
B = {B1 , . . . , Bk } be an intersecting family of bicliques. For each vertex v of each
biclique Bj of B, consider the disk D2 (v). Observe that if v ∈ Bj , the vertices of
Bj are included in the set of vertices of D2 (v). Consequently, the family of disks
{D2 (v) | v ∈ ∪Bj } is also intersecting. As G is disk-Helly, there is a vertex z which
belongs to every D2 (v). We are going to prove that z belongs to every biclique Bi of
642 M. Groshaus, J.L. Szwarcfiter
Proof. Let S = {v1 , v2 , v3 } be a set of pairwise non adjacent vertices and G[BS ]
its extension. Consider V2 (S). If V2 (S) = ∅, there is nothing to show. Otherwise,
it is clear that the family of neighborhoods of vertices of V2 (S) is intersecting. By
hypothesis, there is a vertex w which belongs to every neighborhood of the fam-
ily. Moreover, since G is triangle-free, for any s ∈ V2 (S), the vertices vertices w, s,
together with the two vertices of S adjacent to s are included in some biclique of
BS . Thus, w ∈ V (G[BS ]), and is adjacent to all vertices of V2 (S), as required.
Conversely, let G be a graph satisfying the hypothesis. By way of contradic-
tion, assume that G is not neighborhood-Helly. For vi ∈ V (G), denote by Ni the
neighborhood of vi . Let N = {N1 , N2 ,...,Nl }, l ≥ 3, be a minimal subfamily of
neighborhoods of G which is not Helly. Then for each i = 1, ..., l, N \ {Ni } is a Helly
Biclique-Helly Graphs 643
A B C D
Fig. 2. Graphs A, B, C, D
family, and therefore has a non empty intersection. Fix wi ∈ ∩j =i Nj . We claim that
vi = wj , 1 ≤ i, j ≤ l. Clearly, if i = j the claim holds because wi ∈ Nj and vj ∈ / Nj .
Examine the alternative, vi = wi . It implies that vi is adjacent to vk for any k = i.
Moreover, such a k must exist because l ≥ 3. Since Ni and Nk intersect, there exists
a vertex w which forms a triangle with vi and vk , a contradiction. Consequently,
wi = vj . Furthermore, wi , wj must be distinct, for i = j . Finally, we assert that
wi , wj are non adjacent. Otherwise, if wi , wj are adjacent, consider any vk , k = i, j .
By the above claim, vk = wi , wj . Since wi , wj ∈ Nk , it follows that wi , wj , vk form
a triangle, a contradiction. Then S = {w1 , w2 , w3 } is an independent set. Consider
the extension G[BS ]. Observe that every Ni contains at least two of the vertices of
S. Thus, vi is adjacent to at least two vertices of S and then belongs to V2 (S). By
hypothesis, there is a vertex w adjacent to every vertex of V2 (S). Then, w belongs
to Ni for 1 ≤ i ≤ l, a contradiction.
induce the graph A which does not extend to any of the graphs B, C or D, what
leads to a contradiction. Then {N (v), v ∈ X} is an intersecting family. As G is neigh-
borhood-Helly, there is a vertex v adjacent to every vertex w ∈ X, v ∈ V (G[BS ]).
Then, v is an edge dominator in G[BS ]. By Theorem 2, G is biclique-Helly.
Examine the case where S induces the complement of a P3 in G. Similarly as
in Lemma 1, by applying the definition of G[BS ], we conclude that G contains a
triangle or a C5 , contradicting the hypothesis. Finally, the case where S induces a
triangle also does not occur, completing the proof.
Proof. By Theorem 5, we only have to prove that for every independent set of three
vertices S = {v1 , v2 , v3 }, its extension G[BS ] has a vertex v adjacent to V2 (S). Recall
that as G is biclique Helly, by Lemma 1, G[BS ] is a bipartite graph. Let X, Y be a
bipartition of vertices of G[BS ], S ⊆ X. Suppose there is no edge dominator in X. By
Theorem 2, since G is biclique-Helly, G[BS ] has an edge dominator w. Then w ∈ Y .
As none of vertices vi , i = 1, 2, 3 are edge dominators, there exist distinct vertices
w1 , w2 , w3 ∈ Y with vi adjacent to wj precisely when i = j , for i, j = 1, 2, 3.
Then, the vertices v1 , w3 , v2 , w1 , v3 , w2 , w induce the graph C of Figure 2, absurd.
Consequently, G[BS ] has an edge dominator in X.
Acknowledgements. To the referee for the careful reading and for providing valuable sugges-
tions which improved the paper. In particular, he suggested the present version of the proof
of Lemma 1, which is shorter and more elegant than the authors’ previous proof.
References
1. Bandelt, H.-J., Farber, M., Hell, P.: Absolute reflexive retracts and absolute bipartite
graphs. Discrete Applied Mathematics 44, 9–20 (1993)
2. Bandelt, H.-J., Pesch, E.: Dismantling absolute retracts of reflexive graphs. European J.
Combinatorics 10, 211–220 (1989)
3. Bandelt, H.-J., Prisner, E.: Clique graphs and Helly graphs. J. Combin Theory B 51,
34–45 (1991)
4. Berge, C.: Hypergraphs. North Holland Mathematical Library, vol.45, Elsevier Science
Publishers B.V., Amsterdam, 1989
5. Berge, C., Duchet, P.: A generalization of Gilmore’s theorem. In: Fiedler, M. (ed.) Recent
Advances in Graph Theory. Acad. Praha, Prague pp 49–55 (1975)
6. Bondy, A., Durán, G., Lin, M., Szwarcfiter, J.: Self-clique graphs and matrix permuta-
tions. Journal of Graph Theory 44, 178–192 (2003)
7. Burzyn, P., Bonomo, F., Durán, G.: NP-completeness results for edge modification prob-
lems. Discrete Applied Mathematics 154(13), 1824–1844 (2006)
8. Dragan, F.F.: Centers of Graphs and the Helly property. PhD thesis, Moldava State
University, Chisinau, Moldava (1989) (In russian)
9. Escalante, F.: Über iterierte Clique-Graphen. Abhandlungender Mathematischen Sem-
inar der Universität Hamburg 39, 59–68 (1973)
10. Hamelink, B.C.: A partial characterization of clique graphs. J. Combin Theory 5, 192–
197 (1968)
11. Pavol Hell, Personal Communication (2003)
Biclique-Helly Graphs 645
12. Larrión, F., Neumann-Lara, V., Pizaña, M.A., Porter, T.D.: A hierarchy of self-clique
graphs. Discrete Mathematics 282, 193–208 (2004)
13. Müller, H.: On edge perfectness and classes of bipartite graphs. Discrete Math. 149,
159–187 (1996)
14. Peeters, R.: The maximum edge biclique problem is NP-complete. Discrete Appl. Math.
131, 651–654 (2003)
15. Prisner, E.: Bicliques in graphs I. Bounds on their number. Combinatorica 20, 109–117
(2000)
16. Prisner, E.: Bicliques in graphs II. Recognizing k-path graphs and underlying graphs of
line digraphs. Graph Theoretic Concepts in Computer Science, Lecture Notes in Com-
puter Science 1335, 253–287 (1997)
17. Szwarcfiter, J.L.: Recognizing clique-Helly graphs. Ars Combinatoria 45, 29–32 (1997)
18. Tuza, Z.: Covering of graphs by complete bipartite subgraphs: complexity of 0-1 matrices.
Combinatorica 4, 111–116 (1984)
19. Robert, F.S., Spencer, J.H.: A characterization of clique graphs. J. Combin. Theory B 10,
102–108 (1971)