Acessibilidade / Reportar erro

On the convex hull of 3-cycles of the complete graph

Abstract

Let Kn be the complete undirected graph with n vertices. A 3-cycle is a cycle consisting of 3 edges. The 3-cycle polytope is defined as the convex hull of the incidence vectors of all 3-cycles in Kn. In this paper, we present a polyhedral analysis of the 3-cycle polytope. In particular, we give several classes of facet defining inequalities of this polytope and we prove that the separation problem associated to one of these classes of inequalities is NP-complete. Finally, it is proved that the 3-cycle polytope is a 2-neighborly polytope.

polytope; cycle; facet; NP-completeness


On the convex hull of 3-cycles of the complete graph

Michel KovalevI; Jean-François MaurrasII, * * Corresponding author / autor para quem as correspondências devem ser encaminhadas ; Yann VaxèsIII

IFaculty of Applied Mathematics and Informatics University of Belarus Minsk – Belarus

IILaboratoire d'informatique fondamentale de Marseille UMR 6166 CNRS Marseille Cedex 9 – France. jean-francois.maurras@lim.univ-mrs.fr

IIILaboratoire d'informatique fondamentale de Marseille UMR 6166 CNRS Marseille Cedex 9 – France

ABSTRACT

Let Kn be the complete undirected graph with n vertices. A 3-cycle is a cycle consisting of 3 edges. The 3-cycle polytope is defined as the convex hull of the incidence vectors of all 3-cycles in Kn. In this paper, we present a polyhedral analysis of the 3-cycle polytope. In particular, we give several classes of facet defining inequalities of this polytope and we prove that the separation problem associated to one of these classes of inequalities is NP-complete. Finally, it is proved that the 3-cycle polytope is a 2-neighborly polytope.

Keywords: polytope; cycle; facet; NP-completeness.

1. Introduction

A 3-cycle is a cycle with three edges. Consider the following minimum weighted 3-cycle problem: given a graph G=(V,E) and a 'weight' function w:E ® find a 3-cycle C of G such that w(C) is as small as possible. This problem can easily be solved in polynomial time by complete enumeration of the triangles G.

Let P(G) be the polytope defined as the convex hull of the incidence vectors of the 3-cycles of G, that is

P(G)= conv.hull {cC Î {0,1}:C is a 3-cycle of G}.

The minimum weighted 3-cycle problem is clearly equivalent to the linear program

max{wx:x Î P(G)},

as every minimum weighted 3-cycle yields an optimal vertex solution of the linear program and vice versa. Since the minimum weighted 3-cycle problem is solvable in polynomial time, it follows from the work of Grötschel, Lovász & Schrijver (1981, 1993) that there exists a polynomial time algorithm that solves the following problem:

Separation problem (SEP): given a graph G=(V ,E) and a vector y Î decide whether y belongs to P(G) or not, in the later case, find a vector a Î such that ax < ay for all x Î P(G).

This algorithm for problem SEP provides an implicit description for P(G). Motivated by the existence of an implicit description for P(G), we attempt to find an explicit description of P(Kn) by a minimal system of linear inequalities. In this paper, we present several classes of facet-defining linear inequalities for P(Kn), we prove that it is NP-hard to solve the separation problem for one of these classes, we show that the diameter of P(Kn) is one. Unfortunately, we did not succeed in our pursuit for a complete description of P(Kn) by a reasonable number of classes of linear inequalities. Using a computer we were able to verify that the facet-defining inequalities presented provide a complete description for P(K6) (70 facets) and P(K7) (896 facets). See Barahona & Grötschel (1986), Coullard & Pulleyblank (1989) and Seymour (1979) for related studies concerning other cycle polytopes.

Let us introduce some definitions and notations. For a cycle C, define its incidence vector cC Î by letting = 1 if e Î C and 0 otherwise. Throughout this paper, we will confuse a cycle C with its incidence vector, e.g. we will say that a cycle C satisfies an inequality. Let G=(V ,E) be an undirected graph. For any two adjacent vertices u and v, denote by uv the edge between u and v. A cycle C of G will be viewed as a set of edges but denoted by an ordered list of vertices; e.g. (v1,v2,v 3,v4) denotes the cycles containing edges v1v2,v2v3,v3v4,v4v1. A 3-cycle is a simple cycle of length 3. For two subsets U and W of V, we define the subset of edges (U:W) as follows

(U:W) := {uw Î E : u Î U and w Î W}

and d(U):= (U:V-U). For a subset X of vertices, let E(X) be the set of edges in uv with u,v Î X, and vice versa, for a subset F of edges, let V(F) be the set of end-vertices of edges in F. A cycle C is called tight with respect to an inequality ax < b if acC= b. Finally, for a given subset of edges F and a given vector x Î , we adopt the following notation x(F) := xe.

In the next section, we present a few basic properties of P(Kn) and we establish an auxiliary lemma which will be used several times in the rest of the paper for proving that an inequality defines a facet of P(Kn). In Section 3, we provide a complete description of P(Kn) for n < 6 employing three classes of facet defining inequalities. Then, three new classes of facet defining inequalities are introduced. Altogether, they allows to describe completely P(K7) We prove that it is NP-hard to solve the separation problem for one of these classes. Next, we present a class of facet defining inequalities that generalizes four classes introduced before and give an additional classes of facets for P(Kn) with n > 9. Finally, in Section 4 we prove that P(Kn) is a 2-neighborly polytope for all n > 4.

2. Basic results

Let us start with some observations which will be useful later.

Lemma 1. If all 3-cycles of a K4 induced by the subset of vertices {u,v,w,t} Í V satisfy an equality ax = b then

auv = awt = a1,

avw = aut = a2,

auw = avt = a3,

auv +avw + auw = b.

Proof. Let us consider all 3-cycles of a K4

auv + avw + awu = b,

avw + awt + atv= b,

auw = awt = atu = b,

auv +avt + atu = b.

Summing up any two of these equalities and subtracting the two others, we get

auv -awt = 0,

auw - avw = 0,

aut - avw = 0.

Lemma 2.If all 3-cycles of a K5 induced by a subset of vertices S Í V satisfy an equality ax = b, then auv = b/3 for all u,v Î S.

Proof. Applying Lemma 1 to all 4-cliques defined on S, we deduce auv = b/3 for all u,v Î S

Proposition 1. For n > 5, {x Î : x(E) = 3} is the affine hull of P(Kn).

Proof. Suppose that all 3-cycles of Kn satisfy an equality ax = b. By scaling, we may assume that b = 3 and by Lemma 2 ax = b is precisely x(E) = 3.

Remark 1. For n > 5 the dimension of P(Kn) is -1. For n = 5 this dimension is 9. The incidence vectors of the ten 3-cycles of K5 are linearly independent. The polytope P(K5) is a 9-dimensional simplex which is defined x(E) = 3 and

Moreover, these inequalities define facets of P(K5) Indeed, nine of the ten 3-cycles of K5 are tight with respect to a given inequality from (1).

3. Facet defining inequalities

In the rest of the paper, in order to prove that a valid inequality I defines a facet of P(Kn), we proceed as follows. Consider the linear variety defined by x(E) = 3 and I, if the set of 3-cycles that are tight with respect to I does not span this variety, then they belong to a proper subvariety, i.e. they satisfy another equality J º ax = b such that I, J and x(E) = 3 are independent. By adding an appropriate linear combination of x(E) = 3 and I to J we can fix two coefficients of J to 0. Finally, using the fact that all tight 3-cycles with respect to I satisfy J we derive that ae = b = 0 for all e Î E.

Proposition 2. For each edge uv Î E, the linear inequality

defines a facet of P(Kn) whenever n > 6.

Proof. Suppose that all tight 3-cycles with respect to (2) (that is, all 3-cycles not containing the edge uv) satisfy an inequality ax = b. Applying Lemma 2 to all K5 not containing the edge uv we deduce that ae = b/3 for all e Î E - {uv}. Finally, fixing b = 0 and auv = 0 we get ae = b = 0 for all e Î E. 

Lemma 3. Let u,v Î V and n > 6 if all tight with respect to

3-cycles of Kn satisfy ax = b then

Proof. The 3-cycles of Kn-1 not containing u are tight with respect to (3), thus they satisfy ax = b. Using Lemma 2 we derive ae = b/3 = a3 for all e Î E -d(u). Now, all 3-cycles (u,v,w) with w Î V - {u,v} are tight with respect to (3) yielding auw = 2b/3 - auv for all w Î V - {u,v}. 

Proposition 3. For each edge uv Î E the inequality (3) defines a facet of Kn whenever n > 6.

Proof. First apply Lemma 3, then fix two coefficients b = a1 = 0, yielding a2 = a3 = 0. 

The set of all integer solutions of the system x(E) = 3, (2) and (3) is exactly the set of all 3-cycles of Kn.

Proposition 4. For each subset X Í V such that 2 < |X| < |V|/2, the inequality

defines a facet of P(Kn) whenever n> 6.

Proof. Let us suppose that all tight 3-cycles with respect (4) belong to a proper subvariety defined by x(E) = 3, (4) and ax = b. Note that all 3-cycles of a K4 containing two vertices u,v, Î X and two other u,v Î V - X are tight with respect to (4). Applying Lemma 1 to these K4 we obtain

with a1 + 2a2 = b. By fixing a1= a2 = 0 we get b = 0 and ae = 0 for all edge e Î E. 

Using a computer code, we have been able to enumerate all facets of P(K6) This polytope has 70 facets and is completely defined by inequalities (2), (3), (4) and x(E) = 3.

Proposition 5. Let G = (V ,E) be a graph and let

The separation problem for Q(G) is NP-complete.

Proof. We provide a polynomial reduction from the problem MAXCUT which is proved to be NP-hard (Garey, Johnson & Stockmeyer, 1976). Its formulation follows. Given an undirected graph H = (V ,F) and a positive integer k, find a subset of vertices X Í V such that |d(X)|>k. One can transform an instance of the MAXCUT problem in an instance of the separation problem for Q(Kn) as follows. Suppose without loss of generality. that no vertex of H has a degree larger than k (otherwise one can find a cut of cardinality larger than k in linear time). Then, consider a real valued vector x Î defined as follows

Clearly, there is a subset X Í V such that (4) separates x from Q(Kn) if and only if there is a cut of cardinality larger than k in H. This concludes the proof of Proposition 5. 

Proposition 6. For each subset of four vertices {u,v,w,t} Í V, the inequality

defines a facet of P(Kn) whenever n > 7.

Proof. Consider the complete subgraph Kn-1 which does not contain the vertex v. Since it has at least 6 vertices, as in the proof of Proposition 3, we can show that ae = 0 for each edge e of this subgraph, and thus b = 0. Analogously, one can show that the same equality holds for all edges of the subgraph Kn-1 which does not contain w. It remains to fix auw. Note that b = 0 and consider one of the two 3-cycles containing the edge vw and which is tight with respect to (5), namely (u,v,w) or (u,t,w). We obtain auw = 0. 

Proposition 7. For each subset of three vertices {u,v,w} Í V the inequality

defines a facet of Kn whenever n > 7.

Proof. The proof is similar to that of Proposition 3. First, consider the complete subgraph Kn-1 which does not contain the vertex v and then the one which does not contain w. Finally, consider the 3-cycle (u,v,w) which is tight with respect to inequality (6) and contains the edge vw. 

Proposition 8. For a pair of vertices {u,v} Í V, and each simple cycle C containing all vertices of V-{u,v} the inequality

defines a facet of P(K7).

Proof. Consider a K4 induced by u,v and any two non consecutive vertices w and t of the cycle C. Using Lemma 1 we derive awt = auv = a1 and auw = avt = a2 for each v,W Î C. It remains to fix the coefficients of the edges of the cycle C. Let us consider a 3-cycle (v,w,t) which contains only one edge wt of C. This 3-cycle is tight with respect to (7) implying awt = b - 2a1 Finally, we fix a1 = a2 = 0. and by considering a tight 3-cycle (u,v,w), we deduce b = 0 and ae = 0 for each edge e Î C. 

Using a computer code, we have been able to enumerate all 896 facets of P(K7). This polytope is completely defined by inequalities (2)-(7) and equality x(E) = 3. Note that six classes of inequalities are necessary to describe completely P(K7). Note that, for n > 8, the inequality (7) is not valid since it is violated by any 3-cycle consisting of vertices of C and not containing any edge of C.

Now, we present a class of facet defining inequalities that generalizes the classes (3), (5) and (6). Given a positive integer k, a list Si,Ti,(i=1,...,k) of disjoint subsets of V, we define the following subsets of edges

Consider the following inequality

Proposition 9. If n > 5, |H| + > 5, and at least one of the following conditions holds:

1.k > 2

2.|S1|=1

3.|T1| > 2

then (8) defines a facet of P(Kn).

Proof. We distinguish two cases.

Case 1: k> 2.

Consider one after another all K4 obtained by picking a vertex in each subset Si,Ti,Sj for i,j=1,...,k, i ¹ j. By applying Lemma 1 for each of these K4 we get the following equalities

Let {s1,...,sk} and {t1,...,tk} be two subsets of vertices such that sjÎTj and tjÎ Tj for all j=1,...,k. For i=1, ,k, define Pi := H È {si} È {t1,t2,...,tk}. By hypothesis |Pi| = |H|+k+1 > 6, hence by applying Lemma 3 to the complete subgraph induced by Pi we derive

Note that any 3-cycle with two vertices in Tiand one in Si (or the reverse) satisfy (8), therefore

Fixing two coefficients b = a1= 0, we get a2 = a3 = a4= 0. This concludes Case 1.

Case 2: k=1. We distinguish two subcases.

|S1|=1.

If |T1|=1 then the proof of Proposition 3 applies. Otherwise, let t1,q1 be two vertices in T1. We can apply Lemma 1 on the subgraphs induced respectively by the subsets of vertices P1= {s1} È {t1} È H and Q1= {s1} È {q1} È H and derive

Hence, a'2=a2, a'3=a3, and we deduce that a'1=a1. These equalities do not depend on the choice of the vertices t1, q1. Finally, note that all 3-cycles having two vertices in T1 and one in S1 satisfy (8), hence

By fixing b = a1= 0, we get a2=a3=a4=0.

|S1| > 2.

In this subcase, |T1|> 2 because one of the three conditions of the proposition must hold. Choose two vertices s1,s'1ÎS1. For each of them we can provide the same proof as in the case |S1|=1 and show that

We get a'3=a3 and a'4=a4, and deduce a'1=a1 and a'2=a2. Fixing b =a1=0, we concludes a2=a3=a4=0. 

Proposition 10. Let C and C' be two simple cycles covering all vertices of Kn and such that if uv and vw belongs to C, then uw belong to C'. The inequality

defines a facet of P(Kn) whenever n is odd and n > 9.

Proof. Let e=uv, e'=wtÎC be two edges such that the K4 induced by the subset of vertices {u,v,w,t} has only the edges uv and wt in common with C and C'. Every 3-cycles of this K4 are tight with respect to inequality (9). Using Lemma 1 for every such K4, we show that

Next, consider the 3-cycles (e,e',e'') with e,e' Î C and e'' Î C'. They are also tight with respect to inequality (9), yielding that ae''+2a2=b and ae''=a3. Hence, the following equalities holds

a1+2a2= b and a3+2a1= b

Finally, we fix a1=a2= 0, and concludes b = a3=0. 

4. Neighbourhood relation on P(Kn)

A polyhedron P is said to be k-neighborly if each k-subset S Í vert(P) defines a face F = conv(S) such that S = vert(F).

Proposition 11. P(Kn) is a 2-neighborly polytope whenever n > 4.

Proof. Given any two 3-cycles x=(v1v2v3) and x'=(v'1v'2v'3), the incidence vector of the subgraph obtained as the union of x' and x'' cannot be written as a convex linear combination of any other 3-cycles. Therefore, the intersection of conv(vert(P(Kn)-{x',x''})) and aff({x',x''}) is empty. In other words, conv({x',x''}) is a 1-face (an edge) of P(Kn).

Following Grünbaum (1967), we conclude that each 3-face of P(Kn) is a simplex, the diameter of P(Kn) is equal to 1, and the number of 1-faces of P(Kn) is equal to .

Furthermore, notice that for a linear program over P(Kn), the problem of finding the best neighbour of an extreme point is equivalent to the complete enumeration.

Acknowledgement

We would like to thank anonymous referees for many helpful comments.

Recebido em 04/2002, aceito em 11/2002 após 1 revisão

  • (1) Barahona, F. & Grötschel, M. (1986). On the cycle polytope of a binary matroid. Journal of Combinatorial Theory Ser. B, 40-62.
  • (2) Coullard, C.R. & Pulleyblank, W.R. (1989). On cycles cones and polyhedra. Linear Algebra and Its Applications, 114/115, 613-640.
  • (3) Garey, M.R.; Johnson, D.S. & Stockmeyer, L. (1976). Some Simplified NP-Complete Problems. Theoretical Computer Science, 1, 237-267.
  • (4) Grötschel, M.; Lovász, L. & Schrijver, A. (1981). The ellipsoid method and its consequences in combinatorial optimization. Combinatorica, 2, 169-197 [Corrigendum: Combinatorica (1984), 4, 291-295].
  • (5) Grötschel, M.; Lovász, L. & Schrijver, A. (1993). Geometric algorithms and combinatorial optimization. Algorithms and Combinatorics, 2, Springer-Verlag, Berlin.
  • (6) Grünbaum, B. (1967). Convex Polytopes Interscience Publisher, John Wiley & Sons.
  • (7) Seymour, P.D. (1979). Sums of circuits. In: Graph Theory and Related Topics [edited by J.A. Bondy and U.S.R. Murty], Academic Press, New York, 341-355.
  • *
    Corresponding author / autor para quem as correspondências devem ser encaminhadas
  • Publication Dates

    • Publication in this collection
      26 May 2003
    • Date of issue
      Jan 2003

    History

    • Accepted
      Nov 2002
    • Received
      Apr 2002
    Sociedade Brasileira de Pesquisa Operacional Rua Mayrink Veiga, 32 - sala 601 - Centro, 20090-050 Rio de Janeiro RJ - Brasil, Tel.: +55 21 2263-0499, Fax: +55 21 2263-0501 - Rio de Janeiro - RJ - Brazil
    E-mail: sobrapo@sobrapo.org.br