3eme Cours : Chordal Graphs MPRI 2011–2012
3eme Cours : Chordal GraphsMPRI 2011–2012
Michel [email protected]
http://www.liafa.jussieu.fr/~habib
Chevaleret, septembre 2011
3eme Cours : Chordal Graphs MPRI 2011–2012
Schedule
What we already know
More structural insights of chordal graphs
Properties of reduced clique graphs
Interval graphs
Exercises
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Fundamental objects to play with
Definition
A graph is chordal iff it has no chordless cycle of length ≥ 4.
Maximal Cliquesunder inclusion
Minimal SeparatorsA subset of vertices S is a minimal separator if Gif there exist a, b ∈ G with ab /∈ G , such that a andb are not connected in G − S .and S is minimal for inclusion with this property .
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
An example
a
b c ef
d3 minimal separators {b} for f and a, {c} for a and e and {b, c}
for a and d .
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
If G = (V ,E ) is connected then for every a, b ∈ V such thatab /∈ Ethen there exists at least one minimal separator.But there could be an exponential number of minimal separators.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
VIN : Maximal Clique trees
A maximal clique tree (clique tree for short) is a tree T thatsatisfies the following three conditions :
I Vertices of T are associated with the maximal cliques of G
I Edges of T correspond to minimal separators.
I For any vertex x ∈ G , the cliques containing x yield a subtreeof T .
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Helly Property
Definition
A subset family {Ti}i∈I satisfies Helly property if∀J ⊆ I et ∀i , j ∈ J Ti ∩ Tj 6= ∅ implies ∩i∈ITi 6= ∅
Exercise
Subtrees in a tree satisfy Helly property.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Demonstration.
Suppose not. Consider a family of subtrees that pairwise intersect.For each vertex x of the tree T , it exists at least one subtree of thefamily totally contained in one connected component of T − x .Else x would belong to the intersection of the family, contradictingthe hypothesis.Direct exactly one edge of T from x to this part.We obtain a directed graph G , which has exactly n vertices and ndirected edges. Since T is a tree, it contains no cycle, therefore itmust exist a pair of symmetric edges in G , which contradicts thepairwise intersection.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Back to chordal graphs
Chordal graph recognition
1. Apply a LexBFS on G O(n + m)
2. Check if the reverse ordering is a simplicial elimination schemeO(n + m)
3. In case of failure, exhibit a certificate : i.e. a cycle of length≥ 4, without a chord. O(n)
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Subtrees in a tree
Using results of Dirac 1961, Fulkerson, Gross 1965, Buneman1974, Gavril 1974 and Rose, Tarjan and Lueker 1976 :
For a connected graph, the following statements are equivalentand characterize chordal graphs :
(i) G has a simplicial elimination scheme
(ii) Every minimal separator is a clique
(iii) G admits a maximal clique tree.
(iv) G is the intersection graph of subtrees in a tree.
(v) Any MNS (LexBFS, LexDFS, MCS) provides asimplicial elimination scheme.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Two subtrees intersect iff they have at least one vertex in common.By no way, these representations can be uniquely defined !
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Proof of the chordal characterization theorem
I Clearly (iii) implies (iv)
I For the converse, each vertex of the tree corresponds to aclique in G .But the tree has to be pruned of all its unnecessary nodes,until in each node some subtree starts or ends. Then nodescorrespond to maximal cliques.
I We need now to relate these new conditions to chordal graphs.(iii) implies (i) since a maximal clique tree yields a simplicialelemination scheme.(iv) implies chordal since a cycle without a chord generates acycle in the tree.(iv) implies (ii) since each edge of the tree corresponds to aminimal separator which is a clique
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
from (i) to (iv)
Demonstration.
By induction on the number of vertices. Let x be a simplicialvertex of G .By induction G − x can be represented with a family of subtreeson a tree T .N(x) is a clique and using Helly property, the subtreescorresponding to N(x) have a vertex in common α.To represent G we just add a pending vertex β adjacent to α.x being represented by a path restricted to the vertex β, and weadd to all the subtrees corresponding to vertices in N(x) the edgeαβ.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Playing with the representation
Easy Exercises :
1. Find a minimum Coloring (resp. a clique of maximum size) ofa chordal graph in O(n + m).Consequences : chordal graphs are perfect.At most n maximal cliques.
2. Find a minimum Coloring (resp. a clique of maximum size) ofan interval graph in O(n)using the interval representation.
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Exercises
1. Can we use efficiently this representation of chordal graphs asintersection of subtrees ?
2. Same question for path graphs ? (intersection graph of pathsin a tree)
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Which kind of representation to look for for special classesof graphs ?
I Easy to manipulate (optimal encoding, easy algorithms foroptimisation problems)
I Geometric in a wide meaning (ex : permutation graphs =intersection of segments between two lines)
I Examples : disks in the plane, circular genomes . . .
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
First remark
Proposition
Every undirected graph can be obtained as the intersection of asubset family
Proof
G = (V ,E )Let us denote by Ex the set of edges adjacent to x .xy ∈ E iff Ex ∩ Ey 6= ∅We could also have taken the set Cx of all maximal cliques whichcontains x .Cx ∩ Cy 6= ∅ iff ∃ one maximal clique containing both x and y
3eme Cours : Chordal Graphs MPRI 2011–2012
What we already know
Starting from a graph in some application, find its characteristic :
1. 2-intervals on a line (biology), intersection of disks (orhexagons) in the plane (radio frequency), filament graphs,trapezoid graphs . . .
2. A whole book on this subject :J. Spinrad, Efficient Graph Representations, Fields InstituteMonographs, 2003.
3. A website on graph classes :http ://www.graphclasses.org/
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Clique treeclique tree of G = a minimum size tree model of G
for a clique tree T of G :I vertices of T correspond precisely to the maximal cliques of GI for every maximal cliques C ,C ′, each clique on the path in T
from C to C ′ contains C ∩ C ′
I for each edge CC ′ of T , the set C ∩ C ′ is a minimal separator(an inclusion-wise minimal set separating two vertices)
Note : we label each edge CC ′ of T with the set C ∩ C ′.
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Consequences of maximal clique tree
Theorem
Every minimal separator belongs to every maximal clique tree.
Lemma
Every minimal separator is the intersection of at least 2 maximalcliques of G .
Corollary
There are at most n minimal separators.
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Theorem
Every minimal separator belongs to every maximal clique tree.
Lemma
Every minimal separator is the intersection of at least 2 maximalcliques of G .
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Proof of the lemma
Demonstration.
We know that S is a clique. Let us consider G1 a connectedcomponent of G − S . Let x1, . . . , xk be the vertices of G1 having amaximal neighbourhood in S .If k = 1 then x1 must be universal to S , since S is a minimalseparator.Else, consider a shortest path µ in G1 from x1 to xk . Necessarily x1(resp. xk) has a private neighbour z (resp. t) in S .Then the cycle [x1, µ, xk , t, z ] has no chord, a contradiction.Therefore x1 ∪ S is a clique, and is contained in some maximalclique C in G1. We finish the proof by taking another connectedcomponent of G − S .
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Proof of the theorem
Demonstration.
Therefore S = C ′ ∩ C ′′. These two maximal cliques belong to anymaximal clique tree T of G . Let us consider the unique path µ inT joigning C ′ to C ′′.All the internal maximal cliques in µ must contain S . Suppose thatall the edges of µ are labelled with minimal separators strictlycontaining S , then we can construct a path in G from C ′ − S toC ′′ − S avoiding S , a contradiction. So at least one edge of µ islabelled with S .
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Clique graphthe clique graph C(G ) of G = intersection graph of maximalcliques of G
G C(G )
3eme Cours : Chordal Graphs MPRI 2011–2012
More structural insights of chordal graphs
Reduced clique graphthe reduced clique graph Cr (G ) of G = graph on maximal cliquesof G where CC ′ is an edge of Cr (G ) ⇐⇒ C ∩ C ′ is a minimalseparator.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Combinatorial structure of Cr (G )
Lemma 1 : M.H and C. Paul 95
If C1,C2,C3 is a cycle in Cr (G ), with S12,S23 and S23 be theassociated minimal separators then two of these three separatorsare equal and included in the third.
Lemma 2 : M.H. and C. Paul 95
Let C1,C2,C3 be 3 maximal cliques, ifC1 ∩ C2 = S12⊂S23 = C2 ∩ C3 then it yields a triangle in Cr (G )
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Lemma 3 : Equality case
Let C1,C2,C3 be 3 maximal cliques, if S12 = S23 then :
I either the C1 ∩ C3 = S13 is a minimal separator
I or the edges C1C2 and C2C3 cannot belong together to amaximal clique tree of G .
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
a
bc
d fe
g
h
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e
e c,e
c
c
ee
b,c c
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e c,e
ce
b,c
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e
e c,e
c
b,c
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Theorem (Gavril 87, Shibata 1988, Blayr and Payton 93)
The clique trees of G are precisely the maximum weight spanningtrees of C(G ) where the weight of an edge CC ′ is defined as|C ∩ C ′|.
Theorem (Galinier, Habib, Paul 1995)
The clique trees of G are precisely the maximum weight spanningtrees of Cr (G ) where the weight of an edge CC ′ is defined as|C ∩ C ′|.
Moreover, Cr (G ) is the union of all clique trees of G.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Applications
I All clique trees have exactly the same labels, includingrepetitions.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Maximal Cardinality Search : MCS
Data: a graph G = (V ,E ) and a start vertex s
Result: an ordering σ of V
Assign the label 0 to all verticeslabel(s)← 1for i ← n a 1 do
Pick an unumbered vertex v with largest labelσ(i)← vforeach unnumbered vertex w adjacent to v do
label(w)← label(w) + 1end
end
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Maximum spanning trees
Maximal Cardinality Search can be seen as Prim algorithm forcomputing a maximal spanning tree of Cr (G ).
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
I How to compute a clique tree ?
I How to generate all simplicial elimination schemes ?
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
LexBFS, MCS or MNS visit maximal cliques “consecutively” (i.e.when the search explores a vertex x of a maximal clique C thatdoes not belong to any of the previously visited maximal cliquesthen all the unvisited vertices of C will appear consecutively justafter x).
Therefore when applying a search (LexBFS, MCS or MNS) one cancompute a clique tree, by considering the strictly increasingsequences of labels.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
a
bc
d fe
g
h
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e
e c,e
c
c
ee
b,c c
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e c,e
ce
b,c
b,c,e
a,b,c
b,d,e e,g c,e,f
c,hb,e
e c,e
c
b,c
b, a, c, e, d , h, f , g is a LexBFS ordering.we can find the maximal cliques b, a, c then b, c , e then b, e, dthen c, h then c , e, f and e, g .
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Simplicial elimination schemes
1. Choose a maximal clique tree T
2. While T is not empty doSelect a vertex x ∈ F − S in a leaf F of T ;F ← F − x ;If F = S delete F ;
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Canonical simplicial elimination scheme
1. Choose a maximal clique tree T
2. While T is not empty doChoose a leaf F of T ;Select successively all vertices in F − Sdelete F ;
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Remark
Does there exist other simplicial elimination scheme ?
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Size of a maximal clique tree in a chordal graph
I Let G = (V ,E ) be a chordal graph.
I G admits at most |V | maximal cliques and therefore the treeis also bounded by |V | (vertices and edges).
I But some vertices can be repeated in the cliques. If weconsider a simplicial elimination ordering the size of a givenmaximal clique is bounded by the neighbourhood of the firstvertex of the maximal clique.
I Therefore any maximal clique tree is bounded by |V |+ |E |.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Size of CS(G )
Considering a star on n vertices,shows |CS(G )| ∈ O(n2)Not linear in the size of G
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
CS(G ) is not chordal !
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
CS(G ) is not chordal !
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
In fact CS(G ) is dually chordal (almost chordal)and CS(CS(G )) is chordal.
3eme Cours : Chordal Graphs MPRI 2011–2012
Properties of reduced clique graphs
Cannonical representation
I For an interval graph, its PQ-tree represents all its possiblemodels and can be taken as a cannonical representation of thegraph (for example for graph isomorphism)
I But even path graphs are isomorphism complete. Therefore acanonical tree representation is not obvious for chordal graphs.
I Cr (G ) is a Pretty Structure to study chordalgraphs.To prove structural properties of all maximal clique trees of agiven chordal graph.
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
Characterization Theorem for interval graphs
(i) G = (V ,E ) is interval graph.
(ii) It exists a total ordering τ of the vertices of V s.t.∀x , y , z ∈ G with x ≤τ y ≤τ z and xz ∈ E thenxy ∈ E .
(iii) G has a maximal clique path. (A maximal clique pathis just a maximal clique tree T, reduced to a path).
(iv) G is the intersection graph of a family of intervals ofthe real line
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
To recognize an interval graph, we just have to compute amaximal clique tree and check if it is a path ?Difficulty : an interval graph has many clique trees among themsome are paths
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
Many linear time algorithms already proposed for interval graphrecognition ....using nice algorithmic tools :graph searches, modular decomposition, partition refinement,PQ-trees . . .
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
Linear time recognition algorithms for interval graphs
I Booth and Lueker 1976, using PQ-trees.
I Korte and Mohring 1981 using LexBFS and ModifiedPQ-trees.
I Hsu and Ma 1995, using modular decomposition and avariation on Maximal Cardinality Search.
I Corneil, Olariu and Stewart SODA 1998, using a series of 6consecutive LexBFS, published in 2010.
I M.H, McConnell, Paul and Viennot 2000, using LexBFS andpartition refinement on maximal cliques.
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
A partition refinement algorithm working on maximalcliques
1. Compute a tree T using LexBFIf T is not a maximal clique tree ; then G is not chordal,neither interval.
2. Start form the last maximal clique Refine the cliques with theminimal separator.
3. Refine until each part is a singleton
4. If a part is not a singleton start recursively from the last cliqueof this part according to LexBFS.
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
Variations on these representations
I Path graphs = Paths in a treeIn between chordal and interval.
I Directed path graphs = directed paths in a rooted directedtreeappear in some polynomial CSP class.
3eme Cours : Chordal Graphs MPRI 2011–2012
Interval graphs
State of the Art
I We (Juraj Stacho, M.H.) study chordal representations foralgorithms
I A linear time algorithm for path graph recognition is stillmissing
3eme Cours : Chordal Graphs MPRI 2011–2012
Exercises
Exercise 1
Ends of a LexBFS
Many properties can be expressed on the last vertex of a LexBFS.Example : if G is a chordal graph, the last vertex is simplicial .
1. Show that the last maximal clique visited can be taken as theend of some chain of cliques if G is an interval graph.
2. Complexity of the following decision problem :
Data: a graph G = (V ,E ) and a given vertex x ∈ V
Result: Does there exist a LexBFS of G ending at x ?