On Box-Perfect Graphs

Guoli Ding∗
Department of Mathematics, Louisiana State University, Baton Rouge, USA
arXiv:1608.04572v2 [math.CO] 17 Aug 2016

Wenan Zang† and Qiulan Zhao‡

Department of Mathematics, The University of Hong Kong, Hong Kong, China

June 17, 2021

Abstract

Let G = (V, E) be a graph and let AG be the clique-vertex incidence matrix of G. It is

well known that G is perfect iff the system AG x ≤ 1, x ≥ 0 is totally dual integral (TDI). In
1982, Cameron and Edmonds proposed to call G box-perfect if the system AG x ≤ 1, x ≥ 0
is box-totally dual integral (box-TDI), and posed the problem of characterizing such graphs.
In this paper we prove the Cameron-Edmonds conjecture on box-perfectness of parity graphs,
and identify several other classes of box-perfect graphs. We also develop a general and powerful
method for establishing box-perfectness.

1 Introduction

A rational system Ax ≤ b is called totally dual integral (TDI) if the minimum in the LP-duality
equation
max{w T x : Ax ≤ b} = min{y T b : y T A = wT ; y ≥ 0} (1.1)
has an integral optimal solution, for every integral vector w for which the minimum is finite.
Edmonds and Giles [14] proved that total dual integrality implies primal integrality: if Ax ≤ b is
TDI and b is integral, then both programs in (1.1) have integral optimal solutions whenever they
have finite optimum. So the model of TDI systems serves as a general framework for establishing
min-max results in combinatorial optimization (see Schrijver [22] for an comprehensive and in-depth
account). As summarized by Schrijver [20], the importance of a min-max relation is twofold: first, it
serves as an optimality criterion and as a good characterization for the corresponding optimization
problem; second, a min-max relation frequently yields an elegant combinatorial theorem, and allows
a geometrical representation of the corresponding problem in terms of a polyhedron. Many well-
known results and difficult conjectures in combinatorial optimization can be rephrased as saying
that a certain linear system is TDI; in particular, by Lovász’ Replication Lemma [16], a graph G is
∗
Supported in part by NSF grant DMS-1500699.
†
Supported in part by the Research Grants Council of Hong Kong.
‡
Corresponding author. E-mail: [email protected].

1
perfect if and only if the system AG x ≤ 1, x ≥ 0 is TDI, where AG is the clique-vertex incidence
matrix of G. The reader is referred to Chudnovsky et al. [10, 12] for the proof of the Strong Perfect
Graph Theorem and to Chudnovsky et al. [8] for recognition of perfect graphs.
A rational system Ax ≤ b is called box-totally dual integral (box-TDI) if Ax ≤ b, l ≤ x ≤ u is
TDI for all vectors l and u, where each coordinate of l and u is either a rational number or ±∞.
By taking l = −∞ and u = ∞ it follows that every box-TDI system must be TDI. Cameron and
Edmonds [3, 5] proposed to call a graph G box-perfect if the system AG x ≤ 1, x ≥ 0 is box-TDI;
they also posed the problem of characterizing such graphs.
We make some preparations before presenting an equivalent definition of box-perfect graphs.
Let G = (V, E) be a graph (all graphs considered in this paper are simple unless otherwise stated).
For any X ⊆ V , let G[X] denote the subgraph of G induced by X. For any v ∈ V , let NG (v) denote
the set of vertices incident with v. Members of NG (v) are called neighbors of v. By duplicating a
vertex v of G we obtain a new graph G′ constructed as follows: we first add a new vertex v ′ to G,
which may or may not be adjacent to v, and then we join v ′ to all vertices in NG (v).
As usual, let α(G) and χ(G) denote respectively the stable number and chromatic number of
G. Let χ̄(G) = χ(Ḡ), which is the clique cover number of G. For any integer q ≥ 1, let
αq (G) = max{|X| : X ⊆ V (G) with χ(G[X]) ≤ q}, and
χ̄q (G) = min{q χ̄(G − X) + |X| : X ⊆ V (G)}.
Notice that α1 = α and χ̄1 = χ̄. A graph G is called q-perfect if αq (G[X]) = χ̄q (G[X]) holds for
all X ⊆ V (G). This concept was introduced by Lovász [17] as an extension of perfect graphs, since
1-perfect graphs are precisely perfect graphs. Let us call a graph totally perfect if it is q-perfect for
all integers q ≥ 1. Lovász pointed out that comparability graphs, incomparability graphs, and line
graphs of bipartite graphs are totally perfect. However, S3 is not 2-perfect, showing that a perfect
graph does not have to be q-perfect when q > 1.

Figure 1.1: Graph S3 and its complement S̄3

Theorem 1.1 (Cameron [4]). A graph is box-perfect if and only if every graph obtained from this
graph by repeatedly duplicating vertices is totally perfect.

This theorem implies the following immediately.

Corollary 1.2 (Cameron [4]). (1) Induced subgraphs of a box-perfect graph are box-perfect.
(2) Duplicating vertices in a box-perfect graph results in a box-perfect graph.
(3) Comparability and incomparability graphs are box-perfect.

The next proposition contains a few other important observations made by Cameron [4]. A
matrix A is totally unimodular if the determinant of every square submatrix of A is 0 or ±1. A

2
{0, 1}-matrix A is balanced if none of its submtrices is the vertex-edge incidence matrix of an odd
cycle. For each graph G, let BG be the submatrix of AG obtained by keeping only rows that
correspond to maximal cliques of G. Let us call G totally unimodular or balanced if BG is totally
unimodular or balanced. It is worth pointing out that bipartite graphs and their line graphs are
totally unimodular, and every totally unimodular graph is balanced. In addition, as shown by
Berge [1], all balanced graphs are totally perfect. Let S̄3+ be obtained from the complement S̄3 of
S3 by adding a new vertex v and joining v to all six vertices of S̄3 .

Proposition 1.3 (Cameron [4]). (1) S̄3+ is not box-perfect.

(2) Totally unimodular graphs are box-perfect.
(3) Balanced graphs do not have to be box-perfect, shown by S̄3+ .
(4) The complement of a box-perfect graph does not have to be box-perfect, shown by S̄3 .
(5) Box-perfectness is not preserved under taking clique sums, shown by S3 .

As we have seen, many nice properties of perfect graphs are not satisfied by box-perfect graphs.
Another property of this kind is substitution: substituting a vertex of a box-perfect graph by a
box-perfect graph does not have to yield a box-perfect graph, as shown by S̄3+ (which is obtained
by substituting a vertex of K2 with S̄3 ). To our knowledge, almost none of the known summing
operations that preserve perfectness can carry over to box-perfectness – this makes it extremely
hard to obtain a structural characterization of box-perfect graphs!
At this point, the only known box-perfect graphs are totally unimodular graphs, comparability
graphs, incomparability graphs, and p-comparability graphs (where p ≥ 1 and 1-comparability
graphs are precisely comparability graphs) [3, 5]. Cameron and Edmonds [3] conjectured that
every parity graph is box-perfect. In this paper we confirm this conjecture and identify several
other classes of box-perfect graphs, including claw-free box-perfect graphs. In the next section we
construct a class R of non-box-perfect graphs, from which we characterize box-perfect split graphs.
It turns out that every minimal non-box-perfect graph that we know of is contained in a graph
from R. This observation raises the question: is it true that a graph G is box-perfect if and only
if G does not contain any graph in R as an induced subgraph?
In addition to structural description, the other difficulty with the study of box-perfect graphs
lies in the lack of a proper tool for establishing box-perfectness. In section 3 we introduce a so-
called ESP property, which is sufficient for a graph to be box-perfect. Although recognizing box-
perfectness is an optimization problem, our approach based on the ESP property is of transparent
combinatorial nature and hence is fairly easy to work with. For convenience, we call a graph ESP if
it has the aforementioned ESP property. In the remainder of this paper, we shall establish several
classes of box-perfect graphs by showing that they are actually ESP, including all classes obtained
by Cameron [3, 4, 5]. We strongly believe that the ESP property is exactly the tool one needs for
the study of box-perfect graphs.

Conjecture 1.4. A perfect graph is box-perfect if and only if it is ESP if and only if it contains
none of the members of R as an induced subgraph.

We close this section by mentioning a result on the complexity of recognizing box-perfect graphs.

Theorem 1.5 (Cook [13]). The class of box-perfect graphs is in co-NP.

3
2 A class of non-box-perfect graphs

Let Sn be the graph obtained from cycle v1 v2 ...v2n v1 by adding edges vi vj for all distinct even
i, j. It was proved in [4] that S2n+1 is not box-perfect for all n ≥ 1. In this section we construct
a class of non-box-perfect graphs, which include S̄3+ and S2n+1 (n ≥ 1). We will use this result
to characterize box-perfect split graphs (a graph is split if its vertex set can be partitioned into a
clique and a stable set).
Let G = (U, V, E) be a bipartite graph, where U = {u1 , ..., um } and V = {v1 , ..., vn }. The
biadjacency matrix of G is the {0, 1}-matrix M of dimension m × n such that Mi,j = 1 if and only if
ui vj ∈ E. Let Q be the set of bipartite graphs G such that its biadjacency matrix M is not totally
unimodular but all submatrices of M are. The following is a classical result of Camion.

Lemma 2.1 (Camion [6]). Every graph G = (U, V, E) in Q is Eulerian. In addition, G satisfies
|U | = |V | and |E| ≡ 2 (mod 4).

Let R be the class of graphs constructed as follows. Take a bipartite graph G′ = (U, V, E ′ ) ∈ Q
and a graph G′′ = (V, E ′′ ) such that NG′ (u) is a clique of G′′ for all u ∈ U . Let G = (U ∪V, E ′ ∪E ′′ ).
If there exists u ∈ U with NG′ (u) = V then G − u belongs to R; otherwise G belongs to R.
Examples. For each odd n ≥ 3, Sn belongs to R since Sn can be constructed from a cycle
G′ = C2n ∈ Q and a complete graph G′′ = Kn , where no vertex is deleted in the construction.
Graph S̄3+ also belongs to R. In this case a vertex is deleted in the construction, see Figure 2.1.

Figure 2.1: Graph S̄3+ is constructed from a bipartite graph in Q and K4

Lemma 2.2. No graph in R is box-perfect.

Proof. Let G ∈ R be constructed from G′ = (U, V, E ′ ) ∈ Q and G′′ = (V, E ′′ ). Let AG and BG be
the clique and maximal clique matrices of G. Then AG can be expressed as AG = [BCG ]. Let M be
the biadjacency matrix of G′ and let n := |U | (= |V |). Since every u ∈ U belongs to exactly one
maximal clique of G, the column of BG that corresponds to u has precisely one nonzero entry. If
no vertex was deleted in the construction of G then BG can be expressed as BG = [M In
N 0 ], where
the first n columns are indexed by V and the last n columns are indexed by U . If a vertex u0 ∈ U
was deleted in the construction of G, then G′′ has to be a complete graph. In this case, since U
does not have a second vertex adjacent to all vertices in V , BG can be expressed as [M, J], where
I
Jn×(n−1) = [ n−10 ] and the last row of M , which corresponds to u0 , is a vector of all ones. By Lemma
2.1, all entries of 1T M and M 1 are even, and 1T M 1 = 4m + 2, for an integer m > 0. We consider
the dual programs (with A = AG )

max{wT x : Ax ≤ 1; x ≥ l} = min{y T 1 − z T l : y T A − z T = wT ; y, z ≥ 0}. (2.1)

4
 Suppose no vertex was deleted in the construction of G. Let p > 2m + 1 be a prime and let
1 
21
" #
1 T    1  
2 M 1 0 2p 1 0
w= , l= 1 , x= 1 , y =  0 , z = 1 .
0 1 − 2p M 1 1 − 2p M 1 21
0

Then it is routine to verify that w is integral, l ≥ 0, and x, (y, z) are feasible solutions to (2.1).
Moreover wT x = 2m+1 T T
2p = y 1 − z l, so x, (y, z) are optimal solutions. Since the optimal value is
1
not p -integral, while l is, it follows that the dual does not have an integral optimal solution and so
G is not box-perfect. Next, suppose that a vertex was deleted in the construction of G. The proof
for this case is almost identical to the proof for the last case. The only difference is that BG has
2n − 1 columns, instead of 2n columns. Thus we need to truncate the corresponding vectors. To
be precise, let
1 T     1
 1   
2 M 1 0 n 1 2 1 0
w= , l= T , x= T , y= , z= 1 .
0 J (1 − n1 M 1) J (1 − n1 M 1) 0 21

Using the fact that the last row of M is 1T we deduce that x and (y, z) are feasible solutions, and
wT x = 2m+1
n = y T 1 − z T l, which implies that both solutions are optimal. Furthermore, since M 1
1
is even and its last entry is n, we deduce that n is even and thus l is n/2 -integral. However, the
2m+1 1
optimal value n is not n/2 -integral, so G is not box-perfect, which proves the theorem.
To identify all minimally non-box-perfect split graphs, we consider the following subsets of Q.
Let Q1 consist of all bipartite graphs G = (U, V, E) ∈ Q such that U has a vertex adjacent to all
vertices of V . Let Q2 consist of all bipartite graphs G = (U, V, E) ∈ Q\Q1 such that the graph
obtained from G by adding a vertex and making it adjacent to all vertices of V does not contain
any graph in Q1 as an induced subgraph. Let S consist of all graphs in R that are constructed
from a bipartite graph G′ ∈ Q1 ∪ Q2 and a complete graph G′′ . It is clear that all members of S
are split graphs. Moreover, S̄3+ and S2n+1 (n ≥ 1) belong to S.
Theorem 2.3. The following are equivalent for any split graph G.
(1) G is box-perfect;
(2) no graph in S is an induced subgraph of G;
(3) G is totally unimodular.

Proof. Implication (3) ⇒ (1) follows from Proposition 1.3(2) and implication (1) ⇒ (2) follows
from Lemma 2.2 and Corollary 1.2(1). To prove (2) ⇒ (3), let G = (U, V, E) be a split graph,
where U is a stable set and V is a clique. Let G′′ = G[V ] and G′ = G\E(G′′ ). Let G′′′ be the
bipartite graph obtained from G′ by adding a vertex w adjacent to all vertices in V . Let M be the
biadjacency matrix of G′′′ .
We first prove that M is totally unimodular. Suppose otherwise. Then G′′′ has an induced
subgraph H ′ ∈ Q. Let us choose H ′ so that H ′ contains the new vertex w whenever it is possible.
Consequently, H ′ ∈ Q1 ∪ Q2 . Let H be constructed from H ′ and a complete graph H ′′ . Then
H ∈ S and, by the construction of G, G contains H as an induced subgraph. This contradicts (2)
and thus M has to be totally unimodular.
Let N be the biadjacency matrix of G′ . Then BG = [N, I] or [N1 I0 ], depending on if V is a
maximal clique of G. Notice that M = [N1 ]. So BG , and thus G, is totally unimodular.

5
 This theorem shows that all minimally non-box-perfect split graphs are contained in S. In fact,
S consists of precisely such graphs.
Theorem 2.4. A split graph G belongs to S if and only if G is not box-perfect but all its induced
subgraphs are.

Proof. The backward implication follows immediately from Theorem 2.3. To prove the forward
implication, let G ∈ S. By Lemma 2.2, we only need to show that G − w is box-perfect for all
w ∈ V (G). Suppose G is constructed from a bipartite graph G′ = (U, V, E ′ ) ∈ Q1 ∪ Q2 and a
complete graph G′′ = (V, E ′′ ). Let M be the biadjacency matrix of G′ and let n := |U | = |V |.
I
Observe that if G′ ∈ Q1 then BG = [N, J], where N = M and J = [ n−1 ′
0 ]; if G ∈ Q2 then
M In
BG = [N, J], where N = [ 1 ] and J = [ 0 ].
Now it is straightforward to verify that, for each u ∈ U , BG−u = [N ′ , J ′ ] is obtained from BG by
deleting the row and the column indexed by u; for each v ∈ V , BG−v = [N ′ , J ′ ] is obtained from BG
by deleting the column indexed by v and also possibly the last row. In both cases, N ′ is a proper
submatrix of N . This implies that N ′ is totally unimodular and thus so is [N ′ , J ′ ]. Consequently,
G − w is box-perfect (totally unimodular) for all w ∈ V (G), which proves the theorem.
As we observed earlier that S̄3+ and S2n+1 (n ≥ 1) belong to S. Thus these graphs are minimally
non-box-perfect. We point out that, in addition to graphs in S, other minimally non-box-perfect
graphs can also be obtained using Lemma 2.2. For instance, the graph illustrated in Figure 2.2
is constructed from G′ = C10 and G′′ = C5 + e. By Lemma 2.2, this graph G is not box-perfect.
However, G is not minimally non-box-perfect since H = G − {9, 0} is not box-perfect, which is
certified by vectors wT = (1, 1, 1, 1, 1, 0, 0, 0), lT = (0, 0, 0, 0, 0, 12 , 12 , 12 ), xT = ( 41 , 14 , 14 , 14 , 34 , 21 , 12 , 12 ),
y T = (0, 12 , 12 , 21 , 12 , 12 ), and z T = (0, 0, 0, 0, 0, 12 , 12 , 12 ), where the first row of BH is the triangle 123.
It can be shown that H is in fact minimally non-box-perfect because H − x is totally unimodular
for x = 1, 2, 3, 4, 5, 6, 7, and H − 8 has the ESP property defined in the next section which implies
the box-perfectness.
7 2 6

3 1

8 4 5 0
9

Figure 2.2: A new non-box-perfect graph G

3 ESP graphs

In this section we introduce a so-called ESP property, which is sufficient for a graph to be box-
perfect. We shall use this combinatorial property to identify several new classes of box-perfect
graphs. We begin with a few lemmas.
Lemma 3.1 (Chen, Ding and Zang [7]). Suppose a1 and a2 are rational vectors with a1 ≥ a2 ,
and b1 and b2 are rational numbers with b1 ≤ b2 . Then the system Ax ≤ b, aT1 x ≤ b1 , aT2 x ≤ b2 ,

6
x ≥ 0 is box-TDI if and only if the system Ax ≤ b, aT1 x ≤ b1 , x ≥ 0 is box-TDI.

Lemma 3.2 (Cameron [4]). The system Ax ≤ b is box-TDI if and only if the system Ax ≤ b,
x ≤ u is TDI, for all vectors u, where each coordinate of u is either a rational number or +∞.

The next two lemmas are reformulations of Theorem 22.7 and Theorem 22.13 of Schrijver [21].

Lemma 3.3 (Schrijver [21]). Suppose the system Ax ≤ b, x1 ≤ u is TDI for all rational numbers
u, where x1 is the first coordinate of x. Then Ax ≤ b is TDI.

Lemma 3.4 (Schrijver [21]). A rational system Ax ≤ b, x ≥ 0 is TDI if and only if min{y T b :
y T A ≥ w T, y ≥ 0 is half-integral} is finite and is attained by an integral y, for each integral vector
w for which min{y T b : y T A ≥ wT, y ≥ 0} is finite.

The next are two easy corollaries.

Lemma 3.5. A graph G is box-perfect if and only if the system BG x ≤ 1, 0 ≤ x ≤ u is TDI for
all rational vectors u ≥ 0.

Proof. The forward implication follows immediately from the definition of box-TDI and Lemma
3.1. Conversely, Lemma 3.2 and Lemma 3.3 imply that BG x ≤ 1, x ≥ 0 is box-TDI. Then the
result follows from Lemma 3.1.

Lemma 3.6. A graph G is box-perfect if and only if for all rational u ≥ 0 and integral w ≥ 0,
min{y T 1 + z T u| y T BG + z T ≥ 2wT ; y, z ≥ 0 integral}
≥ 2 min{y T 1 + z T u| y T BG + z T ≥ wT ; y, z ≥ 0 integral}. (3.1)

Proof. Observe that, for all vectors u ≥ 0 and w, the three programs
min{y T 1 + z T u| y T BG + z T ≥ wT ; y, z ≥ 0}
min{y T 1 + z T u| y T BG + z T ≥ wT ; y, z ≥ 0 half-integral}
min{y T 1 + z T u| y T BG + z T ≥ wT ; y, z ≥ 0 integral}
are finite. Moreover, replacing w by w + does not change the minimum values of these programs,
where w+ is obtained from w by turning its negative coordinates into zero. Therefore, the result
follows immediately from Lemma 3.5 and Lemma 3.4.
Let G = (V, E) be a graph. For any multiset Λ of cliques of G and any v ∈ V , let dΛ (v) denote
the number of members of Λ that contain v. We call G equitably subpatitionable (ESP) if for every
set Λ of maximal cliques of G there exist two multisets Λ1 and Λ2 of cliques of G (which are not
necessarily members of Λ) such that
(i) |Λ1 | + |Λ2 | ≤ |Λ|;
(ii) dΛ1 (v) + dΛ2 (v) ≥ dΛ (v), for all v ∈ V ; and
(iii) min{dΛ1 (v), dΛ2 (v)} ≥ ⌊dΛ (v)/2⌋, for all v ∈ V .
We call (Λ1 , Λ2 ) an equitable subpartition of Λ, and refer to the above (i), (ii), and (iii) as ESP
property. Note that (i) is equivalent to |Λ1 | + |Λ2 | = |Λ| since we may include empty cliques in Λ1
and Λ2 . Similarly, (ii) is equivalent to dΛ1 (v) + dΛ2 (v) = dΛ (v) for all v, since cliques in Λ1 , Λ2 can
be replaced by smaller ones. Finally, it is also easy to see that in an ESP graph every multiset Λ
of cliques admits an equitable subpartition. We will use these facts without further explanation.

7
Theorem 3.7. Every ESP graph G = (V, E) is box-perfect.

Proof. By Lemma 3.6 we only need to show that inequality (3.1) holds for all rational u ≥ 0 and
all integral w ≥ 0. Let (y T , z T ) be an optimal solution of the first minimum in (3.1). Let C be the
set of maximal cliques of G and let D be the multiset of members of C such that each C ∈ C appears
in D exactly yC times. Let Λ be the set of C ∈ C such that yC is odd. Since G is ESP, Λ admits a
equitable subpartition (Λ1 , Λ2 ). Since every clique can be extended into a maximal clique, we may
assume without loss of generality that members of Λ1 and Λ2 are all in C. Let D0 be the multiset
of members of C such that each C ∈ C appears ⌊yC /2⌋ times. It follows that D = D0 ⊎ D0 ⊎ Λ,
where ⊎ stands for multiset sum. For i = 1, 2, let Di = D0 ⊎ Λi . We deduce from (i) that
(1) |D1 | + |D2 | ≤ |D|.
Let p = y T BG + z T − 2w T and let v ∈ V . Without loss of generality we assume
(2) dD1 (v) ≥ dD2 (v) and pv z v = 0.
Since dD (v) = 2dD0 (v) + dΛ (v), we deduce from (ii-iii) that dD1 (v) + dD2 (v) ≥ dD (v) and
dDi (v) = dD0 (v) + dΛi (v) ≥ ⌊dD (v)/2⌋ (i = 1, 2). Thus we conclude from (2) that
(3) dD1 (v) ≥ ⌈dD (v)/2⌉ and dD2 (v) ≥ ⌊dD (v)/2⌋.
By the definition of D we have dD (v) = y T Bv , where Bv is the column of BG indexed by v. So
(4) dD (v) + z v = pv + 2w v ≥ 2wv .
Since wv is an integer, we deduce that
(5) wv ≤ ⌊(dD (v) + z v )/2⌋.
Setting z 1v = ⌊z v /2⌋ and z 2v = ⌈z v /2⌉, we have
(6) z 1v uv + z 2v uv = z v uv .
We further claim that
(7) dDi (v) + z iv ≥ wv , for i = 1, 2.
To see (7), recall pv z v = 0 from (2). If dD (v) is even, we deduce from (4) that z v is even, which
implies, by (3-4), that dDi (v) + z iv ≥ 12 (dD (v) + z v ) ≥ wv . So we assume that dD (v) is odd. If
z v = 0 then, by (3) and (5), dDi (v) + z iv = dDi (v) ≥ ⌊dD (v)/2⌋ ≥ wv . Else, by (2) and (4), z v is
odd. Thus dDi (v) + z iv ≥ 12 (dDi (v) ± 1) + 12 (z v ∓ 1) = 21 (dDi (v) + z v ) ≥ wv , because of (3), (5),
and the definition of z iv . So (7) holds.
For i = 1, 2, let z i = (z iv : v ∈ V ) and y i ∈ ZC+ be the multiplicity function of Di . It follows
from (7) that y Ti BG + z Ti ≥ wT , which means that both (y 1 , z 1 ) and (y 2 , z 2 ) are feasible solutions
of the second program in (3.1). From (1) and (6) we also conclude that y Ti 1+z Ti u ≤ (y T 1+z T u)/2
holds for at least one i ∈ {1, 2}. Hence inequality (3.1) holds, which proves the Theorem.
For a perfect graph G, being ESP can be characterized as follows. Let Z+ denote the set of
V (G)
nonnegative integers. For any d ∈ Z+ , let Gd denote the graph obtained from G by substituting
each vertex v with a stable set of size d(v). Note that v is deleted when d(v) = 0. Let cG = 1T BG .
In other words, for each v ∈ V (G), cG (v) is the number of maximal cliques of G that contain v.

8
 V (G)
Theorem 3.8. Let G be perfect. Then G is ESP if and only if for every d ∈ Z+ with d ≤ cG
V (G) ′ ′
there exists d′ ∈ Z+ such that ⌊d/2⌋ ≤ d′ ≤ ⌈d/2⌉ and α(Gd ) + α(Gd−d ) ≤ α(Gd ).
V (G)
Proof. To prove the forward implication, let G be ESP and let d ∈ Z+ . Since Gd is perfect, its
vertex set can be partitioned into α(Gd ) cliques. These cliques naturally correspond to a multiset
Λ of α(Gd ) cliques of G. Note that |Λ| = α(Gd ) and dΛ = d. Since G is ESP, Λ admits a equitable
subpartition (Λ1 , Λ2 ). By deleting vertices from cliques in Λ1 and Λ2 we can obtained multisets Λ∗1
and Λ∗2 of cliques of G such that |Λ∗1 |+|Λ∗2 | ≤ |Λ1 |+|Λ2 |, dΛ∗1 +dΛ∗2 = d, and min{dΛ∗1 , dΛ∗2 } ≥ ⌊d/2⌋.
Let d′ = dΛ∗1 . Then ⌊d/2⌋ ≤ d′ ≤ ⌈d/2⌉ and
′ ′
α(Gd ) + α(Gd−d ) ≤ α(GdΛ1 ) + α(GdΛ2 ) ≤ |Λ1 | + |Λ2 | ≤ |Λ| = α(Gd ),
which proves the forward implication.
To prove the backward implication, let Λ be a set of maximal cliques of G. Then d := dΛ ≤ cG
and thus there exists d′ as stated in the theorem. Let d1 = d′ and d2 = d − d′ . For i = 1, 2,
vertices of Gdi can be partitioned into α(Gdi ) cliques, and these cliques correspond to a multiset Λi
of α(Gdi ) cliques of G. Note that dΛi = di . Thus (Λ1 , Λ2 ) is an equitable subpartition of Λ, which
proves the theorem.
We first remark that α(Gd ) is exactly the maximum of
P
v∈S d(v) over all stable sets S of
G. Sometimes this interpretation is more convenient. We also remark that we do not know a
box-perfect graph that is not ESP. It seems reasonable to conjecture that no such a graph exists.

4 Known box-perfect graphs

Cameron [4] identified a few classes of box-perfect graphs. In this section we prove that they are in
fact ESP graphs. Our results could be stronger than the results of Cameron if ESP and box-perfect
are not equivalent. But the main reason for establishing our results is for future applications. We
envision that more ESP graphs (possibly all box-perfect graphs) can be constructed from basic
ESP graphs. Therefore, it is important to make sure that all known box-perfect graphs are ESP.

4.1 Totally unimodular graphs

It is well known (see Theorem 19.3 of [21]) that in a totally unimodular matrix, each set of rows
can be partitioned so that the sum of one part minus the sum of the other part is a {0, ±1}-vector.
If G is totally unimodular then BG has this partition property, which implies immediately that G
satisfies the definition of ESP graphs. Thus we have the following.

Theorem 4.1. Totally unimodular graphs are ESP.

We point out that totally unimodular graphs include graphs like interval graphs, bipartite
graphs, and block graphs (every block is a complete graph).

9
4.2 Incomparability graphs

Theorem 4.2. Every incomparability graph G is ESP.

V (G)
Proof. Since G is perfect, we may apply Theorem 3.8. Let d ∈ Z+ . Note that Gd is again
an incomparability graph. In fact, let P be a poset such that G is the incomparability of P and
let P d be obtained from P by replacing each element v with a chain of size d(v). Then Gd is
the incomparability graph of poset P d . For each positive integer i, let Ai be the set of maximal
elements of P d − (A1 ∪ ... ∪ Ai−1 ). Then (A1 , ..., An ) is a partition of V (Gd ) into cliques, where
n = α(Gd ). Let V1 be the union of Ai for all odd i and let V2 be the union of Ai for all even i. Then
V (G)
Gd [V1 ] and Gd [V2 ] can be expressed as Gd1 and Gd2 , respectively, for some d1 , d2 ∈ Z+ . It is
easy to see that d1 + d2 = d and ⌊d/2⌋ ≤ dj ≤ ⌈d/2⌉ (j = 1, 2). Moreover, each α(Gdj ) is bounded
by the number of Ai s contained in Vj . Therefore, α(Gd1 ) + α(Gd2 ) ≤ α(Gd ), which implies that
d′ = d1 satisfies Theorem 3.8 and thus G is ESP.

4.3 p-Comparability graphs

p-Comparability graphs were introduced in [3] and were shown [3, 5] to be box-perfect. We show
that they are ESP. Let D be a digraph with a special set T of vertices such that every arc is in a
dicycle (directed cycle) and every dicycle meets T exactly once. In particular, D has no arc between
any two vertices of T . If p is an integer with |T | ≤ p, then a p-comparability graph G is defined
from D by adding all chords of all dicycles, then deleting T , and finally ignoring all directions on
edges. Note that 1-comparability graphs are precisely comparability graphs.

Theorem 4.3. Every p-comparability graph G is ESP.

To prove this theorem we will need the following Lemma. Let D = (V, A) be a digraph. For
each dicycle C of D, the incidence vector of C is the vector χC ∈ {0, 1}A such that χC (a) = 1 if
and only if a is on C. A sum of incidence vectors of (not necessarily distinct) dicycles of D is called
a circulation of D. The following is a special case of Corollary 11.2b of [22].

Lemma 4.4. Every circulation f is the sum of two circulations f1 , f2 such that ⌊f /2⌋ ≤ fi ≤ ⌈f /2⌉
holds for both i = 1, 2.

Proof of Theorem 4.3. Let G be constructed from D and T . Let D ∗ be obtained from D by
splitting each vertex v into v ′ and v ′′ such that arcs entering v are now entering v ′ , and arcs leaving
v are now leaving v ′′ . We also add an arc from v ′ to v ′′ . Observe that for every dicycle C of D, D ∗
has a unique dicycle C ∗ such that A(C ∗ ) ∩ A(D) = A(C). Moreover, every dicycle of D ∗ can be
expressed as C ∗ for a dicycle C of D.
We will use a fact proved in [5] that for every clique K of G, there exists a dicycle CK of D
such that K ⊆ V (CK ).
Let Λ be a set of maximal cliques of G. We prove the theorem by showing that Λ admits an
equitable subpartition. Let f be the sum of incidence vectors of CK ∗ over all K ∈ Λ. Since each C
K
meets T exactly once, each CK ∗ must meet T ∗ = {t′ t′′ : t ∈ T } exactly once. As a result, |Λ| equals

10
the sum of f (a) over all a ∈ T ∗ . In addition, since each K ∈ Λ is a maximal clique, we must have
V (CK ) − T = K. This implies that dΛ (v) = f (v ′ v ′′ ) holds for all v ∈ V (D).
Let f1 and f2 be the two circulations of D ∗ determined by Lemma 4.4. For i = 1, 2, let Ci∗ be
the multiset of dicycles of D ∗ such that fi is the sum of χC ∗ over all C ∗ ∈ Ci∗ . Then let Ci be the
multiset {C : C ∗ ∈ Ci∗ } and Λi = {V (C) − T : C ∈ Ci }. By the construction of G, each P member
of Λi is a clique of G. Moreover, dΛi (v) = fi (v ′ v ′ ) holds for all v ∈ V (G), and |Λi | = a∈T ∗ fi (a).
Therefore, (Λ1 , Λ2 ) is an equitable subpartition of Λ, which proves that G is ESP.
Remark. Let us call a graph strong ESP if every set Λ of maximal cliques admits an equitable
subpartition (Λ1 , Λ2 ) with max{|Λ1 |, |Λ2 |} ≤ ⌈|Λ|/2⌉. This proof also proves that (1-)comparability
graphs are in fact strong ESP.

5 Parity graphs

A graph is called a parity graph if any two induced paths between the same pair of vertices have the
same parity. These are natural extensions of bipartite graphs and they are perfect [19]. Cameron
and Edmonds [3] conjectured that every parity graph is box-perfect. The objective of this section
is to present a proof of this conjecture.
To establish our result we need a structural characterization of parity graphs. Let H be a graph
with a stable set S such that all vertices of S have the same set of neighbors. Let B be a bipartite
graph and let T be a subset of a color class of B with |T | = |S|. Let G be obtained from the disjoint
union of H and B by identifying S with T . We call G a bipartite extension of H by B, and we also
call the construction of G from H bipartite extension.
Lemma 5.1 (Burlet and Uhry [2]). Every connected parity graph can be constructed from a single
vertex by repeatedly duplicating vertices and bipartite extensions.
Lemma 5.2. Duplicating a vertex in an ESP graph results in an ESP graph.

Proof. Let ESP graph G have a vertex v. Let G′ be obtained by duplicating v and let v ′ be the
new vertex. For any set Λ′ of maximal cliques of G′ , we prove that Λ′ has an equitable subpartition.
We define Λ as follows. If vv ′ is an edge then Λ = {K − v ′ : K ∈ Λ′ }; if vv ′ is not an edge then
Λ = {K : v ′ 6∈ K ∈ Λ} ⊎ {K − v ′ + v : v ′ ∈ K ∈ Λ′ }. Note that Λ is a multiset of maximal cliques of
G. Since G is ESP, Λ admits an equitable subpartition (Λ1 , Λ2 ). By deleting vertices from cliques
in Λ1 and Λ2 we may assume that dΛ1 + dΛ2 = dΛ and ⌊dΛ /2⌋ ≤ dΛi ≤ ⌈dΛ /2⌉ (i = 1, 2).
If vv ′ is an edge, let Λ′i = {K : v 6∈ K ∈ Λi } ⊎ {K + v ′ : v ∈ K ∈ Λi } (i = 1, 2). Then (Λ′1 , Λ′2 )
is an equitable subpartition of Λ′ because dX (v ′ ) = dX (v) holds for X ∈ {Λ, Λ′1 , Λ′2 }.
Now suppose vv ′ is not an edge. Note that dΛ (v) = dΛ′ (v) + dΛ′ (v ′ ). Also we may assume that
dΛ1 (v) = ⌊dΛ (v)/2⌋ and dΛ2 (v) = ⌈dΛ (v)/2⌉. Let

m1 = ⌊dΛ′ (v)/2⌋, m2 = ⌈dΛ′ (v)/2⌉, m′1 = dΛ1 (v) − m1 , m′2 = dΛ2 (v) − m2 .

Then
m1 + m2 = dΛ′ (v), m′1 + m′2 = dΛ′ (v ′ ), min{m′1 , m′2 } ≥ ⌊dΛ′ (v ′ )/2⌋.

11
For i = 1, 2, let Λ′i be obtained from Λi by turning m′i cliques K that contain v into K − v + v ′ .
Then the above equalities and inequalities imply that (Λ′1 , Λ′2 ) is an equitable subpartition of Λ′ .
Remark. Clearly, this proof also proves that duplicating a vertex in a strong ESP graph results
in a strong ESP graph.

Theorem 5.3. Parity graphs are ESP.

Proof. By Lemma 5.2, we only need to show that if G is a bipartite extension of an ESP graph H
by a bipartite graph B = (X, Y, E), then G is ESP. Let X0 ⊆ X be the intersection of H and B.
Let Λ be a set of maximal cliques of G. Naturally, Λ can be partitioned into ΛH and ΛB , which are
maximal cliques of H and edges of B, respectively. Now we find an equitable subpartition (Λ′B , Λ′′B )
of ΛB and an equitable subpartition (Λ′H , Λ′′H ) of ΛH such that (Λ′B ∪ Λ′H , Λ′′B ∪ Λ′′H ) is an equitable
subpartition of Λ. Let X0 be partitioned into (X1 , X2 ) such that X1 consists of x ∈ X0 with both
dΛB (x) and dΛH (x) odd. Since (Λ′B , Λ′′B ) and (Λ′H , Λ′′H ) are always compatible on vertices in X2 ,
we only need to focus on vertices in X1 .
Without loss of generality, let ΛB = E. Suppose B has 2t vertices of odd degree. Then E can be
partitioned into cycles and t paths P1 , ..., Pt . Let (Λ′B , Λ′′B ) be defined by assigning edges to the two
parts alternatively along the cycles and paths. Then (Λ′B , Λ′′B ) is an equitable partition. Note that
we have the following freedom in the assignment. Let x ∈ X1 and let Pi be the path with x as an
end. If the other end of Pi is not in X1 , then we may choose dΛ′B (x) to be ⌊dΛB (x)/2⌋ or ⌈dΛB (x)/2⌉,
as we wish (without changing dΛ′B (z) and dΛ′′B (z) for any other z ∈ X1 ). If the other end of Pi is a
vertex x′ in X1 , then we may assume that dΛ′B (x) = ⌊dΛB (x)/2⌋ and dΛ′B (x′ ) = ⌈dΛB (x′ )/2⌉. Let
(x1 , x′1 ), ..., (xk , x′k ) be these pairs in X1 .
Let H1 be obtained from H by deleting x′1 , ..., x′k and let Λ1 be obtained from ΛH by replacing
each x′i with xi . Note that dΛ1 (xi ) = dΛH (xi ) + dΛH (x′i ) for all i, while dΛ1 (v) = dΛH (v) for all
other vertices v of H1 . Since H is ESP, so is H1 . Let (Λ′1 , Λ′′1 ) be an equitable subpartition of Λ1 .
Without loss of generality, we assume dΛ′1 (xi ) = dΛ′′1 (xi ) = dΛ1 (xi )/2 for all i. Let Λ′H be obtained
from Λ′1 by turning ⌊dΛH (x′i )/2⌋ of its cliques K that contain xi into K − xi + x′i (for every i). Then
dΛ′H (xi ) = ⌈dΛH (xi )/2⌉ and dΛ′H (x′i ) = ⌊dΛH (x′i )/2⌋. Let Λ′′H be obtained analogously. Now it is
straightforward to verify that, the freedom on partition (Λ′B , Λ′′B ) allows us to make adjustments
so that (Λ′B ∪ Λ′H , Λ′′B ∪ Λ′′H ) is an equitable subpartition of Λ.

6 Complements of line graphs

In the rest of this paper we allow some graphs to have loops and parallel edges. We call these
multigraphs and we reserve the word graph for simple graphs. If a multigraph H is obtained from a
graph H0 by adding loops and parallel edges, then H0 is called a simplification of H and is denoted
by si(H).
Let L(H) denote the line graph of a multigraph H. Under this circumstance, we always make
the following implicit assumptions:
(i) H has no isolated vertices (deleting an isolated vertex does not affect L(H));
(ii) H has no loops (replacing a loop with a pendent edge does not affect L(H));

12
 (iii) H has no distinct vertices x, y, z such that z is the only neighbor of x and the only neighbor
of y (replacing edges between y and z by edges between x and z does not affect L(H)).
The complement of L(H) will be denoted by L̄(H). Our results in the next two sections imply
a characterization of box-perfect line graphs. The goal of this section is to characterize box-perfect
graphs that are complements of line graphs.
Theorem 6.1. Let G = L̄(H) be perfect. Then G is box-perfect if and only if G is {S3 , S̄3+ }-free.

Our proof of this theorem is divided into a sequence of lemmas. We first determine the structure
of {S3 , S̄3+ }-free perfect graphs of the form L̄(H), and then we confirm that all such graphs are
ESP. We will see that some of these graphs are in fact strong ESP.
We need a result of Gallai [15] which identifies eight classes and ten individual graphs such that
a graph is a comparability graph if and only if it does not contain any of these identified graphs as
an induced subgraph. We will use the following immediate consequence of Gallai’s theorem. Let Γ
be the graph obtained from a 6-cycle v1 v2 v3 v4 v5 v6 v1 by adding two edges v1 v3 and v1 v5 .
Lemma 6.2. Let G be claw-free and perfect. Then G is an incomparability graph if and only if G
does not contain any of S3 , S̄3 , Γ, and C2n (n ≥ 3) as an induced subgraph.

Let K4+ denote the graph obtained from K4 by adding two pendent edges to two of its distinct
+
vertices. Let K2,n denote the graph obtained from K2,n (n ≥ 3) by adding a pendent edge to a
degree-2 vertex and an edge between the two degree-n vertices.
Lemma 6.3. Let L̄(H) be {C5 , S3 , S̄3+ }-free. If H contains S̄3 as a subgraph then si(H) is either
+
K4+ or a subgraph of K2,n for some n ≥ 3.

Proof. Since S̄3 is a subgraph of H, we assume V (H) = {x1 , x2 , x3 , y1 , y2 , y3 , z1 , ..., zm } such that
x1 x2 x3 is a triangle and xi yi ∈ E(H) (i = 1, 2, 3). If m = 0 then it is straightforward to verify the
conclusion of the lemma, using the fact that H does not contain C5 as a subgraph. So we assume
m > 0. Let K1,3 ∗ denote the graph obtained from K
1,3 by subdividing each edge exactly once. Note
∗
that K1,3 is not a subgraph of H since L̄(K1,3 ∗ ) = S . As a result, each z is adjacent to none of
3 i
+
y1 , y2 , y3 , and at most two of x1 , x2 , x3 . Furthermore, since L̄(H) is S̄3 -free, the entire neighborhood
of each zi must be a subset of {x1 , x2 , x3 } of size one or two (here we also use assumption (i) above).
By assumption (iii) above we may assume that each zi is adjacent to exactly two of x1 , x2 , x3 . Since
C5 is not a subgraph of H, all zi ’s must have the same set of neighborhood. Now, since m > 0, it
+
is straightforward to verify that si(H) is a subgraph of K2,m+3 .
Let C be an even cycle of length ≥ 4. Let X be a stable set of C and let Y = V (C)−X −NC (X),
where X is allowed to be empty. We construct a bipartite graph from C by adding a pendent edge
to each vertex in Y and by repeatedly duplicating vertices in X. Let C consist of all graphs that
can be constructed in this way.
Lemma 6.4. Let L(H) be perfect and S̄3 -free. Suppose H is connected and H does not contain S̄3
as a subgraph. If L(H) contains an induced Γ or C2n (n ≥ 3), then si(H) is a subgraph of a graph
in C ∪ {K3,3 }.
Proof. Suppose Γ is an induced subgraph of L(H). Then H has a subgraph with a 4-cycle x1 x2 x3 x4
and two pendent edges x1 y1 , x2 y2 . Note that x1 x3 and x2 x4 are not edges of H since S̄3 is not

13
a subgraph of H. Let z1 , ..., zm be the remaining vertices of H. If m = 0, then either si(H) is a
subgraph of K3,3 or H contains a 5-cycle. So we assume m > 0. Like in the proof of the last lemma,
∗ are not subgraphs of H, for each i we must have N (z ) = {x , x } or {x , x },
since C5 and K1,3 H i 1 3 2 4
or {xj } for some j. In addition, NH (yi ) ⊆ {xi , xi+2 } (i = 1, 2) and |NH (y1 ) ∪ NH (y2 )| ≤ 3. Now,
since H does not contain K1,3∗ , it is routine to check that si(H) is a subgraph of a graph in C.

Next, suppose L(H) is Γ-free. Then H contains a 2n-cycle x1 x2 ...x2n (n ≥ 3). Note that this
cycle has no chord (otherwise L(H) contains an induced Γ, S̄3 , or C2k+1 with k ≥ 2). Let z1 , ..., zm
be the remaining vertices of H. Using the same argument we used in the last paragraph it is
straightforward to show that each NH (zi ) is {xj } or {xj , xj+2 } for some j (where x2n+t is xt ). In
addition, if NH (zi ) = {xj , xj+2 } then NH (xj+1 ) = NH (zi ). Therefore, si(H) is a subgraph of a
graph in C.
Lemma 6.5. Suppose G has a vertex u such that G − u is bipartite and G − N (u) is edge-less.
Then G is totally unimodular.

Proof. By Theorem 19.3 of [21], we only need to show that each set Λ of maximal cliques admits
an equitable partition (Λ1 , Λ2 ), meaning that min{dΛ1 (v), dΛ2 (v)} ≥ ⌊dΛ (v)⌋, for all v ∈ V (G).
Suppose to the contrary that some Λ does not admit such a partition. We choose Λ with |Λ| as
small as possible.
Let A, B, C, D be a partition of V (G) − u such that A ∪ C, B ∪ D are stable and N (u) = B ∪ C.
Let G′ be the subgraph of G formed by edges in K − u, over all K ∈ Λ. We claim that G′ is a
forest. Suppose G′ has a cycle x1 x2 ...xn . Note that for each i, exactly one of xi xi+1 and uxi xi+1
is a clique in Λ. Let Λ′ be the rest cliques in Λ. By the minimality of |Λ|, Λ′ admits an equitable
partition (Λ′1 , Λ′2 ). Let us extend Λ′j (j = 1, 2) to Λj by including xi xi+1 or uxi xi+1 (whichever
belongs to Λ) for all i with i − j even. Then it is easy to see that (Λ1 , Λ2 ) is an equitable partition
of Λ. This contradicts the choice of Λ and thus the claim is proved. The same argument also shows
that G′ has no maximal path with two ends both in A ∪ B or both in C ∪ D. Thus all components
of G′ are paths with one end in A ∪ B and one end in C ∪ D. If G′ has only one path then the
same argument still works. If G′ has two or more paths then we can take any two of them and
treat their union as a cycle and again apply the same argument.
Recall that a graph G is strong ESP if every set Λ of maximal cliques of G admits an equitable
subpartition (Λ1 , Λ2 ) with max{|Λ1 |, |Λ2 |} ≤ ⌈|Λ|/2⌉. The next lemma follows immediately from
this definition.
Lemma 6.6. (1) If G is strong ESP then so are all its induced subgraphs.
(2) Let G1 , G2 be strong ESP and let G be obtained from the disjoint union of G1 , G2 by adding
all edges between them. Then G is also strong ESP.

In a (loopless) multigraph G, the degree of a vertex v, denoted dG (v), is the number of edges
incident with v. The next is the key step for proving Theorem 6.1.
Lemma 6.7. For every H ∈ C ∪ {K3,3 }, L̄(H) is strong ESP.
E(H)
Proof. For each µ ∈ Z+ , let µH denote the multigraph with vertex set V (H) such that the
number of edges between any two vertices x, y is zero (if xy 6∈ E(H)) or µ(xy) (if xy ∈ E(H)). Note
that µH is bipartite since H is bipartite. Let ∆(µ) denote the maximum degree of µH. By Konig’s

14
edge-coloring theorem, E(µH) is the union of k matchings if and only if k ≥ ∆(µ). Because of this
theorem and the one-to-one correspondence between cliques of L̄(H) and matchings of H, to prove
the lemma it is enough for us to show that
E(H) E(H)
(∗) for any µ ∈ Z+ there exist µ1 , µ2 ∈ Z+ such that µ1 + µ2 = µ, µi ≥ ⌊µ/2⌋ (i = 1, 2),
∆(µ1 ) ≤ ⌈∆(µ)/2⌉, and ∆(µ2 ) ≤ ⌊∆(µ)/2⌋.
In the following we construct a partition (E1 , E2 ) of E(µH) such that the multiplicity functions
µi of Ei (i = 1, 2) satisfies (∗). This partition will be constructed in several steps. In the process
we determine a partition (E1 , E2 , E3 ) of E(µH), where we begin with (E1 , E2 , E3 ) = (∅, ∅, E(µ(H))
and we keep moving edges from E3 to E1 , E2 until E3 becomes empty. For i = 1, 2, 3, let Hi denote
the subgraph of µH formed by edges in Ei .
First, for each edge e = xy of H, among all µ(e) edges of E3 that are between x and y, we
move ⌊µ(e)/2⌋ of them to E1 and ⌊µ(e)/2⌋ of them to E2 . At the end of this process, H3 becomes
a simple graph. It follows that µi ≥ ⌊µ/2⌋ (i = 1, 2) and this inequality will be satisfied no matter
how edges of H3 are moved to E1 and E2 in later steps.
If H3 has a cycle C, since H is bipartite, E(C) can be partitioned into two matchings M1 , M2 .
We move Mi from E3 to Ei (i = 1, 2). We repeat this process until H3 become a forest. At this
point, H1 and H2 have the same degree on every vertex.
Let S = {v : dµH (v) = ∆(µ)}. Suppose H3 has a leaf v that is not in S. Let P be a maximal path
of H3 starting from v. Let E(P ) be partitioned into two matchings M1 , M2 , where we assume the
edge of P that is incident with the other end u of P belongs to M1 . Then we move Mi from E3 to Ei
(i = 1, 2). After this change, dH1 (u) = ⌈dµH (u)/2⌉ ≤ ⌈∆(µ)/2⌉, dH2 (u) = ⌊dµH (u)/2⌋ ≤ ⌊∆(µ)/2⌋,
and dHi (v) ≤ ⌈dµH (v)/2⌉ ≤ ⌊∆(µ)/2⌋ (i = 1, 2). In addition, dH1 (w) = dH2 (w) for all w 6= u, v,
and dHi (u), dHi (v) will remain unchanged in the remaining process. By repeating this process we
may assume that all leaves of H3 are in S. As a consequence, ∆(µ) is odd. Note that the same
argument works if H3 has a maximal path with an odd number of edges. Thus we further assume
that in every component of H3 , all leaves are in the same color class (of any 2-coloring of H3 ).
We first consider the case H = K3,3 . We claim that each component of H3 is a path. Suppose
a component H3′ of H3 is not a path. Then H3′ has at least three leaves. Since all these leaves are
in the same color class, H3′ must have exactly three leaves z1 , z2 , z3 and they form a color class of
H. Consequently, H3′ = H3 = K1,3 . Moreover, in the previous steps of reducing H3 , no path was
ever deleted because otherwise H3 would be a subgraph of K2,2 . It follows that dµH (v ∗ ) is even,
where v ∗ ∈ V (H) − V (H3 ). However, the fact z1 , z2 , z3 ∈ S implies that µH is ∆(µ)-regular, and
thus dµH (v ∗ ) = ∆(µ) is odd. This contradiction proves our claim. Now, since each non-leaf v of
H3 has degree two, its degree in µH is even and thus v 6∈ S. It follows that moving all edges of E3
to E1 results in the required partition.
Next suppose H ∈ C. Let H3′ be a component of H3 . Then H3′ is a caterpillar since K1,3∗ is not

a subgraph of H. Therefore, H3′ has a path x1 x2 ...x2k+1 such that every leaf of H3′ is adjacent to
some x2i+1 . We assume that H3′ is not a path because otherwise we may move the entire path from
E3 to E1 . We make two observations before we continue. First, dH (v) > 1 holds for every leaf v
of H3′ , because otherwise the only edge of H that is incident with v would be the only edge of H3′
(as v ∈ S). Second, if u, v ∈ V (H3′ ) are of degree-2 in H and are contained in a 4-cycle uxvy of
H, then at most one of u, v is in S. This is because otherwise µ(ux) = µ(vy), µ(uy) = µ(vx), and

15
both x, y ∈ S, which implies that H3′ is a subgraph of the 4-cycle uxvy. It follows from these two
observations and the construction of graphs in C that each x2i+1 is adjacent to at most two leaves
of H3′ . For the same reasons, there must exist i0 ∈ {0, 1, ..., k} such that dH3′ (x2i0 +1 ) = 2.
For i = 1, 2, 3, 4, let Vi = {v : dH3′ (v) = i}. Note that V2 ∪ V3 ∪ V4 = {x1 , ..., x2k+1 }. Let
M be the matching {x2i−1 x2i : i = 1, ..., i0 } ∪ {x2i x2i+1 : i0 + 1, ..., k}. From H3′ we move M
to E2 and the rest of E(H3′ ) to E1 . Now we verify that, after this change, dH1 (v) ≤ ⌈∆(µ)/2⌉
and dH2 (v) ≤ ⌊∆(µ)/2⌋ hold for all v ∈ V1 ∪ V2 ∪ V3 ∪ V4 . For each v ∈ V1 it is easy to see
that in fact dH1 (v) = ⌈∆(µ)/2⌉ and dH2 (v) = ⌊∆(µ)/2⌋. For each even i, we have xi ∈ V2 and
dH1 (xi ) = dH2 (xi ) = dµH (xi )/2 ≤ ⌊∆(µ)/2⌋. For each odd i we consider two cases. If xi ∈ V3 then
dH1 (xi ) = (dµH (xi ) + 1)/2 ≤ ⌈∆(µ)/2⌉ and dH2 (xi ) = (dµH (xi ) − 1)/2 ≤ ⌊∆(µ)/2⌋. If xi ∈ V2 ∪ V4
then dH1 (xi ) ≤ (dµH (xi ) + 2)/2 ≤ ⌈∆(µ)/2⌉ and dH2 (xi ) ≤ dµH (xi )/2 ≤ ⌊∆(µ)/2⌋. Therefore, we
may apply this split to all components of H3 and create the required partition E1 , E2 .

Proof of Theorem 6.1. The forward implication is obvious so we only show that G = L̄(H) is
ESP when G is perfect and {S3 , S̄3+ }-free.
Suppose L(H) contains an induced S3 . Then H contains S̄3 as a subgraph. By Lemma 6.3,
si(H) is either K4+ or a subgraph of K2,n+
for some n ≥ 3. In both cases, it is straightforward to
verify that L̄(si(H)) satisfies the assumptions in Lemma 6.5. So L̄(si(H)) is totally unimodular
and thus is also ESP. By Lemma 5.2, L̄(H) is ESP.
Now suppose L(H) is S3 -free. We claim that L̄(H ′ ) is strong ESP for every component H ′ of
H. If L̄(H ′ ) is a comparability graph, then the claim follows immediately from the Remark at the
end of Section 4. So we assume that L̄(H ′ ) is not a comparability graph. By Lemma 6.2, L(H)
contains an induced Γ or C2n (n ≥ 3). This implies, by Lemma 6.4, that si(H ′ ) is a subgraph of a
graph in C ∪ {K3,3 }. Then the claim follows from Lemma 6.7, Lemma 6.6(1), and the Remark of
Lemma 5.2. Finally, this claim and Lemma 6.6(2) imply that L̄(H) is ESP.

7 Trigraphs

Our next objective is to characterize claw-free box-perfect graphs. To accomplish this goal, we will
need a result of Chudnovsky and Plumettaz [9] on the structure of claw-free perfect graphs. The
purpose of this section is to explain their result, which requires many definitions.
A trigraph G consists of a finite set V of vertices and an adjacency function θ : V2 → {1, 0, −1}

such that {uv : θ(uv) = 0} is a matching. Two distinct vertices u and v of G are strongly adjacent
if θ(uv) = 1, strongly antiadjacent if θ(uv) = −1, and semiadjacent if θ(uv) = 0. We call u, v
adjacent if θ(uv) ≥ 0, and antiadjacent if θ(uv) ≤ 0. Note that every graph can be considered
as a trigraph with {uv : θ(uv) = 0} = ∅. In other words, graphs are exactly trigraphs with no
semiadjacent pairs. The result of Chudnovsky and Plumettaz is in fact about trigraphs.
For any trigraph G = (V, θ), let G≥0 denote the graph (V, {uv : θ(uv) ≥ 0}). Conversely, for
any graph G = (V, E), let tri(G) denote the set of all trigraphs (V, θ) such that for any distinct
u, v ∈ V , θ(uv) ≥ 0 if uv ∈ E and θ(uv) ≤ 0 if uv 6∈ E.
Let G = (V, θ) be a trigraph. We call G connected if G≥0 is connected. For each v ∈ V , let

16
NG (v) = NG≥0 (v). We often write N (v) for NG (v) if the dependency on G is clear. For any X ⊆ V ,
let G|X be
the trigraph such that its vertex set is X and its adjacency function is the restriction
of θ to X2 . If a trigraph H is isomorphic to G|X for some X ⊆ V , then we call H a subtrigraph
of G and we say that G contains H.
A trigraph is a hole if it belongs to tri(Cn ) for some n ≥ 4. A trigraph (V, θ) is an antihole if
(V, −θ) is a hole. A hole or antihole is odd if its number of vertices is odd. A trigraph is Berge if
it contains neither odd hole nor odd antihole. A trigraph is a claw if it belongs to tri(K1,3 ). A
trigraph is claw-free if it does not contain any claw. In general, if H is a set of trigraphs, then a
trigraph is H-free if it does not contain any trigraph in H. The result of Chudnovsky and Plumettaz
characterizes {claw, holes, antiholes}-free trigraphs, that is, claw-free Berge trigraphs. To describe
the resulting structure we need more definitions.
Let G = (V, θ) be a trigraph. For any two disjoint X, Y ⊆ V , we say that X is complete
(resp. strongly complete, anticomplete, strongly anticomplete) to Y if every x ∈ X and every y ∈ Y
are adjacent (resp. strongly adjacent, antiadjacent, strongly antiadjacent). A clique (resp. strong
clique) of G is a set C ⊆ V such that any two distinct vertices of C are adjacent (resp. strongly
adjacent). A stable set (resp. strong stable set) of G is a set S ⊆ V such that any two distinct
vertices of S are antiadjacent (resp. strongly antiadjacent).
A trigraph H is a thickening of a trigraph G if V (H) admits a partition (Xv : v ∈ V (G)) such
that
• if v ∈ V (G) then Xv 6= ∅ is a strong clique of H;
• if u, v ∈ V (G) are strongly adjacent in G then Xu is strongly complete to Xv in H;
• if u, v ∈ V (G) are strongly antiadjacent, then Xu is strongly anticomplete to Xv in H;
• if u, v ∈ V (G) are semiadjacent, then Xu is neither strongly complete nor strongly
anticomplete to Xv in H.
Let C be the class of all trigraphs illustrated in Figure 7.1, where
• |Bij | ≤ 1 for all i, j ∈ {1, 2, 3}
• |B21 ∪ B31 |, |B12 ∪ B32 |, |B13 ∪ B23 | ∈ {0, 2}
• if θ(a1 a3 ) = 0 then B21 ∪ B31 = ∅
• there exists xi ∈ Bi1 ∪ Bi2 ∪ Bi3 for i = 1, 2, 3, such that {x1 , x2 , x3 } is a clique.
It turns out that there are two kinds of claw-free Berge trigraphs. The first are thickenings of
trigraphs in C. The second are constructed (in a way like constructing line graphs) from certain basic
trigraphs. In the following, we first define the building blocks and then describe the construction.
Let G have three vertices v, z1 , z2 such that θ(vz1 ) = θ(vz2 ) = 1 and θ(z1 z2 ) = −1. Then the
pair (G, {z1 , z2 }) is a spot. Let G have four vertices v1 , v2 , z1 , z2 such that θ(v1 z1 ) = θ(v2 z2 ) = 1,
θ(v1 v2 ) = 0, θ(z1 z2 ) = θ(z1 v2 ) = θ(z2 v1 ) = −1. Then the pair (G, {z1 , z2 }) is a spring.
A trigraph is a linear interval if its vertices can be ordered as v1 , ..., vn such that if i < j < k and
θ(vi vk ) ≥ 0 then θ(vi vj ) = θ(vj vk ) = 1. Let G be such a trigraph with n ≥ 4. We call (G, {v1 , vn })
a linear interval stripe if: v1 and vn are strongly antiadjacent, vi and vi+1 are adjacent for every
i ∈ {1, ..., n − 1}, no vertex is complete to {v1 , vn }, and no vertex is semiadjacent to v1 or vn .
Let (G, {p, q}) be a spring or a linear interval strip. Let H be a thickening of G and let Xv
(v ∈ V (G)) be the corresponding sets. If |Xp | = |Xq | = 1, then (H, Xp ∪ Xq ) is called a thickening

17
 a2
not shown theta = 1
theta = -1
theta 0

B31 B11 2
B1

B32 B32

2 3
B2 B3
1 1
B2 B3

a3 a1

Figure 7.1: Trigraphs in C

of (G, {p, q}).

Let C ′ be the class of all pairs (H, {z}) such that H is a thickening of a trigraph G ∈ C and
i+2
z ∈ Xai for some i ∈ {1, 2, 3} for which Bi+1 ∪ Bii+2 = ∅ and N (z) ∩ (Xai+1 ∪ Xai+2 ) = ∅ (here we
use the notation from the definitions of C and thickening).

P A signed graph (G, s) consists of a multigraph G = (V, E) and a function s : E → {0, 1}. If
e∈E(C) s(e) is even for all cycles C of G, then (G, s) is an evenly signed graph. In the following
we define another three classes of signed graphs. For any F ⊆ E, let G[F ] = (V, F ).
Let F1 be the class of loopless signed graphs (G, s) such that si(G) = K4 and s ≡ 1. Let F2 be
the class of loopless signed graphs (G, s) such that si(G) is obtained from K2,n (n ≥ 1) by adding
an edge e∗ between its two degree-n vertices, and edges in {e : s(e) = 0} are all parallel to e∗ (while
s(e∗ ) = 1). We remark that our F1 is F2 of [9] and our F2 is F1 ∪ F3 of [9]
In a connected multigraph G with E(G) 6= ∅, a subgraph B is a block of G if B is a loop or
B is maximal with the property that B is loopless and si(B) is a block of si(G). A signed graph
(G, s) is called an even structure if E(G) 6= ∅ and for all blocks B of G, (B, s|E(B) ) is a member of
F1 ∪ F2 or an evenly signed graph or a loop.
Now we describe how the pieces defined above can be put together. A trigraph G = (V, θ) is
called an evenly structured linear interval join if it can be constructed in the following manner:
• Let (H, s) be an even structure.
• For each edge e ∈ E(H), let Ze ⊆ V (H) be the set of ends of e (so |Ze | = 1 or 2).
Let Se = (Ge , Ze ) such that Ge is a trigraph with V (Ge ) ∩ V (H) = Ze and
∗ if e is not on any cycle then Se is a spot or a thickening of a linear interval stripe,
∗ if e is on a cycle of length > 1 and s(e) = 0 then Se is a thickening of a spring,
∗ if e is on a cycle of length > 1 and s(e) = 1 then Se is a spot,
∗ if e is a loop then Se ∈ C ′ .
• For all distinct e, f ∈ E(H), V (Ge ) ∩ V (Gf ) ⊆ Ze ∩ Zf .
• Let V = ∪e∈E(H) V (Ge )\Ze and let θ be given by: for any u, v ∈ V
∗ if u, v ∈ V (Ge )\Ze for some e ∈ E(H) then θ(uv) = θGe (uv)

18
 ∗ if u ∈ NGe (x) and v ∈ NGf (x) for distinct e, f ∈ E(H) with a common end x, then θ(uv) = 1
∗ in all other cases, θ(uv) = −1.
• We will write G = Ω(H, s, {Se : e ∈ E(H)}).
Theorem 7.1 (Chudnovsky and Plumettaz [9]). A connected trigraph is claw-free and Berge if
and only if it is a thickening of a trigraph in C or an evenly structured linear interval join.

In the following we produce a different formulation of this result. A vertex x of a trigraph is

simplicial if N (x) 6= ∅ and {x} ∪ N (x) is a strong clique. For i = 1, 2, let Gi = (Vi , θi ) be a trigraph
with a simplicial vertex xi and with |Vi | ≥ 3. The simplicial sum of G1 , G2 (over x1 , x2 ) is the
trigraph G = (V, θ) such that V = (V1 − x1 ) ∪ (V2 − x2 ) and, for all distinct v1 , v2 ∈ V ,
• θ(v1 v2 ) = θi (v1 v2 ) if {v1 , v2 } ⊆ Vi for some i = 1, 2
• θ(v1 v2 ) = 1 if vi ∈ NGi (xi ) for both i = 1, 2
• θ(v1 v2 ) = −1 if otherwise.
We point out that both G1 and G2 are contained in G. Moreover, using the language of [9], G
admits either a 1-join or a homogeneous set of size ≥ 2.
Lemma 7.2. Let G be a simplicial sum of G1 , G2 . Then G is claw-free if and only if both G1 , G2
are; and G is Berge if and only if both G1 , G2 are.

We omit the proof since it is straightforward. This lemma suggests that we can characterize
claw-free Berge trigraphs by determining all such trigraphs that are not simplicial sums. In the
following we describe these trigraphs.
Let I be the class of linear interval trigraphs. Let L be the class of trigraphs G such that G≥0 is
the line graph of a bipartite multigraph and every triangle (a clique of size 3) of G is a strong clique.
Let J1 be the first graph in Figure 7.2. We consider J1 as a trigraph with no semiadjacent pairs.
Let J1 consists of trigraphs obtained from J1 by deleting k of its cubic vertices (0 ≤ k ≤ 4). Let
J2 (n) be the second trigraph in Figure 7.2, where Q1 , Q2 , and all vertical triples are strong cliques,
θ(uv) could be 0, 1, or −1, and all other pairs are strongly antiadjacent. Note that J2 (0) ∈ I. Let
J2 consist of trigraphs of the form J2 (n) − X for all n ≥ 1 and all X ⊆ {u, v}. Let J = J1 ∪ J2 .

Q1
u
1 2 n

v
Q2

Figure 7.2: J1 and J2

Theorem 7.3. A connected trigraph is claw-free and Berge if and only if it is obtained by simplicial
summing thickenings of trigraphs in C ∪ L ∪ I ∪ J .

We need a few lemmas in order to prove this theorem. A 1-separation of a multigraph H is a pair
(H1 , H2 ) of edge-disjoint proper subgraphs of H such that H1 ∪ H2 = H and |V (H1 ) ∩ V (H2 )| = 1.

19
Suppose G = Ω(H, s, {Se }). Then a 1-separation (H1 , H2 ) of H is called trivial if there exists
i ∈ {1, 2} such that Hi = K2 and Sf is a spot, where f is the only edge of Hi .

Lemma 7.4. Suppose G = Ω(H, s, {Se }) and suppose H has a nontrivial 1-separation (H1 , H2 ).
Then G is a simplicial sum of two trigraphs.

Proof. Let x be the common vertex of H1 , H2 . For i = 1, 2, let Hi′ be obtained from Hi by adding
a new vertex xi and a new edge xxi . Let si be the signing of Hi′ which agrees with s on Hi , and
si (xxi ) = 1. Since all blocks of Hi′ (other than xxi ) are blocks of H, (Hi′ , si ) is an even structure.
Let Sxxi be a spot and let Gi = Ω(Hi′ , si , {Se : e ∈ E(Hi′ )}). Since separation (H1 , H2 ) is nontrivial,
Gi must have ≥ 3 vertices. Now it is straightforward to verify that xi is a simplicial vertex of Gi
(i = 1, 2) and G is the simplicial sum of G1 and G2 over x1 and x2 .

Lemma 7.5. Let H be a thickening of G.

(i) H is claw-free if and only if G is claw-free.
(ii) H is Berge if and only if G is Berge.

Proof. Part (ii) is (6.4) of [9] and part (i) is easy to verify, as pointed out in [11].
A trigraph G is quasi-line if N (v) is the union of two strong cliques for every v ∈ V (G). It is
easy to see that if G is quasi-line then G is claw-free. A trigraph G is cobipartite if V (G) is the
union of two strong cliques. Clearly, if G cobipartite then G is quasi-line and thus is claw-free. It is
also clear that every connected cobipartite trigraph with ≥ 2 vertices is a thickening of a two-vertex
trigraph. Thus every cobipartite trigraph is Berge.

Proof of Theorem 7.3. To prove the backward implication, by Lemma 7.2 and Lemma 7.5, we
only need to consider trigraphs G ∈ C ∪ L ∪ I ∪ J . If G ∈ C then the result follows from Theorem
7.1. If G ∈ I then G is claw-free [11] and Berge [9]. If G ∈ L ∪ J then G is quasi-line and thus G is
claw-free. If G ∈ J , then deleting simplicial vertices from G results in a cobipartite trigraph, which
implies that G is Berge. Finally, assume G ∈ L and G≥0 = L(B) is the line graph of a bipartite
multigraph B. We need to show that G is Berge. Since no semiadjacent pairs are contained in a
triangle, every hole of G must come from a cycle of B and thus G contains no odd holes. If G has an
antihole v1 v2 ...vn v1 with n ≥ 7, then we consider the restriction of G on v1 , ..., v6 . If θ(vi vi+1 ) = −1
for all i = 1, ..., 5, then the graph X formed by {vi vj : θ(vi vj ) ≥ 0} would be the complement of a
path on six vertices, which is one of the minimal non-line-graphs. This is impossible since X is an
induced subgraph of L(B). So θ(vi vi+1 ) = 0 holds for some i, which makes vi , vi+1 , xk a triangle
for some k. This contradiction (two semiadjacent vertices are contained in a triangle) shows that
G contains no antihole of length ≥ 7. Thus G is Berge, which completes the proof of the backward
direction.
To prove the forward implication, by Theorem 7.1, we assume G = Ω(H, s, {Se }). Since G is
connected, H is connected as well. By Lemma 7.4, we also assume that all 1-separations of H are
trivial. Let U be the set of all degree-one vertices u of H for which if e is the only edge incident
with u then Se is a spot. We assume V (H) 6= U because otherwise H = K2 and G = K1 and thus
the result holds. Let H0 = H − U . Note that H0 is connected, as H is connected. Moreover, by its
construction, H0 dose not have a 1-separation. Thus either H0 = K1 or H0 is a block of H.
Suppose H0 is K1 or K2 . It follows that H is a tree with 1, 2, or 3 edges. Moreover, Se is a

20
thickening of a linear interval strip for at most one e, and every other Se is a spot. In all cases, it
is routine to check that G is a thickening of a trigraph in I.
Suppose H0 is a loop e. Let Se = (Ge , {z}) ∈ C ′ and let Ge be a thickening of C ∈ C. If H has
≥ 2 edges then H consists of e and a pendent edge f with Sf a spot. It follows that G = Ge , which
is a thickening of a trigraph in C. So e is the only edge of H and G = Ge − z. If z is not the only
vertex of Xai (here we use the notation in the definition of C ′ ) then G is also a thickening of C. If
z is the unique vertex of Xai then G is cobipartite. In this case G is a thickening of a two-vertex
trigraph and thus G is a thickening of a trigraph in I.
Suppose none of the last two cases occurs. Then H0 is a block in which every edge is on a cycle
of length ≥ 2. Let s0 be the restriction of s on H0 . Then (H0 , s0 ) is either in F1 ∪ F2 or evenly
signed. First we assume (H0 , s0 ) is evenly signed. Then (H, s) is also evenly signed. Moreover, Se
is a thickening of a spring for every edge in E0 = {e ∈ E(H0 ) : s(e) = 0}, and Se is a spot for every
other edge of H. Let Se′ be a spring for each e ∈ E0 and let Se′ = Se for every other edge of H.
Then G is a thickening of G′ = Ω(H, s, {Se′ }). Now we only need to show that G′ ∈ L. Let H ′ be
obtained from H by subdividing each edge in E0 exactly once. Then H ′ is bipartite. It follows from
the construction of Ω that adjacent pairs of G′ are exactly adjacent pairs of the line graph L(H ′ ).
In addition, all semiadjacent pairs of G′ come from a spring, and thus no such pair is contained in
a triangle. Therefore, G′ belongs to L, as required.
It remains to consider the case (H0 , s0 ) ∈ F1 ∪ F2 . If (H0 , s0 ) ∈ F1 , then H is obtained from
K4 by adding parallel edges and adding pendent edges to distinct vertices. Moreover, every Se is
a spot. It follows that G is an ordinary graph (meaning that G has no semiadjacent pairs) and
this graph is exactly L(H). Now it is clear that G is a thickening of L(si(H)), which belongs to
J1 . So we assume (H0 , s0 ) ∈ F2 . Let V (H0 ) = {x1 , x2 , y1 , ..., ym } (m ≥ 1) such that xi (i = 1, 2)
is adjacent to all other vertices. Like before, we assume that H0 has no parallel edges, except for
two possible edges e0 , e1 between x1 , x2 , and such that s(e0 ) = 0 and s(e1 ) = 1. We also assume
that Se0 is a spring, if e0 is present. Suppose H is obtained by adding pendent edges to y1 , ..., yn
(n ≥ 0) and to k of x1 , x2 (0 ≤ k ≤ 2). If e0 is present, then G is a thickening of J2 (n), where
θ(uv) = 0. So assume that e0 is not in H, and thus G = L(H). For i = 1, 2, let Qi be the clique of
G formed by edges of H incident with xi . Let Q′i = Qi − {x1 x2 , xi y1 , ..., xi yn }. If Q′1 6= ∅ is neither
complete nor anticomplete to Q′2 6= ∅, then again G is a thickening of J2 (n) with θ(uv) = 0. In the
remainder cases (which are: some Q′i is empty, or Q′1 6= ∅ is complete or anticomplete to Q′2 6= ∅),
if n = 0 then G is a thickening of K3 , and if n ≥ 1 then G = J2 (n) − X for some X ⊆ {u, v}.

8 Claw-free box-perfect graphs

In this section we prove the following.

Theorem 8.1. A claw-free perfect graph is box-perfect if and only if it is S3 -free.

We divide the proof into several lemmas. Let G be a trigraph. We call G a sun if G ∈ tri(S3 ).
We call G an incomparability trigraph if G≥0 is an incomparability graph. We call G elementary if it
is a thickening of a trigraph in L. We remark that when an elementary trigraph has no semiadjacent
pairs then they are exactly elementary graphs discussed in [18].

21
Lemma 8.2. Let G be a connected Berge trigraph. If G is {claw, sun}-free then G is obtained by
simplicial summing incomparability trigraphs and elementary trigraphs.

Proof. Since G is connected, Berge, and claw-free, by Theorem 7.3, G is obtained by simplicial
summing thickenings of trigraphs in C ∪ L ∪ I ∪ J . Therefore, we may assume that G is a thickening
of a trigraph G0 ∈ C ∪ L ∪ I ∪ J . If G0 ∈ L then G is elementary and we are done. If G0 ∈ C then
G0 |{a1 , a2 , a3 , x1 , x2 , x3 } (here we are using the notation in the definition of C) is a sun and thus
G contains a sun, which is impossible. So we assume that G0 ∈ I ∪ J . In the following we prove
that G is an incomparability trigraph.
Suppose G0 ∈ I. Then vertices of G0 can be ordered as v1 , ..., vn such that if i < j < k and
θ0 (vi vk ) ≥ 0 then θ0 (vi vj ) = θ0 (vj vk ) = 1. Using the notation in the definition of thickening, we
let Xvi = {xi,j : j = 1, ..., ni } (1 ≤ i ≤ n). Now we define a binary relation ≺ on V (G) such that
xi1 ,j1 ≺ xi2 ,j2 if θ(xi1 ,j1 xi2 ,j2 ) = −1 and (i1 , j1 ) is lexicographically smaller than (i2 , j2 ). We claim
that ≺ is transitive. Suppose xi1 ,j1 ≺ xi2 ,j2 ≺ xi3 ,j3 . Since each Xvi is a strong clique, we must
have i1 < i2 < i3 . It follows that θ0 (vi1 vi2 ) ≤ 0 and θ0 (vi2 vi3 ) ≤ 0. As a result, θ0 (vi1 vi3 ) = −1 and
thus θ(xi1 ,j1 xi3 ,j3 ) = −1, which proves our claim. This claim implies that the complement of G≥0
is the comparability graph of poset (V (G), ≺), which proves that G≥0 is an incomparability graph
and thus G is an incomparability trigraph.
Now suppose G0 ∈ J . We claim that G0 is a thickening of a trigraph in I. This claim clearly
implies that G is a thickening of trigraph in I, and thus the last paragraph proves that G is an
incomparability trigraph.
Before proving the claim we make an observation. It is clear that every cobipartite trigraph is
a thickening of a trigraph that has exactly two vertices and that the two vertices are semiadjacent.
Since this two-vertex trigraph is in I, our claim holds if G0 is cobipartite.
We first consider the case G0 ∈ J1 . If G0 has two or more cubic vertices then G0 contains an
induced S3 . So G0 contains at most one cubic vertex and in this case G0 is cobipartite. Next we
assume G0 ∈ J2 and let G0 = J2 (n) − X (see the definition of J2 ). Let xi yi zi (1 ≤ i ≤ n) denote
the vertical triangles of J2 (n), where yi ∈ Q1 and zi ∈ Q2 . If n ≥ 3 then G0 |{w, x1 , y1 , z1 , y2 , z3 } is
a sun. So we have n ≤ 2. Now it is straightforward to verify that either G0 is cobipartite, or G0
contains a sun (found in a similar way), or G0 = J2 (2) − {u, v}. In the last case, G0 is a thickening
of the trigraph G∗ = ({t1 , t2 , t3 , t4 , t5 }, θ ∗ ), where θ ∗ (ti ti+1 ) = 1 (i = 1, 2, 3, 4), θ ∗ (t2 t4 ) = 0, and
θ ∗ (ti tj ) = −1 for all other pairs. This completes the proof of our claim and also completes the
proof of the lemma.
Although simplicial sum was defined for trigraphs, this operation can be naturally inherited by
ordinary graphs. Moreover, we have the following.

Lemma 8.3. The simplicial sum of two ESP graphs is ESP.

Proof. Let G be the simplicial sum of G1 and G2 over x1 and x2 , where Gi is an ESP graph with a
simplicial vertex xi for i = 1, 2. Let Λ be a set of maximal cliques of G. Note that NG1 (x1 )∪NG2 (x2 )
is the only maximal clique of G that contains edges between NG1 (x1 ) and NG2 (x2 ). For i = 1, 2, let
Λi consist of members of Λ that are cliques in Gi . Since Gi is ESP, Λi has an equitable subpartition
(Λi1 , Λi2 ). If Λ does not contain the clique NG1 (x1 ) ∪ NG2 (x2 ), then (Λ11 ∪ Λ21 , Λ12 ∪ Λ22 ) is clearly
an equitable subpartition of Λ. Now assume that Λ contains NG1 (x1 ) ∪ NG2 (x2 ). For i = 1, 2, let

22
Λ′i = Λi ∪{{xi }∪NGi (xi )}. Note that dΛ′i (xi ) = 1. Since Gi is ESP, Λ′i has an equitable subpartition
(Λ′i1 , Λ′i2 ). Without loss of generality, suppose dΛ′i1 (xi ) = 1 and dΛ′i2 (xi ) = 0 for i = 1, 2. Let Λ′ be
obtained from Λ′11 ∪ Λ′21 by replacing the two cliques containing x1 or x2 by NG1 (x1 ) ∪ NG2 (x2 );
and set Λ′′ = Λ′12 ∪ Λ′22 . Then (Λ′ , Λ′′ ) is an equitable subpartition of Λ.

Lemma 8.4. Let Λ be a set of cliques of a graph G, for which V (G) is partitioned into two cliques
X, Y . Then G has a multiset Λ′ of cliques such that
(i) |Λ′ | = |Λ| and dΛ′ (v) = dΛ (v), for all v ∈ V (G);
(ii) members of Λ′ can be enumerated as Q1 , ..., Q|Λ| such that every v ∈ X appears in the first
dΛ (v) terms and every v ∈ Y appears in the last dΛ (v) terms.

Proof. For each i = 1, ..., |Λ|, let Xi = {x ∈ X : i ≤ dΛ (x)} and Yi = {y ∈ Y : i ≥ |Λ| − dΛ (y) + 1}.
Then for every i, Qi = Xi ∪ Yi is a clique since dΛ (x) + dΛ (y) ≤ |Λ| holds for all non-adjacent x ∈ X
and y ∈ Y . Now it is clear that Λ′ = {Q1 , ..., Q|Λ| } satisfies the requirements.

Lemma 8.5. Elementary graphs are ESP.

Proof. Let elementary graph H be obtained by thickening a trigraph G, where G≥0 is the line
graph of a bipartite multigraph B and such that semiadjacent pairs of G are not contained in
any triangle. Let (Z1 , Z2 ) be a partition of V (B) into two stable sets. Let u1 v1 , ..., un vn be the
semiadjacent pairs of G. Let (Xv : v ∈ V (G)) be the partition of V (H) over which G is thickened.
For i = 1, ..., n, let Hi = H[Xui ∪ Xvi ]. Since no semiadjacent pairs of G are contained in a triangle,
it is easy to see that for each maximal clique C of H, either C is a maximal clique of some Hi or
C = ∪{Xv : v ∈ Q} for some maximal strong clique Q of G. On the other hand, since G≥0 = L(B),
for each maximal clique Q of G there exists a vertex z of B such that members of Q are precisely
edges of B that are incident with z. We will say that Q and C = ∪{Xv : v ∈ Q} come from z.
Note that, if Q, Q′ are maximal cliques of G with ui ∈ Q − Q′ and vi ∈ Q′ − Q for some i, then Q
and Q′ come from vertices that both belong to Z1 or both belong to Z2 .
Let Λ be a set of maximal cliques of H. We need to show that Λ admits an equitable subpartition.
For i = 1, ..., n, let Λ(i) = {C ∈ Λ : C ⊆ V (Hi )}. Let Λ(0) = Λ − Λ(1) ... − Λ(n) . We assume by
(i) (i)
Lemma 8.4 that members of each Λ(i) are enumerated as C1 , ..., Cni such that every x ∈ Xui
appears in the first dΛ(i) (x) terms and every x ∈ Xvi appears in the last dΛ(i) (x) terms. In the
following we define a partition (Λ1 , Λ2 ) of Λ. To verify that (Λ1 , Λ2 ) is an equitable subpartition
of Λ we only need to verify min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋, for all x ∈ V (H).
We first consider Λ(0) . If C ∈ Λ(0) then C comes from a vertex z of B. In this case we put C
into Λi if z ∈ Zi (i = 1, 2). Since each v ∈ V (G) is contained in at most two maximal cliques, we
deduce that dΛ (x) ≤ 2 for all x ∈ V (H) − V (H1 ) − ...V (Hn ). For these x, our partition of Λ(0)
guarantees that min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋.
For cliques in each Λ(i) (i = 1, 2, ..., n) we consider three cases. If none of Xui , Xvi is contained
in any clique of Λ(0) , then dΛ (x) = dΛ(i) (x) for all x ∈ V (Hi ). Moreover, for each x ∈ V (Hi ), since
(i) (i) (i)
cliques containing x appear consecutively in the sequence C1 , ...., Cni , putting Cj into Λ1 for all
(i)
odd j and putting Cj into Λ2 for all even j lead to min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋.

If exactly one of Xui , Xvi is contained in a clique of Λ(0) , we assume by symmetry that Xui is
contained in a clique C ∈ Λ(0) . We also assume without loss of generality that C has been placed

23
into Λ2 . For each x ∈ V (Hi ), since cliques containing x appear consecutively in the sequence
(i) (i) (i) (i)
Xui , C1 , ...., Cni , putting Cj into Λ1 for all odd j and putting Cj into Λ2 for all even j lead to
min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋.
If both Xui , Xvi are contained in cliques, say C, D, of Λ(0) , by discussion in the first paragraph
of this proof, we assume that both C, D have been placed into Λ2 . We consider two subcases.
Suppose ni is odd. For each x ∈ V (Hi ), since cliques containing x appear consecutively in the
(i) (i) (i) (i)
sequence Xui , C1 , ...., Cni , Xvi , putting Cj into Λ1 for all odd j and putting Cj into Λ2 for all
even j lead to min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋. Now suppose ni is even. For each x ∈ V (Hi ),
(i) (i) (i)
note that cliques containing x appear consecutively in the sequence C1 , Xui , C2 , ...., Cni , Xvi ,
(i) (i)
unless x ∈ C1 ∩ Xvi . In this case we put Cj into Λ2 for all odd j > 1 and we put the rest into
(i)
Λ1 . For each x ∈ V (Hi ) − (C1 ∩ Xvi ), it is clear that min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋. For each
(i)
x ∈ C1 ∩ Xvi , we have dΛ (x) = 1 + ni . Our partition yields dΛ2 (x) = ni /2, which also leads to
min{dΛ1 (x), dΛ2 (x)} ≥ ⌊dΛ (x)/2⌋.

Proof of Theorem 8.1. The forward implication is clear, so we only need to consider the backward
implication. Let G be perfect and {claw, S3 }-free. By Lemma 8.2, each component of G is obtained
by simplicial summing incomparability graphs and elementary graphs. By Theorem 4.2 and Lemma
8.5, incomparability graphs and elementary graphs are ESP. Thus G is ESP by Lemma 8.3, which
proves that G is box-perfect by Theorem 3.7.

References
[1] C. Berge, The q-perfect graphs. II, Matematiche (Catania) 47 (1992), 205-211.
[2] M. Burlet and J. Uhry, Parity graphs, Ann. of Discrete Math. 66 (1982), 1-26.
[3] K. Cameron, Polyhedral and Algorithmic Ramifications of Antichains, Ph.D. Thesis, Univer-
sity of Waterloo, 1982. (Supervisor: J. Edmonds)
[4] K. Cameron, A min-max relation for the partial q-colourings of a graph. II. Box perfection,
Discrete Math. 74 (1989), 15-27.
[5] K. Cameron and J. Edmonds, Coflow polyhedra, Discrete Math. 101 (1992), 1-21.
[6] P. Camion, Characterization of totally unimodular matrices, Proc. Amer. Math. Soc. 16 (1965),
1068-1073.
[7] X. Chen, G. Ding, and W. Zang, The box-TDI system associated with 2-edge connected
spanning subgraphs, Discrete Appl. Math. 157 (2009), 118-125.
[8] M. Chudnovsky, G. Cornuéjols, X. Liu, P. Seymour, and K. Vušković, Recognizing Berge
graphs, Combinatorica 25 (2005), 143-186.
[9] M. Chudnovsky and M. Plumettaz, The structure of claw-free perfect graphs, J. Graph Theory
75 (2014), 203-230.
[10] M. Chudnovsky, N. Robertson, P. Seymour, and R. Thomas, The strong perfect graph theorem,
Ann. of Math. (2) 164 (2006), 51-229.

24
[11] M. Chudnovsky and P. Seymour, Claw-free graphs. V. Global structure, J. Combin. Theory
Ser. B 98 (2008), 1373-1410.
[12] M. Chudnovsky and P. Seymour, Even pairs in Berge graphs, J. Combin. Theory Ser. B 99
(2009), 370-377.
[13] W. Cook, On box totally dual integral polyhedra, Math. Programming 34 (1986), 48-61.
[14] J. Edmonds and R. Giles, A min-max relation for submodular functions on graphs, Ann. of
Discrete Math. 1 (1977), 185-204.
[15] T. Gallai, Transitiv orientierbare graphen, Acta Math. Acad. Sci. Hungar. 18 (1967), 25-66.
[16] L. Lovász, Normal hypergraphs and the perfect graph conjecture, Discrete Math. 2 (1972),
253-267.
[17] L. Lovász, Perfect graphs, in: Selected Topics in Graph Theory II (L. Beineke and R. Wilson,
eds.), Academic Press, 1983, pp. 55-87.
[18] F. Maffray and B. Reed, A description of claw-free perfect graphs, J. Combin. Theory Ser.
B 75 (1999), 134-156.
[19] H. Sachs, On the Berge conjecture concerning perfect graphs, in: Combinatorial Structures
and Their Applications, Gordon and Breach, New York, 1970, pp. 377-384.
[20] A. Schrijver, Min-max results in combinatorial optimization, in: Mathematical Programming:
The State of the Art, Springer, Berlin, 1983, pp. 439-500.
[21] A. Schrijver, Theory of Linear and Integer Programming, John Wiley & Sons, New York, 1986.
[22] A. Schrijver, Combinatorial Optimization: Polyhedra and Efficiency, Springer-Verlag, Berlin,
2003.

25