Discrete Applied Mathematics 8 (1981) 295-305 295

North-Holland

RECOGNIZING GRAPHS WITH FIXED INTERVAL

NUMBER IS NP-COMPLETE

Douglas B. WEST
University of Illinois, Urbana, IL 61801, USA

David B. SHMOYS*
University of California, Berkeley, CA 94720, USA

Received 20 August 1982

Revised I June 1983

A t-interval representation of a graph expresses it as the intersection graph of a family of

subsets of the real line. Each vertex is assigned a set consisting of at most t disjoint closed inter-
vals, in such a way that vertices are adjacent if and only if some interval for one intersects some
interval for the other. The interval number i(G) of a graph G is the smallest number t such that
G has a t-representation. We prove that, for any fixed value of t with t ~2, determining whether
i(G)5 t is NP-complete.

1. Introduction

The notion of the interval number of a simple undirected graph (henceforth

‘graph’) generalizes the concept of interval graphs. A graph on n vertices is an inter-
val graph if it is the intersection graph of a collection of finite intervals on the real
line. In other words, each vertex is assigned a real interval such that two vertices
of G are adjacent if and only if the corresponding intervals intersect.
To generalize this concept, we may assign more than one interval to each vertex.
That is, given an integer t, we say that G has a t-interval representation (or simply
a t-representarion) if to each vertex of G we can assign at most i disjoint closed inter-
vals such that two vertices u and w are adjacent in G if and only if some interval
for u intersects some interval for w. Without loss of generality we assume that dif-
ferent intervals do not share endpoints. The interval number of G, denoted i(G),
is the smallest integer t such that G has a t-interval representation. Thus the interval
graphs are the graphs G with i(G)< 1.
Several authors have remarked that interval graphs and interval numbers are use-
ful in scheduling and allocation problems. Here we describe an application where
‘multiple-interval’ graphs may be more appropriate than interval graphs. Benzer [l]
used interval graphs to discuss the linear arrangement of nucleotides in genes. Many
genes are encoded in a single molecule of DNA and each gene contains information
*This work was supported by an NSF Fellowship, the Eugene C. Gee and Mona Fay Gee Scholarship,
NSF Grant No. iMCS78-20054 and DARPA Grant No. NOOO39-82-C-0235.

0166-218X/84/S3.00 0 1984, Elsevier Science Publishers B.V. (North-Holland)

296 D.B. West. D.B. Shmqvs

for the assembly of a specific protein. DNA is a linear molecule composed of a

sequence of sub-units, each of which contains one of four possible nucleotides. The
specific sequence of these nucleotides encodes the genetic information. For many
years biologists thought that the nucleotides encoding a single gene appear in an un-
broken sequence, covering a single ‘interval’ on the chain of nucleotides. Recently,
Chambon [3] and others have shown that many genes are not represented as single
unbroken sequences but rather as a collection of unbroken sequences on the DNA
strand. This is analogous to a family of intervals on the real line. Thus, interval
numbers may be a useful tool for studying the structure of genes; if an upper bound
can be given on the number of intervals used for each gene, then finding interval
numbers can be used to test gene compositions.
Bounds on the interval number in terms of other parameters of the graph appear
in [5], [6], [7], [S], and [ 111. Several of those papers pose the related problems of
characterizing graphs with interval number t (even for t = 2) and finding an efficient
algorithm for computing the interval number. In this paper, we show that for any
fixed value of t with t 22, determining whether i(G)5 t is NP-complete. To do this,
we reduce the problem of recognizing triangle-free 3-regular graphs having a
Hamiltonian circuit to the problem of recognizing graphs having a 2-representation.
Then we reduce the (t - 1)-representation problem to the f-representation problem.
(The t-representation problem is similar in some respects to problems proved
NP-complete in [9] and [lo]. However, there does not seem to be any immediate
relationship among these problems.)
To put this result in perspective, we note that there exists a linear time algorithm
for the case t = 1, which is the problem of interval graph recognition [2]. Our result
then implies that, for the class of recognition problems parametrized by a fixed
value of the interval number, the boundary between P and NP-complete has been
completely determined. This NP-completeness result also implies that a ‘nice’
characterization of t-interval graphs is unlikely to exist for any t 5 2.

2. A preliminary reduction

We wish to reduce from the Hamiltonian circuit problem in 3-regular triangle-

free graphs to the 2-representation problem. It is well known that testing for a
Hamiltonian circuit in a 3-regular graph (also called a cubic graph) is NP-complete,
as listed in [4]. To show that this remains true when triangles are forbidden from
the graph, we provide a preliminary reduction.

Hamiltonian Circuit in Cubic Graph (3HC)

Instance: A graph G(K!E) in which every vertex has three neighbors.
Question: Does G have a Hamiltonian circuit?

Hamiltonian Circuit in Triangle-free Cubic Graph (d3HC)

 Interval number is NP-complere 297

Instance: A graph G = (V,E) in which every vertex has three neighbors but no
three vertices form a cycle.
Question: Does G have a Hamiltonian circuit?

Lemma 1. d3HC is NP-complete.

Proof. We reduce 3HC to d3HC. Given an instance G of 3HC, we construct in

polynomial time a triangle-free cubic graph G’ such that G’ has a Hamiltonian
circuit if and only if G has a Hamiltonian circuit.
Begin by examining all the triples of vertices in the graph to determine which ones
induce triangles. This can be done in 0(n3)-time. By looking at the remaining
neighbors of the vertices in each triangle, we can determine whether or not the
triangle shares an edge with another triangle. If so, we call such a triangle matched;
otherwise it is ‘isolated’. If G is 3-regular and G#K,, then every triangle must be
matched or isolated.
To form G’, replace each isolated triangle by bisecting its edges with new vertices
and adding a fourth new vertex joined to the new vertices on the three old edges.
For each matched pair of triangles, replace one of the two triangles as above. These
local replacements are shown in Fig. 1. No new triangles are created, so G’ is
triangle-free.
To show that G has a Hamiltonian circuit if and only if G’ also does, note that
any Hamiltonian circuit that enters one of the components shown in either graph
must visit all vertices in that component before leaving the component, since it has
at most three connections to the rest of the graph. The heavy lines in Fig. 1 indicate
the only way (up to symmetry) that a Hamiltonian circuit can contain all of the
vertices in one of these components. By making the indicated substitutions we can
transform a Hamiltonian circuit in either graph into a Hamiltonian circuit in the
other. 0

A-A
Fig. 1. Local replacement for triangle-free cubic graphs.

3. Useful facts about interval numbers

Complete bipartite graphs play a significant role in the subsequent reductions. We

refer to the maximal independent sets of a complete bipartite graph as the parts of
the graph. To facilitate discussion of interval representations, we let&u) denote the
298 D.B. West, D.B. Shmo.vs

union of the intervals assigned to v. A f-representation of a graph G is displayed

if each set f(v) contains an open interval disjoint from all other sets f(w). We say
that a t-representation is of depth r if no point on the real line is contained in inter-
vals for r+ 1 of the vertices.
Trotter and Harary [12] showed that i(K,,.) = f(m,n), where t(m,n) =
[(mn + I)/@+ n)l. In fact, they showed that Km,n has a t(m, n)-representation
that is displayed. To obtain an important fact about the representations of certain
bipartite graphs, we give the argument for the lower bound. Since the depth is 2,
when we read the representation from left to right we obtain at most one new edge
at the left endpoint of each interval. No new edge arises if the left endpoint of this
interval is contained in no other interval, such as with the leftmost interval in the
representation. Since there are at most t intervals per vertex, this yields an upper
bound on the number of edges that can be represented: JE Jzs(m + n)t- 1. Hence
i(K,,,.)r(mn + l)/(m+n). If (mn+ l)/(m+n) is an integer t, then we say that
K,,,, is t-tight. To represent all the edges in a t-tight K,,,., the argument above im-
plies that an intersection of intervals must occur at the left endpoint of every interval
after the first. This means that their union is a single closed interval. We say that
such a set of intervals ‘appears contiguously’.

Lemmat. ThegraphK,:+,_,,,+, ist-tight. IfK,z+,_I,I+, isaninducedsubgraphof

a graph G, then in any t-representation of G the intervals for vertices of K[2+t _ 1,f +,
appear contiguously. Furthermore, if u and v are any specified vertices from opposite
parts of K,,,.,,, then K,,,. has a displayed t(m, n)-representation in which u and v are
assigned the leftmost and rightmost intervals in the representation, respectively.

Proof. Since [(t + l)(t’+ t - 1) + l]/(t’+ 2t) = t, KCz+t_I,I+I is t-tight. In any

t-representation of a graph G, the intervals corresponding to the vertices of an
induced subgraph of G must give a t-representation of the subgraph, so the intervals
for K~+r-l,~+~ must appear contiguously.
For the second part of the lemma, take a displayed t-representation of Km,nr as
guaranteed by the Trotter-Harary construction. Note that, for either of the parts
in Km.,,, the leftmost of the intervals assigned to vertices in that part can be ex-
tended leftward to become the leftmost interval in the representation. Therefore, by
permuting vertex labels we can have the leftmost interval assigned to u. We can do
the same to assign the rightmost interval to any vertex v belonging to the other
part. 3

In fact, the lemma remains true if we drop the requirement that u and v belong
to different parts in the bipartite graph. However, this is harder to prove, and our
reductions require only the version stated above. We will make critical use of the
graphs K2+r-l,r+l, such as &,s, to ensure that representations can be constructed
only in certain ways.
 Interval number is NP-complete 299

4. The main reduction

Consider the following decision problem.

f-Interval Representation (tREP)

Instance: A graph G = (V,E).
Question: Does G have a representation as the intersection graph of sets con-
sisting of at most c intervals on the real line?

Theorem 1. 2REP is NP-complete.

Proof. We prove this by reducing d3HC to 2REP. Let G be a triangle-free cubic

graph. We will construct a supergraph G’ of G such that G’ has a 2-representation
if and only if G is Hamiltonian.
First we note that it is easy to construct a 2-representation of a cubic Hamiltonian
graph G. The edges of G can be partitioned into a Hamiltonian cycle and a complete
matching. Suppose the Hamiltonian cycle is uo, ulr . . . , v,, uo. The Hamiltonian path
from u. to u, can be represented contiguously, using one interval per vertex, such
that intervals for consecutive vertices in the path overlap. Complete the Hamiltonian
cycle by using a second interval for uo. Suppose that (u,, oi) is the edge incident to
u. in the complete matching. Extend the second interval for u. to overlap a second
interval for Dj. For any vertex whose matched edge has not been represented, only
one interval has been assigned. Therefore, we represent these remaining edges by
pairs of intervals that intersect each other. This representation, which we call the
H-representation of G, is shown in Fig. 2. Note that a cubic graph has an H-repre-
sentation (for some labeling of the vertices) if and only if it is Hamiltonian.

“0 “I
---
“2 “3 ._. “n “0 “I
E
U
c
. . .
---
W

Fig. 2. The H-representation of a cubic Hamiltonian graph.

For an arbitrary cubic graph, many different 2-representations may be possible.

We want to restrict the representation of G so that it must be an H-representation.
The first step is to consider only triangle-free graphs, which restricts the depth to
two. In addition, we add ‘gadgets’ to G so that any H-representation remains
feasible, but other possible representations are eliminated. More precisely, if G has
an H-representation, then the supergraph G’ will have a 2-representation of G, and
G’ will have no 2-representation in which the intervals for G appear in any other
way.
Roughly speaking, there are two steps to this restriction. First, we want to add
gadgets to separate the intervals for vertices of G into two sets, each containing one
of the intervals for each vertex. We call these the ‘inside’ and ‘outside’ intervals,
for reasons that will soon be apparent. Second, we construct the gadgets so that the
300 D. B. West, D. B. Shmoys

outside intervals can intersect at most one other interval for a vertex in G. This will
force the matching of an H-representation to occur among the outside intervals and
a Hamiltonian cycle to occur among the inside intervals. The reader may wish to
refer to Fig. 3 as we describe the gadgets to be added.

Fig. 3. The transformed graph G’.

The idea behind the gadgets added for the first step is straightforward. Add a
vertex t joined to every vertex of G. This forces every vertex of G to have an interval
intersecting an interval for E. These we call the ‘inside’ intervals. Actually, to en-
force the condition that all 2-representations of G’ contain an H-representation for
G, we want all of these ‘inside’ intervals to lie within a single interval for z. So, we
attach additional gadgets to z to insure that only one of the intervals for z contains
any of the ‘inside’ intervals, and to control what happens at the endpoints of the
intervals in&). The details of this will be described more explicitly below.
To insure that each vertex of G has another interval outsidef(z), we need only
add, for each vertex u E G, some structure M(u) containing a vertex adjacent to IJ
but not to z. We choose a structure M(u) that accomplishes the second step men-
tioned above. In particular, for each UE V(G) let M(o)=&, and join u to one
vertex of M(o). No other edges join M(o) to any other vertices in G’. Note that
K,,, is 2-tight, so if the G’ being constructed has 2-representation, then the inter-
vals of each M(o) appear contiguously. The only other interval that can intersect
these is the outside interval for u. If it lies wholly withinf(M(u)), it cannot be used
for anything else. So, using Lemma 2, we may assume that the outside interval for
u overlaps f(M(u)) at one of its endpoints.
To complete the construction of G’, we describe the gadgets mentioned earlier
that are used to pin down the ways in which the intervals for t can appear. Add three
more copies of K,,, to the graph, called Hi, Hz, and Ws. Join z to two vertices in
the same part of Hi, and join one vertex from the other part of Hi to a vertex in
Hz. Also, place one edge from z to a vertex of H3_ We have described all of the
vertices of G’, and this completes the construction except for a few more edges
needed to maintain feasibility of the H-representations, as will be seen shortly.
To understand the role of the Hi, consider the intervals representing the sub-
graph induced by z and the vertices in the Hi. A 2-representation of G’ must contain
 Interval number is NP-complete 301

a 2-representation of this induced subgraph. The H, are 2-tight, so the intervals in

their representations appear contiguously. Therefore, the edge between H, and Hz
can be represented only by an intersection between end-intervals in the representa-
tions of H, and Hz. Assume that the intervals for Hz appear to the left of those
for H,.
Since z neighbors only two independent vertices in H,, and since each vertex of
HI has already been assigned two intervals, it follows that two intervals for z must
be used to represent its edges to HI. Since one endpoint of f(H,) is already covered
byf(H,), one of the intervals inf(z) must be swallowed in a displayed portion of
its neighbor in H,. Since z has other neighbors in G’ which do not neighbor any-
thing in Hz, the other interval for z must overlap the other end off(H,) and be
displayed. By Lemma 2, this can be arranged, since the neighbors of z in H, are in
the opposite part from the vertex with a neighbor in HI. Finally, H3 is also 2-tight,
and its edge to z can only be represented by having the right end of the displayed
interval for z overlap the left endpoint of f(H3).
Consider the graph G’ constructed from a Hamiltonian graph G. The reader
should refer to Fig. 4, where the 2-representation of G’ obtained from the H-repre-
sentation of G is illustrated. Now, let us check the feasibility of the H-representation
shown, given these gadgets. The configuration of intervals representing z and the
Hi can be inserted to enclose the Hamiltonian cycle part of the H-representation -
almost. We want the second interval for u. to intersect the second interval for its
matched neighbor ui, but the latter is the outside interval for Uj. We add edges
from u. to each vertex in H3, so that this interval for u. can (and must) pass out-
sidef(z) and pasty to intersect the outside interval for uj. Now, M(u,) can on-
ly be added to the H-representation by attaching its (contiguous) representation to
the interval for u. at the other end of the cycle, which is withinf(z). To enable this,
we add edges from z to each vertex of M(uo). Finally, the other M(u)‘s can be
added to the representation before and after the matched pairs of intervals.

r M(vo)
- vo v, vz
. . . -.vn-I vn VO v- Mb) M(u) u w M(w)
-- I o---z ...
>
Hz HI “3
Fig. 4. The 2-representation of G’.

Summarizing the representation, from left to right, we have: (l)f(H2), (2)f(H,)

containing an interval for z, (3) the displayed interval for z containing f(M(u,))
and the Hamiltonian circuit from u. to uo, and (4) f(H,). The second interval for
u. extends throughout f(H3) and intersects the outside interval for its non-circuit
neighbor. Finally, the remaining M(Ui), Ui, Uj,M(Uj) pairs complete the representa-
tion.
We have described all of the vertices and edges added to G to obtain G’, and in
the process we have shown that if G is Hamiltonian, then it has an H-representation
and G’ has a corresponding 2-representation. Note that we have distinguished a
302 D.B. Wesf, D. B. Shmoys

vertex u. by the edges added in the previous paragraph. Since a Hamiltonian

circuit can ‘start’ at any point, it does not matter which vertex we call uo. If G
has n vertices, then it has 3n/2 edges (being 3-regular), and G’ has 9n +25 ver-
tices and 37n/2+65 edges, so the transformation can be performed in polynomial
time.
To complete the proof, we need only show that if G’ has a 2-representation, then
it contains an H-representation of G, and thus G has a Hamiltonian circuit. In moti-
vating the construction we have already shown the required structure of the 2-repre-
sentation for the subgraph induced on t and all the Hi, as well as the structure for
the subgraph induced on just o and M(o) for all vertices u. Since z does not neighbor
the vertices of M(o), and since o#uo does not neighbor any vertices in Hi, the
representations of each of the subgraphs on u and M(u) must lie outside f(z), for
u # uo. Since each vertex of G neighbors t, each f(u) must have an interval inside
the displayed portion off(z). For u#uo, we have forced an outside interval and an
inside interval, but for u. we still have a ‘free’ interval.
Now, how can the edges of the original graph G appear in the representation?
First, we show that for any DE V(G), the interval for u that intersects f(M(u))
cannot be used to represent more than one of the edges of G. First consider u # uo.
No interval for u can entirely contain the outside interval for any u B {u, uo},
because this would produce an intersection between f(u) and f(M(u)). Therefore,
any intersection between f(u) and the outside interval for f(u) must use the ‘free’
endpoint of that interval. Since G has no triangles, it follows that only one of the
edges incident to u in G can be represented using its outside interval, unless that
interval entirely contains an interval for uo. Suppose the outside interval for u
entirely contains an interval for uo. Then the interval for u. that intersects f(M(uo))
must intersect intervals for the two other neighbors of u. in G. Showing this is
impossible will complete all parts of the claim. Arguing as before, this interval for
u. must entirely contain the inside interval for some vertex WE V(G). But then the
outside interval for w must intersect outside intervals for two vertices of G other
than uo, which is impossible.
Now, consider the interval for ufu, inside f(z). We claim it can intersect at
most two other intervals for vertices of G. Suppose not. Since G has no triangles,
one of these three intervals must be interely contained inf(u); suppose it belongs
tof(w). Then the other interval for w must intersect f(M(w)), and it must be used
to represent two edges of G. By the fact proved in the preceding paragraph, this is
impossible.
We conclude that, for every u # uo, the inside interval intersects intervals for two
neighbors in G, and the outside interval intersects an interval for one neighbor in
G. This means that the inside intervals group into paths. The only place such a path
can end is at an interval for uo, and there are only two intervals for uo, so all the
inside intervals must belong to a single path with intervals for u. at the ends. The
intersection graph for these intervals is a Hamiltonian circuit in the original graph
G. At this point, having assumed a 2-representation for G’, we have forced a
Hamiltonian circuit in G. This completes the proof. cl
 Interval number is NP-compiere 303

4. Further NP-completeness results and additional problems

The result of Theorem 1 holds equally well for f > 2; 2-representations merely pro-
vide a convenient place to get started.

Theorem 2. For any integer t 12, t REP is NP-complete.

Proof. Given Theorem 1, we need only reduce (t - I)-REP to t REP. Let G be any
graph. We will construct a supergraph G’ of G such that G has a (t- l)-representa-
tion if and only if G’ has a f-representation. For each o E G, create three copies of
&>+I-t,t+1. Call them Hi(o), i = 1,2,3. Join o to one vertex of Hz(u). Add two
more edges, one joining Hz(u) to each of H,(o) and H,(u), using vertices from the
two different parts of Hz (see Fig. 5). If G has n vertices and m edges, then G’ has
n(1 + 3tZ + 6t) vertices and m + 3nt*(t + 2) edges, so this is a polynomial transfor-
mation.

H,(VI

Hz(v)

Fig. 5. The transformed graph G’.

Consider how a t-representation of G’ must be arranged. All the Hi are f-tight.

Hence in any f-representation of G’ the Hi(u) appear contiguously for a fixed
vertex O, using c intervals per vertex. The edges between the Hi(u) can only be
realized by overlapping the intervals at the ends of thef(Hi(o)). By Lemma 2, this
can be done. Having done this, the endpoints of f(H,(o)) are covered, and the
edge from u to H*(u) must be represented by inserting an interval for u into the
displayed portion of its neighbor in H,(u). This interval for u cannot intersect any
other intervals, because u and its neighbor in H,(u) have no mutual neighbors.
Creating this part of the representation for each vertex u E G represents all of the
new edges but none of the old edges and uses one interval for each u E G. None of
these intervals can intersect any other intervals in the representation. Therefore, G’
has a t-representation if and only if the subgraph G has a (t- 1)-representation. Cl

Another related result follows easily at this point. Recall that the depth of a repre-
sentation is the maximum number of intervals in it that cover a single point on the
line, and the depth-r interval number i,(G) is the smallest t such that the graph has
a depth-r f-representation.

Depth-r f-Representation (t,REP)

Instance: A graph G = (c*:E).
Question: Is i,(G) I t?
304 D. B. West, D.B. Shmoys

Theorem 3. For every t r2 and rz3, t,-REP is NP-complete.

Proof. Note that the 2-representation of G’ constructed in the proof of Theorem 1

has depth 3, so for any rz3 it is true that G has a Hamiltonian circuit if and only
if G’has a 2-representation of depth (at most) t-. Furthermore, note that in the proof
of Theorem 2 the additional gadgets have t-representations of depth 2, so if the
representation of G or G’ is of depth at least 2, the representation of the other has
the same depth. The theorem follows immediately from these two facts. 0

Several interesting special cases remain. What happens for depth 2? In particular,
any t-representation of a triangle-free graph must have depth 2. This leads us to ask
whether there is an efficient algorithm to determine the interval number of a
triangle-free graph. To reach NP-completeness we need not have anything larger
than triangles, because the graph generated in the main transformation has no K4.
Similarly, Scheinerman and West [ 1l] proved that the interval number of a planar
graph never exceeds 3. Given a planar representation of a graph, their proof gives
an efficient algorithm to find a 3_representation, but it does not determine whether
the interval number is 3 or smaller. Hence we ask, is there an efficient algorithm
to determine the interval number of a planar graph?

Acknowledgement

We wish to thank our good friend Pattie Suyemoto for suggesting this collabora-
tion.

References

[l] S. Benzer, On the topology of the genetic fine structure, Proc. Nat. Acad. Sci. 45 (1959) 1607-1620.
[2] KS. Booth and G.S. Leuker, Testing for the consecutive ones property, interval graphs, and graph
planarity testing using PQ-tree algorithms, J. Comput. System. Sci. 13 (1976) 335-379.
[3] P. Chambon, Split genes, Scientific American 244 (May 1981) 60-71.
[4] M.R. Garey and D.S. Johnson, Computers and Intractibility, A Guide to the Theory of NP-Com-
pleteness (Freeman, San Francisco, 1979).
[5] J.R. Griggs, Extremal values of the interval number of a graph, II, Discrete &lath. 28 (1979) 37-47.
[6] J.R. Griggs and D.B. West, Extremal values of the interval number of a graph, SIA,M J. Algebraic
Discrete Methods 1 (1980) l-7.
[7] L. Hopkins and W.T. Trotter Jr., A bound on the interval number of a complete mutipartite graph,
in: G. Chartrand. Y. Alavi, D.L. Goldsmith, L. Lesniak-Foster and D.R. Lick, eds., The Theory
and Applications of Graphs (Wiley, New York, 1981) 391-407.
[8] L. Hopkins, W.T. Trotter Jr. and D.B. West, The interval number of the complete multipartite
graph, Discrete Appl. Math. 8 (1984) 163-187.
191 L.T. Kou, Polynomial complete consecutive information retrieval problems, SIAM J. Comput. 6
(1977) 67-75.
 Interval number is NP-complete 305

[IO] L.T. KOU, L.J. Stockmeyer and C.K. Wong, Covering edges by cliques with regard to keyword con-
flicts and intersection graphs, Comm. ACM 21 (1978) 135-138.
[I 11 E.R. Scheinerman and D.B. West, The interval number of a planar graph: three intervals suffice,
J. Combin. Theory (B), to appear.
[12] W.T. Trotter Jr. and F. Harary, On double and multiple interval graphs, J. Graph Theory 3 (1979)
205-2 I 1.