COMBINATORICA 1 (1981) 13-24

MINIMAL DECOMPOSITIONS OF GRAPHS

INTO MUTUALLY ISOMORPHIC SUBGRAPHS
by
F. R . K. CHUNG
Bell Laboratories
Murray Hill, New Jersey 07974, USA
P . ERDŐS*
Mathematical Institute of the Hungarian
Academy of Sciences
Budapest, Hungary H-1053
and
R. L. GRAHAM
Bell Laboratories
Murray Hill, New Jersey 07974, USA
Received 14 May 1979

Suppose 9. = !{pGl , . . ., Gk} is a collection of graphs, all having n vertices and e edges. By a
U-decomposition of T„ we mean a set of partitions of the edge sets E(G,) of the G,, say E(G,)=
= E E,f , such that for each j, all the E,j , 15 is k, are isomorphic as graphs . Define the function
U(T„) to be the least possible value of r a U-decomposition of W„ can have . Finally, let Uk (n) denote
the largest possible value U(~V ) can assume where T ranges over all sets of k graphs having n vertices
and the same (unspecified) number of edges .
In an earlier paper, the authors showed that

2
U, (n) = n + o (n).
3

In this paper, the value of Uk(n) is investigated for k>2. It turns out rather unexpectedly that the
leading term of Uk (n) does not depend on k . In particular we show
3
Uk (n) =
4 n + o k (n), k ? 3.

1 . Introduction
Let T = (G,, G 2 , . . ., Gk ) be a collection of graphs,** all having the same
number of edges . By a U-decomposition of 9 we mean a set of partitions of the
edge sets E(G) of the G i , say E(G i) = Y E;j , such that for each j, all the Ejj ,1 s i s k,
J=i
are isomorphic as graphs . Under the above hypothesis, T always has such a decom-
position since we can always take all the E;j to be single edges . Define the function
U( ) to be the least possible value of r a U-decomposition of
T can have . Finally, let Uk (n) denote the largest possible value U(T) can assume

* This work was performed while the author was a consultant at Bell Laboratories .
** In general we follow the terminology of [1].
AMS subject classification (1980) : 05 C 35


14 F. R. K. CHUNG, P. ERDŐS, R . L . GRAI-TAM

where 9 ranges over all sets of k graphs each having n vertices and the same (un-
specified) number of edges.
In previous work [2], [3], the function U2 (n) was investigated rather thoroughly .
In particular, it was shown that
(1) U,(n) =3 n+o(n) .

An example of a pair of graphs (G,, G 2) achieving the bound in (1) is given by taking
G, to be a star S„ with n edges and G2 to be (3) K3 (see Fig. 1) .

C&

131 = Sn 02 =(3 )
K3

Fig . 1 . A pair of graphs T with U(I) -? n

3

We should point out here that strictly speaking, S„ has n + 1 vertices and further-
more, when n g 0 (mod 3), (1)
3 K3 is undefined. However, here as throughout the
entire paper, such statements are always to be taken with the understanding that
the graphs may have to be adjusted slightly by adding or deleting (asymptotically)
trivial subgraphs so as to make the stated assertion technically correct .
It was also shown in [2] that when T is restricted to bipartite graphs, the
corresponding function U$(n) satisfies

U2 (n) = 2 n+o(n)
with an extreme pair given by taking G, to be a star S„ (together with 2 isolated
vertices) and G2 to be 2 disjoint edges .
s

In this paper we study Uk (n) for k_-3 . Already for the case k=3 it is not
hard to find graphs G,, G2 , G3 on n vertices which require asymptotically more than
2 n subgraphs in any U-decomposition . For example, taking G1 = S„ , G2 = (3 ) K3


DECOMPOSITIONS OF GRAPHS INTO ISOMORPHIC SUBGRAPHS IS

and G, = (
n 2)
S, U Ky., then for T _ (G,, G2 , GO we have

U(g) = 4 n+o(n) .
In fact, as we will show, this is the worst possible behavior since

U, (n) = 3 n+o(n) .

What was completely unexpected is that it does not get any worse than this as k
increases . Indeed, the main result of the paper is that for any fixed kz3,

U, (n) = 4 n + o (n) .
Before proceeding to the proof of this, we remind the reader of the following
notation : S„ denotes the star on n edges, i.e ., the graph consisting of n vertices of
degree 1 and one vertex of degree n ; K„ denotes the complete graph on n vertices ;
nG denotes n disjoint copies of G ; G c H indicates that G is a (partial) subgraph
of H ; and finally V(G) and E(G) denote the vertex set and edge set, respectively,
of a graph G and v(G) and e(G) denote the corresponding cardinalities IV(G)l
and JE(G)J .

2. The main result

The bulk of the paper will be devoted to proving the following result .
Theorem. For any fixed k -- 3,
3
(3) Uk (n) = n + o (n).

The proof of (3) is somewhat complicated . An outline of the plan of attack

is as follows . We first choose an arbitrary fixed e ::-O . We assume we begin with
graphs (G,, . . ., Gk)=ig each having n vertices for a (sufficiently) large value of n,
and eo edges . We will then successively remove isomorphic subgraphs H from the
G;, thereby decreasing the number a of edges currently remaining in each of the
original graphs . Just what the subgraphs H= H(e) are which will be removed will
depend on the current value of e . There will be basically six distinct ranges for e,
which we show in Fig. 2 .
The STEP k notation indicates the process by which H(e) is chosen . Each of
the steps requires rather different arguments ; the preparation for these arguments
will now be made in a series of lemmas .
Let us denote by h(g) the maximum number of edges in any subgraph H
with HCG„ 1si-k .

16 F. R. K . CHUNG, P . ERDŐS, R . L. GRAHAM

Lemma 1 .
eok
(4) h (T) _' n ,k-1
/
2 JJ

Proof. Let A j denote the set of all 1-1 mappings of V(Gi) into V(Gj . For A jEA j ,
ej E E(GI), 1 s i s k, define
1 if A j maps ej onto el ,
Iz,, . . .' z k (el' ' ek) _
{ 0 otherwise,

STEP 1

STEP 2

STEP 3

1 STEP 4
C. n
STEP 5

STEP 6
1
Fn
STEP 7
0
Fig. 2. Ranges for e

where we say that A j maps e j onto e l if it maps the endpoints of e j onto those of
el . Then

S = Z Z Iz,, ...,zk(el, . .., ek) _ _Y (2(n-2)!)k-1 = eó(2(n-2)!)k-r .

el E E(GI) zs E 4$ e, E E(Gl)

ek E E(Gk) dk ek E E(Gk )
Zk E


DECOMPOSITIONS OF GRAPHS INTO ISOMORPHIC SUBGRAPHS 1,7

Since JAJ = n! for all i, then for some ~2 EA 2i . . ., ~k EA,,,

1
h,_ . ., zk (e1 , . . . , ek) S -
e,EE(Gl) IA21 JAkI
ek E
E(G,,)

eo ek
k-1 (2 (n - 2)!)' - ' k-1
(n ! )
(z)

The íI, now determine a subgraph H common to all of the G, which has .at least
eó
k-1 edges and the lemma is proved .
()
k
Suppose G is a graph with v(G)=n and e(G) =e=mn .
Lemma 2 .
(i) If n'13 s m then 3 V rn ) S,R c G;

(ü) If m < n'l3 then ( I M) S,„ G.

Proof. Let X= {x,, . . ., x,) 9 V(G) denote the set of centers of a maximum set of
disjoint S,„'s . Let Y consist of all vertices in V(G)-X which form the endpoints
of these S„,'s . Thus, IYJ=rm . Suppose deg (v) :(r+1)(m+1) for some vEY. Since
v is joined to at most r vertices of X and rm vertices of Y then v is connected to at
least m vertices in Z= V(G)-X- Y, say W= {wI , . . ., w,„) . We now remove x,,
center of the S,„ to which v belongs, and the m vertices in Y attached to x, . Also,
add v to X and W to Y. Thus, we still have X- {x,)U {v) as the centers of disjoint
5,,,'s, the endpoints of which are in Y.
We now keep repeating this process . Suppose at some stage the vertex x,E X
we remove also has deg (xi) ?(r + 1) (m + 1) . Thus, x, is connected to some vertex
uE Z - W since
JXJ-1+IYI+IWI : r-l+rm+m < deg(x) .

In this case we can add v to X, W to Y and also add u to Y (to give x, a complete
disjoint S,,, of its own), forming r+ 1 disjoint S.'s in G . However, this contradicts
the definition of r .
Thus, we must eventually reach a stage at which all vE Y have deg (v) <
(r+ 1) (m + 1). Since IZ j =n and S,„ Z then e(Z) { 2 mn . The number of edges
incident to some point in Y is at most

JYJ •m aY deg (v) s rm(r+l)(rp+1) .

Finally, the number of edges incident to some point in X is at most rn .

2*
18 F. R. K. CHUNG, P, ERD6S, R . L. GRAHAM

Therefore
e = mn s rm(r+l)(m+1)+rn+e(Z)

3
rm(r+1)(m+1)+ mn+rn .
This implies
3
(5) mn s r(r+1)m(m+l)+rn .

Case (i) . ni ts s m . Suppose r 3 m . Then

V
r(r+l)m(m+l)+rn

812
191
.m(m+1)+ 3 . m 1/2
m

sn(
0
3 m+3 m) 3 mn
for n sufficiently large which contradicts (5) .
Case (ü) . mLnl t3. Suppose r<3 m . Then

r(r+l)m(m+l)+rn

191
s ms (m+1)+3 mn

=m ( 1 9 n+3 n)~3mn

for n sufficiently large which again contradicts (5) . This proves the lemma.
Lemma 3 . If e(G)y2dt then either
Sd E- G or tSl c G.
Proof. For t=1 the assertion is clear. Suppose for some t> 1 that S d G . Choose
an edge yE E(G) . Remove y and all incident edges from G, forming G' . Since deg (v) s
d-1 for all vE V(G) then
e(G') z 2dt-2d+1 > 2d (t-1) .
By induction, (t-1) Sl c G' . Thus, since y is disjoint from this (t-1) S, then
tSI 9 G .

Lemma 4. If e(G) z 2 +3dt then either

Sd 9 G or tS2 9 G.


DECOMPOSITIONS OF GRAPHS INTO ISOMORPHIC SUBGRAPHS 19

Proof. For t=1 the assertion clearly holds . Suppose for some t71 that Sd G .
Choose an S2 9 G. Remove it and all incident edges, forming G' . Since deg (v) S
d-1 for all vE V(G) then
e(G~ 2 +3d(t-1) .
Thus, by induction (t-1) S 2 E- G'. Since the S 2 originally removed from G is dis-
joint from this (t-1) S2 then tS2 9 G and the lemma is proved .
We are now ready for the proof of the Theorem. What we will do is to describe
and analyze each step in the decomposition process as the current number of edges
e passes through the previously indicated ranges . In particular, at any time esen
we immediately go to STEP 7, which is simply the removal of subgraphs consisting
of a single edge.
STEP 1 : n2-11k < e s ( 2) . In this step, we repeatedly apply Lemma 1, removing a
ek
common subgraph having at least _ 1 edges . Thus, if ei denotes the number
n
( 2)
of edges remaining in each graph after i repetitions have been performed then
e;
(6) e,+, 5 e, - n k-1

(2)
e
Let a t =e, Then ao = and
(n ) (n )
2 2
a,+1 s ai - ai =f(ai)•
Thus,
f '(x) = 1- kxk-1
1
1 k-1
and so, f(x) achieves a maximum at xo =( k and f(x) is monotone increasing
for Osx<xo .
Suppose
1
k-1
ai - xo for some i z 1.
5 (-Li)
1i
Then
1 I k 1 1
) (-I )k-1
f (00 f
~( 1 k-1) = 1 )k -, ( 1 ) k_1 ( 1 ) k-1 ( 1 i) ( i+ 1
Therefore
1
1 k-1
a,+1 Sf(ai) 5
(i+l)


20 F. R. K. CHUNG, P. ERDŐS, R . L. GRAHAM

Also, we have
1
1 lk-1
a, f (ao) f (xo) _ (k) s 1
so that by induction,
1
l k-1
ai ~~_ for all i,
( 1)
.

( n)
2
ei
(7)- ` 1
ik-1

Choosing io =n1 - 'I', we see that

( l n l k )~k-1)-1 n
e t, s
2n )///j (
1 t

Thus, at most n" -11k

subgraphs are removed during STEP 1.
STEP 2 : n4/s
-esn /
In this step, we repeatedly apply Case (i) of Lemma 2 .
2-1 k

Abusing notation slightly, let eo denote the number edges each graph has at the
beginning of this step . In general, if ei denotes the number of edges remaining after i
applications of the lemma, then
(8) ei+1 ei- 3
In
since the number of edges in 3 S. is essentially 3 nm = 3 ~. Let ej =9ei .
V
Equation (8) then becomes
(9) et+1 et - eí = g(e ;) .
Note that g(x)=x-C is a parabola (at a 45° tilt) which is monotone increasing
for xz1/4 . Suppose for some tLi/2>-0 that
z.
ei s (t -
2)
Then
s g (( t- 1)2l i i2 1 ( i+1
e'i+i s e')
- 9(e') ( t- 2 i
- ( t- 2 = (t - t- 2
2 J2 J J2- 4 J2
Since e,,*sn 2-1 /k by hypothesis then taking t=n1-1/2` we have by induction

e, ,~ ei < ( n 1-1/2k - i
la f.
2

We apply this process only as long as e i >- n'/$ so that at most 2(n1-1/2`-n2/$) sub-
graphs are removed in this step.


DECOMPOSITIONS OF GRAPHS INTO ISOMORPHIC SUBGRAPHS 21

a
STEP 3 : Cen-<e n413 for a large constant C depending on s. In this step, we repeat-
t
edly apply Case (ü) of Lemma 2 . Again, let e denote the number of edges remaining
in each graph after Lemma 2 (ü) has been applied i times . Then

1 er 2
(1 0) ej+i s ej- 3 (n)

By letting fl j =e j/3n2 the inequality becomes

fli+i 5 Nj - Ni •
By performing an analysis parallel to that used for the a j in STEP 1 (with k=2),
we deduce
3n2
C for i--l .
i
Actually, we could take. advantage of the fact that & n413 and strengthen (11) but
it would have no effect on the final estimates . Hence, to reach esC n requires the a
removal of at most C subgraphs .
a

a
STEP 4 : es C n . The first part of this step consists in successively removing (log n) S I
from all the graphs as long as possible. The second part consists of successively
removing S,ogn from all the graphs as long as possible . If . after this process stops,
the number a of remaining edges is less than en then we go directly to STEP 7 .
Thus, we may assume that a>-sn . Let us denote by Hi , . . ., H those graphs having j
no (log n) SI as a subgraph and by H + I j , . . ., k
H those subgraphs having no S n log
as a subgraph . By Lemma 3, if (log n) SI qj H then S an 9 H i Slog n 9 H, i, i.e.,
2logn
(with a similar argument applying if S,ogn gj Hj). Since e >-&n then we must have
1 ~j<k . This completes STEP 4 .

Note that the number of subgraphs removed in this step is at most og n .

t a

STEP 5 : 2 (l +E)-<e a
C n. Let d denote the largest degree of any vertex in any

Hi, 1 . i s k . Since e > 2 (1+8) then by Lemma 3, (log n) Sl i

H implies S n C
41ogn
Hl, i.e.,
n
(12) d
4 log n
Define
j
X, _ {vE V (H ) : d - deg (v) -s 1),

Yj = maximum set of disjoint S2 's in Hj .

By definition

e z 2 JXJ (d -1) for all i,



22 F . R. K . CHUNG, P. ERDŐS, R. L. GRAHAM

IXi J S C' log n .

Also, for j + 1 s i s k, since S,ag„ qj Hi then by Lemma 4,
e n
S2 C
6 log n ) Hi '
i .e.,
E n
(13) 1Y,1 J+1 s i k.
6 log n'
Define x* = max JX1 I . Thus
is~sl
x* s C' log n .
For some io sj, IXi o 1=x * . Therefore
.
(14) e = e(H, o) z (d-1)x*- C' log n
2

Now, define Z i _ {vE V(Hi) : deg (v) z F} for 1 s i j. Suppose JZ; I x* -1 . Con-
sider the graph Hi induced by V(Hi ) -Zj . Note that~'
(log n) Sl $ H;
and
e(Hi) e(H,) - IZJA
.logn)-(x
z (d-1)x* -(C * -1)d

z d -x* - C" log2 n

n- Cs loge n .
4 log n
Thus, by Lemma 3, for

(15) m -C8 loge n)/2 log n
(41og n

we have S,„ 9 H•' . However, the expression in (15) exceeds C for large n which means
that Sy C H, . This contradicts the definition of Z i . Hence, we may assume

(16) Iz,I X ..
s
Finally, for 1 h5j, we define Xl to be Xi U Z{ where Zi g Z; is disjoint from X;
and so that
WIN = IX,I + Iz", I = x*

(this is always possible by (15)) .

It is now easy to see that we can remove x*S2 from each H; so that d is
decreased by 2 . This can be done by choosing each x ;EXi as a center for an S2 for
15 i sj (since deg (xi) y . and x* s C' log n , then this is always possible) . For
j + 1 s i s k, (13) guarantees that x* S 2 E- Hi .


DECOMPOSMONS OF GRAPHS INTO ISOMORPHIC SUBGRAPHS 23

STEP 5 consists in successively removing x* S2 's (of course, each time the
value of x* may change) until eL (1+ E) . Each time a subgraph is removed, the
2
maximum degree d decreases by 2.

STEP 6 : En se7fi
2 (1+ E) . The plan in this step is similar to that of the previous
step . In this case we will remove each time the subgraph x* S, so that d always
decreases by 1 . To see that this is possible, define X i as in the previous step, i .e.,
Xi = (vE V(Hi) : d-deg (v) 1 }. For j+ 1 s i k, define Yt to be a maximum set of
disjoint S,'s in Ht . As before, it follows that
x* = max JXiJ s Ct log n
lsisj
and
I 8 n
Yi 2 log n
Defining Z i as in the preceding step and extending X i .to Xi with (Xt =x*, 1 I

it is not hard to see that x*S, can be removed from each Hi so that d decreases by 1 .
STEP 6 consists in successively removing x*S, in this way until esőn .

STEP 7 : e---&n. This final step consists in successively removing S l , the subgraph
consisting of a single edge . Of course, in this step at most an subgraphs are removed .
We are now ready to count the number N of subgraphs into which each of
the original graphs has been partitioned . Let ai denote the number of subgraphs
removed during STEP i . Then

N 2 ai
i=1

n1-11k+2n1-1/2k+ 3n + C.n
+a5+66+En s
Ce log n
a5+a 6 +2En

3
for C,y and n sufficiently large. However, because of the guaranteed reduction
E
in d (which at the beginning of STEP 5 is certainly less than n), we have
(18) 2v 5 +vs < n.

6 each subgraph has at least one edge and (e :

Also, since in STEP
2 (1 +8) during
this process),
(19) a6 =2 (1+E) .

Adding (18) and (19) we obtain

(20) 2(a, +a,) 2 (3+s).



24 F . R. K . CHUNG, P . ERDŐS, R. L. GRAHAM

Substituting in (17), we have for =(G,, . . ., Gk ),

U(f) s N { ( 3 +3e) n .
Since both a and T were arbitrary then we conclude

U(n) s 3 n +o (n) .

Since we have already given an example of three graphs

( n- Cn
( 3 ) K3 , ) S, U Ky = 9 with U(~ ) =
2 4 n + o (n)
then the Theorem follows .
It would be interesting to know if the o (n) term could be strengthened, say,
to O (1) .

3. Concluding remarks
If we restrict all the G, to be bipartite then it turns out .that the bound on the
corresponding function H,*(n) is the same as that for U, t (n) when k= 3, in contrast
to the bound we mentioned previously :

U2 (n) = 2 + o (n) .
In other words,

U,*(n) = 4 n + o (n)
for all ~ 3 . An example of three bipartite graphs which achieve this bound is given
k
(n -2 2n
by taking G, = S„ , G2 and G3 = ) Si U KyR12, .12
(4) K2,2
In another direction, one can ask the same questions for r-uniform hyper-
graphs . Here, the answers required are harder to obtain and are known with less
precision. For example, in the case of two r-uniform hypergraphs on n vertices, say
.W=(Hi , H2), it can be shown that r even,
cl nri2 < U(del) < c2 n r t2
for suitable positive constants c, . This topic will be treated more fully in a later paper-
Finally, it is natural to ask how close Lemma 1 is to the "truth", i .e ., is this
essentially the right order for h(G,, . . ., Gk)? This too we leave for later .

References
[1] J. A. BoNDY and U . S. R. MURTY, Graph Theory with Applications, American Elsevier, New
York (1976) .
[2] F. R . K . CHUNG, P. ERDős, R. L . GRAHAM, S . M. ULAM and F. F. YAo, Minimal decomposi-
tions of two graphs into pairwise isomorphic subgraphs, Proc. 10" Southeastern Conf. on
Comb., Graph Th . and Comp ., (1979) 3-18 .
[3] F. FRANCES YAo, Graph 2-isomorphism is NP-complete, Inf. Proc. Letters 9 (1979) 68-72.