3

On the maximal number of disjoint circuits of a graph

By P. ERDŐS and L . PÓSA (Budapest)

Throughout this paper Gg " will denote a graph with n vertices and k edges
where circuits consisting of two edges and loops (i . e . circuits of one edge) are not
permitted and G'" will denote a graph of n vertices and k edges where loops and
circuits with two edges are permitted . v(G) (respectively v(G)) will denote the num-
ber of edges of G (respectively G) .
If x,, x" . . ., x,, are some of the vertices of G, then (G-x, - . . . -x k) will
denote the graph which we obtain from G by omitting the vertices x,, . . ., x k and
all the edges incident to them . By G(x,, . . ., x k ) we denote the subgraph of G spanned
by the vertices x,, . . ., xk . The valency of a vertex x - v (x) - will denote the number
of edges incident to it . (A loop is counted doubly .) The edge connecting x, and x,
will be denoted by [x,, x,], edges will sometimes be denoted by e,, ez , . . . . (x,, x,, . . .xk )
will denote the circuit having the edges [x,, x,], . . ., [xk _,, .vk], [x k x,] .
A set of edges is called independent if no two of them have a common vertex .
A set of circuits will be called independent if no two of them have a common ver-
tex . They will be called weakly independent if no two of them have a common
edge .
In a previous paper ERDŐS and GALLAI [l] proved that every

2k ' 1 k 2 11]
(l) G,+i where 1=max (k-1)n-(k-1) +

contains k independent edges .

In the present paper we shall investigate the following question of Turanian
type (see I) : how many edges are needed that a graph should have to contain k
independent or weakly independent circuits? Put
f(n, k) _ (2k -l) n - 2k 2 + k .
Our principal result will be that for n --- n o (k), k--l every Gr( „_k) contains k inde-
pendent circuits except if it contains 2k - 1 vertices of valency n -1 (its structure
is then uniquely determined) . If k=l trivially every vn"' contains a circuit, but
there are of course graphs G ; ;"_', where no vertex has valency n - I and the graph
nevertheless contains no circuit . Thus the restriction k > I is necessary.
Clearly n o (k)-- 3k (since a circuit contains at least three vertices) . For k=2
and k=3 n o (k)=3k, but in general n o (k) > 3k, but we will prove n o (k) ~ 24k.
 4 P . Erdős and L . Pósa

Perhaps the following result analogous to (l) holds : Every

3k l
(2) GM with l=max [( )+n 3k+2 (2k 1) n 2kz+k
2 J
contains k independent circuits .
Denote by g(k) the smallest integer so that every Gr+g(k) contains k weakly
independent circuits . We will show that g(2)=4 and that for every k

c,k log k--g(k) < c z k log k

where c l and c 2 are suitable absolute constants (the c's throughout this paper will
denote suitable absolute constants) . The exact determination of g(k) seems to be
a very difficult problem and we cannot even show that g(k)llog k tends to a cons-
tant. Further we do not know the value of g(3) .
It is easy to see that g(2)--4 i . e . we will show that for every n 6 there exists
a G„+3 which does not contain two weakly independent circuits . To see this let
the vertices of our graph be x,, . . ., x" and its edges

[x i ,xf ], I_-i_-3-<j-6 and [xk , x k+ ,], 6_-k :n-1 .

A simple argument shows that this G ;,+3 does not contain two weakly independent
circuits.
After completing our paper we found out that some of our results were known
to G . DIRAC but he published nothing on this subject . In particular he proved that
for n~6 every G3„_ ; contains two independent circuits and that every G„'+ 4 con-
tains two weakly independent circuits . He also proved that for n--6 every G ( " )
where the valency of every vertex is 3 and the valency of every vertex with at
most one exception is ----4 contains two independent circuits and conjectured that
for n 3k every G ( " ) which is 2k-fold connected (i . e . which remains connected
after the omission of any 2k -1 of its vertices) contains k independent circuits .
Theorem 1 . Let k > 1, n 24k then every Gk-) either contains k independent
circuits or 2k - 1 vertices of valency n -l .
Our Theorem clearly implies that for n-24k every G"') k +i contains k inde-
pendent circuits (since a simple computation shows that a G f(;, k - which has 2k -l
vertices of valency n - I has all its other vertices of valency 2k-1 and its structure
is thus uniquely determined) . n -- 24k could easily be improved a great deal, but
our method does not give any hope of best possible estimates .
Theorem 1 . will be proved by a fairly complicated induction process and to
make this as painless as possible we will restate Theorem 1 . in a very much more
complicated form but which will be more suitable for our induction process .
Theorem 1' . Every Gt,"~ contains a circuit (k =l) . For k > 1 put

f(n, k) + (24k-n) (k-1) for n : 24k

g (n, k) =
f(n, k) for n = 24k .
 On circuits of a graph 5

Then if 3k n 24k - 1 every G ( n ) , k) contains k independent circuits, and if n ~ 24k

and la -g(n, k) then every Gio) I) contains k independent circuits except only if
la =g(n, k) and Gi,' contains 2k - 1 -vertices of valency n -1 and n - (2k -1) vertices
of valency 2k - 1 .
Since g(n, k)=f(n, k) for n ~24k Theorem F implies Theorem 1 .
Theorem V is trivial for k =l . It is also trivial for k > 1 if 3k n 6k since
by a simple calculation
32 1
g (n, k) j 22
i
and for n - 3k the complete graph contains k independent circuits .
First we prove two Lemmas .
Lemma 1 . Let n 6k and assume that Gt"t contains 2k vertices x,, x 2 , . . ., x, k
of valency v(x i)--n-k (l-_i--2k) . Then G"' contains k independent quadrilaterals .
Denote by y,, . . ., Y„-,,, the other n-2k vertices of G"' ) . Consider a maxim a
system of independent quadrilaterals of the form

(X-1i-I ^ Y2i-11 x-i1 Y2i), I-k .

We shall show l = k . Assume l < k . Each of the vertices x , , +, and x,,, 2 are connec-
ted with at least n - k vertices . Thus every vertex except possibly 2k vertices are
connected with both x 2 ,-, and x,t+2, i . e . these are at least n-2k~4k of them
which are connected with both and x, t +, . Since 2k+2l--4k-2 there are
two further vertices and y,1+2 which are connected with both X21 +, and
. Thus the quadrilateral (x2a+I~ Y21-I-, x21 +2,Yet+2) is independent of the
x,,+2
others, which contradicts our maximality assumption, which proves Lemma l .
Lemma 2 . Let n-2k and assume that every vertex of G" ) has valency --2k,
then G ( " ) contains k independent edges .
Lemma 2 can be proved from first principles in a few lines as follows :) Let
e ;,-[x2i_, .x2i], 1t be a maximal set of independent edges . Assume t<k
(otherwise there is nothing to prove) . But then since n ~:- 2k there are two vertices
of G t "I y, and Y2 distinct from the x ;, 1--i-_2í, y, and Y2 can be joined only to
the a- i 1--i-2t (by our maximality assumption), and by the same assumption
if y, is connected to an endpoint of e t l--i--t then y 2 can not be connected to
the other endpoint . Thus v (y,) v (y,) -_ 2t < 2k which contradicts v (y i ) --k . This
contradiction proved the Lemma .
Now we prove Theorem V by induction . Let k > 1 and assume that Theorem Y
holds for k - 1 and assume that it holds for every 6k < m < n . (We already remarked
that it trivially holds for 3k-_m-_6k) . Then we shall prove it for n, and if we have
succeeded in this the proof of Theorem F and therefore Theorem 1 will be complete .

1) If l0>~2) Gi n ' will denote the complete graph of n vertices .

2) This proof is due to G . DIRAC (written communication) .
6 P . Erdős and L . Pósa

Assume first that our G9(,),, k> contains 2k vertices of valency --n-k.
Then by Lemma 1 . our graph contains k independent quadrilaterals and thus
Theorem 1' is proved in this case .
Henceforth we can assume that G contains at most 2k -1 vertices of valency
-n-k . If all the other vertices have valency <2k then the number of edges of
G is at most
(2k-1)(n-l)+(2k-1)(n-2k+1)_f(n
k)
2
and equality can occur only if G contains 2k -1 vertices of valency n -1 (i . e. these
vertices are connected with all the vertices of the graph), and G contains no other
edge, otherwise it would contain another vertex of valency ~--- 2k) . Thus the structure
of our G is uniquely determined (G can have this structure only if n-24k) and Theo-
rem I' is proved in this case too .
Therefore we can now assume that G has a vertex - say x o - of valency l'
satisfying
2k :1'<n-k.

Let x l , . . ., x,, , 2k -- Z' < n - k be the vertices of G connected with x o by an

edge . Assume first that in the graph G(x l , . . ., x, , ) there is a vertex - say x l -
of valency -<k . It may be assumed that x l is not connected by an edge to any of
the vertices x,+1, . . ., x,•, where r--k. Define the graph G, with n-1 vertices as
follows : Omit the vertex x o and all the 1' edges incident to it, and add the edges
[x l , x,+,] 1-_ i l' - r (i . e . x l is connected in G, to all the vertices to which xo
is connected in G [except of course xj) . Clearly
v (G,) ~ v (G) - k
or
(3) v (G,) ~-- g (n, k) - k ~--- g (n -1, k) .
Thus by our induction hypothesis G, contains k independent circuits (from the
first inequality of (3) it follows that v (G 1 ) >f(n -l, k) thus G l cannot have 2k -1
vertices of valency n -1 and n - 2k vertices of valency 2k - 1, i . e . the second alter-
native of Theorem 1' . is excluded) . But then G must also contain k independent
circuits . To see this let Cl , . . ., Ck be the k independent circuits of G 1 , at most one
of these circuits - say C l - contains one or two of the new edges [x l , x, + ,],
(if none of these circuits contains any of these edges, then C l , . . ., Ck
are k independent circuits of G) . If C l contains only one of the new edges - say
[x l , x, ,+1] - then we obtain Cl by omitting [x j , x,+1] from C l and replacing it
by [x o , x l ] and [x o x, ± 1] . If C l contains [x l , x,] and [xl , x,+1] then in Ci these
are replaced by [xo , x,] and [x o , .xr+1 ] . In any case Ci, . . ., Ck are k independent
circuits of G . Thus Theorem l' is proved in this case too .
Assume next that all the vertices of G(x l , . . ., v, , ) have valency --k . Then
(since l'--2k) by Lemma 2 . G(x l , . . ., xr) contains k independent edges e,_
_ [x 2 ,- 1, x2,], 1 i ~<k . Assume first that each of the e, are contained in at least
k-l triangles (x_zi _1,x2E,yri% 1--t--k-1, 1--i-k where the y,r> are all dif-
ferent from xo> X1, xzk-
 On circuits of a graph 7

In this case Theorem 1' easily follows since G contains k independent triangles .
To see this observe that it immediately follows from our assumptions that there
are k - I independent triangles (x2 i _ 1 , x2 b y ,), 1 : i t k - l . (xo, x2x-1, x20 is the
k-th independent triangle .
Henceforth we can thus assume that - say el = [x l , x2] - is contained in
at most k - 2 triangles in the graph (G - x o - x 3 - . . . - x2 j . Put G 2 = (G - x o -x l - x 2 ) .
Now we estimate v (G2) from below, by estimating from above the number of edges
of G incident to x o , x l and x 2 . v(x o)<n-k by our assumption . In G(x l , . . ., x 2 0
the vertices x l and x2 are incident to at most 4k - 3 edges . Finally every vertex
of (G-xo-xl- . . . -x 20 is connected with at most one of the vertices x1 and x 2i
except possibly k - 2 vertices which might be connected with both . Thus we obtain
at most n - k - 3 further edges . Thus the total number of edges incident to x o , x 1
and x 2 is at most
n-k-1+4k-3+n-k-3=2n+2k-7 .
Hence for k>2
v(G Z ) - v(G) -2n-2k+7= (n-3) (2k- 3) -2(k- 1) -' +k+
(4)
+g(n, k)-f(n, k)>g(n-3, k-1)
since clearly
g (n, k) - f(n, k) - g (n - 3, k - 1) -f(n - 3, k -1) .
For k=2, we obtain

(4') v(G2) -- v(G) - 2n + 3 -- n - 3 .

Thus (4) and (4') imply that by our induction hypothesis G, contains k- 1
independent circuits . These and (xo , x l , x 2 ) are together k independent circuits
contained in G . Theorem l' is now proved .
If k = 2 then the assertion of Theorem 1 holds for all n -- 6 . The reader can
verify it for n = 6 and then adapt our induction process to prove it for n 6 .
Perhaps the following result is of some interest .
Theorem 2 . Let n~4k, then every G~2k-1>n-t2x-n'-+1 which contains no tri-
angle contains k independent circuits .
For k = 1 the Theorem is trivial . We use induction and assume that it holds
for k - l . Let (x l , . . ., xi ) = C l be the shortest circuit of G and denote by x1,+1, .. . . Xn
the other vertices of G . No two non-neighbouring vertices of C l can be connected
by an edge (for otherwise C l would not be the shortest circuit of G) . Assume first
that 1 1 >4 . Then every xr , 1, <r--n can be connected to at most one vertex
of Cl (for otherwise C l would not be the shortest circuit) . Thus the vertices
x l , . . ., x,, are incident to at most n edges . Let C2=(xt,+1, . . ., xr,+a,) be the
shortest circuit of (G-xl- . . .-xi,) and (xa,+i 2 +1, • • • ,
xi1 +z_,+0 the shortest
circuit of (G-xl- . . .-xi,+1,) etc . Thus weC3obtain = the circuits C 1i . . ., Cr of length
4<1 1 . . . Ir and we assume that the graph contains
no circuit . If r k our Theorem is proved . Assume r < k . By our previous
argument we obtain that in the vertices of C ; +1 are
8 P. Erdős and L. Pósa

incident to at most n - 1, edges . Since (G - x1 x,, + . . .+ cr) has no circuit

t= 1
r
it has fewer than n - S 1, edges . Thus the total number of edges of G is less than
r-~
r r
n+ (n-1,(r-~-1)n- 1, :5 kn-5<(2k-1)n-(2k-1)z

for n--4k, an evident contradiction .

Assume next that 1, =4 . Then every vertex v,, 4 r n is connected with at
most two of the vertices x,, x,, x 3 , x 4 . Thus the number of edges incident to x,,
x,, x 3i x4 is at most
4+2(n-4)=271 -4 .
Hence
v(G-x1 x,-x3-x 4 )-n (G)-2n+4= (2k - 1) n - (2k - 1)2 - 2n + 5 =
=(2k-3) (n-4)-(2k-3)2 + l .
By our induction hypothesis (G-x1 -x, -x3 - .a-4 ) contains k-1 indepen-
dent circuits which together with (x,, x,, x 3 , x 4) gives k independent circuits of G .
Thus the proof of Theorem 2 . is complete .
Now we show that Theorem 2 . is best possible . Let G be a graph whose ver-
tices are x,, . . ., x,,, and whose edges are [x ;, xJ ] where 1 -_ i n 2k - 1 -:::: .j --n . Clearly
G has (2k-1)n-(2k-1) 2 edges and does not contain k independent circuits .
If k > 1 then this graph is the only G ("2~•~ ._ i)n-(zip-1)'- which contains no triangles
and does not contain k independent circuits, we leave the proof to the reader .
G . DIRAC [2] proved that for n--4 every- G'2','_2 contains a topological complete
quadrilateral (i . e . it contains four vertices x,, x 2 , x 3 , _Y. any two of which are con-
nected by pairwise disjoint paths) . We shall give a simple proof of this theorem
by our method . For n=4 the theorem clearly holds . We will assume that it holds
for n- 1 and prove it for n . Our G`_';,'_2 clearly contains a vertex .x o of valency not
exceeding 3 . If v(t - o )<3 then i , (G-x,) --2n-4 and thus by our induction
hypothesis (G-x,) and therefore G contains a topological complete quadrilateral .
Assume that v(x,)=3 and let x,, x,, x 3 be the vertices connected with x o by an
edge . If [x,, .a- 2 ], [x,, x 3 ], [x. 2 , x3] are all edges of G then G contains the complete
quadrilateral {x,, x 2 , x 3 , x 4 } . Thus we can assume that one of these edges - say
[x,, x,] - does not occur in G . Add the edge [x,, x .] to (G-x o ), thus we obtain
a graph G' having n - 1 vertices and 2n - 4 edges . By our induction hypothesis
G, contains a topological complete quadrilateral ((Y,, Y2, Y3, Y41 . But it is immediate
that {Y,, y 2 , y3, Y4 } is a topological complete quadrilateral of G . To see this observe
that the new edge [x,, x,] can occur in at most one of the connecting paths and
there it can be replaced by [x,, .Y,] and [xo , x2 ] . G . DIRAC showed by simple exam-
ples that not every G2,'„!- 3 contains a complete topological quadrilateral, e . g . the
vertices are x,, a•„ the edges [x„ xJ], 2 _ j - n, [x 2 , r 1 ], 3 j n .
One could perhaps conjecture that for n ~5 every Gin' 5 contains a complete
topological pentagon, but the above proof breaks down and we can not even show
that there exists an absolute constant C so that for n--5 every G,°„ contains com-
plete topological pentagon .
 On circuits of a graph 9-

Theorem 3 . Every G ;,"+ 4 contains two weakly independent circuits.

In other words g(2)=4 (see the introduction). We use induction on n . Our
Theorem clearly holds for n =l . We will assume that it holds for n -1 and prove
it for n . If our graph contains a circuit of four or fewer edges, then our Theorem
is immediate, since by omitting the edges of this circuit a G") 4_ ; remains with
n+4-i--n, thus it contains another circuit thus giving our two weakly indepen-
dent circuits . Thus we can suppose that our graph contains no circuit with fewer
than five edges . If our graph contains a vertex of valency one we omit this vertex
and obtain a " which by our induction hypothesis contains two weakly inde-
pendent circuits . If x o is a vertex of valency two and x i , x, are the vertices connected
to x o by an edge then we define G i as the graph which we obtain from (G-x o)
by adding the edge [x,, x,] . Clearly G, has n -1 vertices and n +3 edges and thus
our Theorem again follows . If all vertices of G have valency _-3 then it has at least
3 n edges, or n=n+4, which implies n--8 . But it is well known and easy to show
that every graph with fewer than 10 vertices every vertex of which has valency --3
contains a circuit of at most 4 edges (for 10 vertices this is false as is shown by the
well known Petersen graph) . This completes the proof of Theorem 3 .

Theorem 4 . For every k > 1

(5) c,k log k -g(k) < c,k log k
tit-here c i and c, are suitable absolute constants .

First we prove the upper bound in (5), (no attempt will be made to get a good.
estimation for c,) . We shall use induction with respect to k . For k=2 the inequa-
lity follows from Theorem 3 . Assume that it holds for k-l, we shall prove it
for k. As in the proof of Theorem 3 . we can assume that every vertex of our graph
u> >
G« +[,,k k] has valency -3 . But then
r(G) 3 T1
or
(6) n 2c,k log k .
First we prove
-.(")e23L_,vEtmrnayxgofphwGicsvaleny
log 2l
contains a circuit of at most n I edges .
Clog ' J
If our graph contains a loop or a circuit of two edges our Lemma is trivial.
Thus assume that such circuits do not occur in our graph . Let a- i be any vertex
of G ( " ) . If G'") contains no circuit of ~2t edges, then all the vertices which can
be reached from x, in t or fewer edges are all distinct . Since every vertex of G'" 1
has valency _-3 a simple argument shows that in t steps we can reach at least
1+3 + . . .+32'-i=2'-'>n
Clog log n
vertices if t = 1 . Thus G ( "' contains a circuit of length not exceeding 2
g 2] [log 2]
as stated .
10 P . Erdős and L. Pósa

From (6) and Lemma 3 . our graph contains a circuit of length not greater
than
log n (2e,k log k) < ác
[log 2 log k
2 [log 2 - 2 log 2
for sufficiently large c2 . If we omit the edges of this circuit we obtain a graph of
at most n vertices and more than

n + [c 2 k log k] - log k > n + c 2 (k -1) log k

Z

edges . By our induction hypothesis our new graph contains k - 1 weakly inde-
pendent circuits, thus together with our first circuit we have our required k weakly
independent circuits, which completes the proof of the right side of (5) .
To prove the lower bound in (5) we need
Lemma 4 . There exists a constant c 3 > 0 so that for every m there exists a G(2'm
which contains no circuit of length less than C3 log 111,
The proof of the Lemma is implicitely contained in a paper by ERDŐS [3], but
for the sake of completeness we give it here in full detail .
Consider all graphs of in labelled vertices having 2rn edges . The number I of
M
((2) '
these graphs clearly equals Denote these graphs by G 1 , . . ., G r and denote
3m
by f(G) the number of distinct circuits of length not exceeding [c 3 log m] contained
in G r . We are going to estimate
r
M= I f(G)
:,Z
=1

from above . A simple combinatorial argument show's that the number of graphs G t
which contain a given circuit of k edges equals
rrz k
(7) 2
,3m-k,
The number of circuits of length k is clearly less than

(8) k! ( k) n1k .

Thus from (7) and (8) we obtain by a simple argument

(rnl
- k
1 [C3log 1,11 (2) log m 3m (3m - l) . . . (3n-a - k + 1)
(9) M< I S In [C3
z mk l rr
k=3 _ 3rn-kk k-a
12~ -k+11
12 1C ,2
1C3 log "'1
1W <m ; ,
k-3
 On circuits of a graph 11

if c 3 is a sufficiently small absolute constant (again no attempt is made to . get a

good estimate for c 3 since as in the previous cases there seems no hope at present
to obtain the best possible value for c 3 ) . From (9) we obtain that at least one of
our graphs - say G, - contains fewer than m'/= circuits of length < [e, log in] .
Omitting one edge from each of these circuits we obtain a graph of m vertices and
more than 3m -m' _ -- 2m edges which contains no circuit of length less than c 3 log m,
which proves Lemma 4 .
Assume first k >k o and let c l >0 be a sufficiently small absolute constant
and put m =[c,k log k] . Then by Lemma 4. there is a G(2',.) which contains no circuit
of length <c3 log in . Therefore our G2 "m contains at most

2m
k
c 3 log m

weakly independent circuits, if c l is sufficiently small . On the other hand our graph
has 2m = m + [c, k log k] edges, which completes the proof of the left side of (5)
for k > k o . But clearly for 2---5k--k, g(k) --g (2) = 4, thus if c l is sufficiently
small (5) holds for all k--2 and thus Theorem 4 . is proved .
If we have already constructed a G;n+t,,kiogkj which does not contain k we-
akly independent circuits, we can construct such a Gnu-r[c,klogkj for every n-m
by adding a path of n - i i new vertices and edges to our G,,,+[c,k lag k] .
Finally we consider the following question : Let m--n and consider a graph
G,,,'n,' . Define h(G,„") as the length of the shortest circuit Of Our G'." ) . Put

f(n, m) = max h(G ;,;')

where the maximum is taken over all graphs G,, °) .

Trivially f(n, n) = n and it is not difficult to show that

2n+2
` 3 .
]
The determination, or even the estimation, off(n, m for general n and m seems
a difficult problem .
Theorem 5 . Plat in =n+d,d,l . Then ,tie have

(n +d) log d
(10) f(n, n I d) < c 4
d

and to every constant C > 0 there exists an A (C) depending only on C so that

log d
,(11) f(n, n+d)>A(C) (yi+d)
d
(11) shows that for d-< Cn (10) gives the correct order of magnitude for f(n, m) .
12 P . Erdős and L . Pósa : On circuits of a graph

From Theorem 4 . every G„ +d contains c ; d/log d weakly independent circuits,

thus at least one of them has length not exceeding
(n+d) log d (n+d) log d
, c4
cs d d
for sufficiently large c, which proves (10) .
We shall only outline the proof of (11) . Assume first d < Cn . By the same
4
method as used in Lemma 4 . we can construct a G;,"ld the smallest circuit of which
n
has more than c 6 log edges (c, depends on C and tends to 0 as C tends to infinity),
which implies (11) by a simple calculation .
Assume next d < 4. By Lemma 4, there exists a Gzd1 all circuits of which have

length =c ; log d. Put on each edge of this graph

Thus we obtain a graph of m : n vertices and m +d
~d J- 1 vertices of valency 2 .
edges the smallest circuit of
which has length not less than

(t~+d log d
(12) logQd>c,
\Cz,I-l}cs
By adding a path of n-d new edges and vertices to this graph we obtain a
Gn"+d the shortest circuit of which satisfies the inequality (12), thus (11) and the-
refore Theorem 5, is proved,
1t would be easy to strengthen Lemma 3, as follows : Let C---~- then there
exists an e c which tends to 0~ as C tends to infinity so that every G'rc, l contains
a circuit of length less than e c log n, but we are far from being able to determine
the exact dependence of sc from C,

Bibliography

[1] P . ERDŐS and T . GALLAI, On maximal paths and circuits of graphs, Acta _Path
. Acad. Sci . Hun-
. 10 (1959), 337-356 .
gar
[2] G. DIRAC, In abstrakten Graphen vorhandene vollstandige 4-Graphen and ihre Unterteilun-
gen, Math. Nachr . 22 (1960), 61 -85 .
[3] P . ERDŐS, Graph theory and probability, Canad. J. Math . 11 (1959), 34-38 .

(Received September 30, 1961 .)