On random graphs I.

Dedicated to 0, Vargo, at the occasion of his 50th birthday.

By P. ERDdS and A. R&WI (Budapest).

Let us consider a random graph r,:l,~v having n possible (labelled)

vertices and N edges; in other words, let us choose at random (with equal

probabilities) one of the (I; possible graphs which can be formed from
t N, 1
the n (labelled) vertices PI, P?, . . ., P,, by selecting N edges from the ;
i)
possible edges Px (1 5 i < j 5 n), Thus the effective number of vertices of
C,,,l- may be less than n, as some points Pi may be not connected in r,, X
with any other point P,; we shall call such points Pi isolafed points. We
consider the isolated points also as belonging to Z7,,~Xy.r,,, X is called com-
pletely connected if it effectively contains all points PI, Pt, . . . , P,, (i, e. if it
has no isolated points) and is connected in the ordinary sense. In the present
paper we consider asymptotic statistical properties of random graphs for
11++ 30. We shall deal with the following questions:
1. What is the probability of r,,. T being completely connected?
2. What is the probability that the greatest connected component (sub-
graph) of r,,, s should have effectively n-k points? (k=O, 1, . . .).
3. What is the probability that rp,N should consist of exactly kf I
connected components? (k = 0, 1, . +.).
4. If the edges of a graph with n vertices are chosen successively so
that after each step every edge which has not yet been chosen has the same
probability to be chosen as the next, and if we continue this process until
the graph becomes completely connected, what is the probability that the
number of necessary sfeps v will be equal to a given number I?
As (partial) answers to the above questions we prove ihe following
four theorems. In Theorems 1, 2, and 3 we use the notation

-&n log n+cn

N,=
(I 1
where c is an arbitrary fixed real number ([xl denotes the integer part of x).
 P. Erdijs and A. RCnyi: On random graphs I. 291

Theorem 1. Let P,(n, N,) denofe the probability of r,,, A7cbeing corn-.
pletety connected. Then we have

Theorem 2. Let P,(n, NJ (k= 0, 1, . . .) denote the probability of the

greatest connecfed component of r,,. A: consisting effectively of n-k poinfs.
(Clearly P,(n, N,) has the same meaning as before.) Tiren we have
(e-oL)?<e-c-)f
(3) lim Pk(n, A&)= k, ,
ii-t-02
i. e. the number of points outside the greatest connected componenf of r,,, N
is disiributed in the limit according to Poissons law wifh mean value e-~2,
Theorem 3. Let I&($ N,) denote the probability of Tit: ~7~consisting
exactly of k + 1 disjoint connected components. (Clearly I&(n, I%) = Po(n, N,).)
Then we have
(e-l,.),;e-e. L
(4) lim Z&~(n, NJ = k, ,
,l->+C
i. e. the number of connected components of r7,Ys diminished by one is in the
limif distributed according to Poissons law with mean value e-zi!.
Theorem 4, Let the edges of a random graph with possible vertices
PI, p2, .., P,I be chosen successively among the possible edges PF; in such
a manner, that at each stage every edge which has not yet been chosen has
the same probability to be chosen as the next, and let us continue this process
until the graph becomes (completely) connected. Let I,, denote the number of
edges of the resulting connected random graph r. Then we have

(5)
for 11;= O(n) and
I,~-- + n log n
-al.
(6) lim P <x =e-c - .
)I++a, i n 1
As regards the first question, previously P. ERDQS and H. WHITNEY have
obtained some less complete results. They proved that if N >
where e > 0 then the probability of r,?;N being connected tends to 1 if
292 P. Erd6s and A. Rhyi

n + + DO, but if N < G--E n log n with e > 0 then the probability of r,,, 3v
( i
being connected, tends to 0 if n --+ + 00. They did not publish their result.
This result is contained in Theorem 1.
Let G,,, N denote a graph having n vertices and N edges. Let C(n, N)
denote the number of (completely) connected graphs GN. Of course the
solution of the above problems would be easy if a simple formula were
known for C(r;, N). As a matter of fact, e. g. the probability P(,(n, N) of I;,.,?
being completely connected is given simply by

Unfortunately such a simple formula is not known. The special case N= n-l
has been considered already by A. CAYLEY[l] who proved that C(n, n-1) =
= w-~ (for other proofs see 0. DZIOBEK [2], A. PR~IJFER [3], Ci. P~LYA [4]).
For the general case, it has been shown by R. I. RIDDELLand G. E. UHLEG
BECK[5] (see also E. N. GILBERT [SJ), that

Relation (8) is a consequence of the following recursive formula for C(n, N):

Unfortunately neither (8) nor (9) helps much to deduce the asymptotic pro-
perties of C(n, N). In the present paper we follow a more direct approach.
All four theorems are based on the rather surprising Lemma given below.
Let us call the graph r?,, X~ with the n possible vertices P,, p2, . . ., P,,
and N, edges, of type A, if it consists of a connected graph having n-k
effective vertices and of k isolated points (k = 0, 1, . . .). Any graph m, Sc
which is not of type A shall be called to be of type 2. Then the following
lemma holds.
Lemma. Let P(A, n, NJ denote the probability of I,,, ~7~being of type A.
Then we have
(10) lim P(A, n, N,) = 0.
?I++@2
Thus for a large n almost all graphs r,,,, are of type A.
 On random graphs I, 293

PROOF OF THE LEMMA. Let M be a large positive number, which will

be chosen later. All graphs G,,,x~ having the possible vertices PI, P?, . . ,, P,
and NC edges, where N, is defined by (1) shall be divided into two classes:
Those in which the greatest connected componeni consists of not less than
n-M points shall constitute the class Ex, all others the class EH. Let
cM(EAV,n, NC) denote the number of graphs G,,, belonging to the class &,
i. e. the number of graphs in which the greatest connected component con-
sists of less then n-M points.
If the graph consists of f connected components having Ii points
(i= 1, 2,. . ., I-) then

and therefore if L= max & we have

(0

--L--l ,a
2 -n
2 A?,
and thus L>---- this implies that
n

Therefore we have

because if the n--s points belonging to the greatest connected component

of the graph are fixed, the a@--s) edges connecting one of the points of
this component with a point outside this component can not belong to the
graph. Thus, denoting by P(.&, n, NJ the probability of F,,:N, belonging to
the class &, we have

D 20
294 P. Erdtk and A. Rhyi

We obtain for n 2 n,, by using some elementary estimations,

-s{n-s)
0;
n NC e(3-2c)s

(14) 5 for s&-

i;is i S! L

02
t NC1
and

02-s(n-s) ~ &2c)(n-s)
Hi
n
s i NC

02
= (n-s)! for s&- 2

i NC 1
We obtain from (13), (14) and (15)

(16) for nSn,.

Thus we have

Thus to prove our Lemma it is sufficient to show that denoting by 2 EIOglog1E

the- class of those graphs which belong to both x and Elog~ogn and by
P(AG,, logn) the probability that r,, ATC
belongs to the class ~ElogloglL we have

Now we have

(1%

because if the n-s points forming the greatest connected component of a

graph belonging to the class %Z loglOgraare fixed, then if r is the number of
edges connecting some of the s points outside this connected component we

must have ~ZG 1, and the r edges in question can be chosen in ways,
 On random graphs I. 295

and the remaining NC- r edges which connect points of the connected com-

ponent having n --s points can be chosen in less than ways. Thus

we obtain

which proves (18). Thus our lemma is proved.

Now we turn to the proof of our theorems.
PROOF OF THEOREM 1. Let 9&,(12,N,) denote the number of completely
connected graphs Gn, 3~~.Then 9$(,, IV,) is equal to the number of graphs
G,,, XCof type A having no isolated points. Now let %,(n, N,) denote the
number of all graphs G,,, xc (including graphs of type A) having no isolated
point. Then clearly

As the Ieft hand side of (21) is contained between any two consecutive par-
tial sums of the sum on the right of (al), and for any fixed value of k we
have

(22) lim
Il-+fCC

Thus we obtain
2 (-lye-
= & .
(23) k!
I,=I,

But clearly

(24

Thus, applying our Lemma, Theorem 1 follows,

296 P. Ertiiis and A. RCnyi

PROOF OF THEOREM 2. According to our Lemma we have to consider

only graphs of type A. Now the number of graphs G,, h-c of the type A
having a connected component consisting of n-k points is clearly equal to
multiplied by the number of connected graphs Gn+ x~. Thus it follows that
0;

(25) P,(n-k, NJ.

Taking into account that for a fixed value of k

Nc-;(n-k) lo&+-k)
lim -ec
12*m n-k
with respect to (22) and (2) we obtain the assertion of Theorem 2. I)
PROOF OF THEOREM 3. As we may restrict ourselves to graphs of type
A, Theorem 3 follows from Theorem 2 immediately.
PROOF OF THEOREM 4. Clearly if Y, = -2-n log n + 1=N+ 1 then be-
[ I
fore choosing the last edge, we had a disconnected graph G,z,s which can
be made completely connected by adding one edge. With respect to our
Lemma we may suppose that G,,,x consists of a connected graph having
n-l vertices and an isolated point. As the last edge can be chosen in n-l
ways among the remaining -N edges it follows from Theorem 2

provided that J- IS bounded.

1) CompareTheorem 2 with the following known result (see [7], Exercise No. 7 of
Chapter IV, p. 134): If N balls are distributed at random in n boxes and R&I, N) denotes
the probability that exactly k boxes remain empty, then we have for any fixed real x and
for k=O, I,...
 On random graphs I. 297

Thus (5) is proved. To prove (6) we have only to sum the probabi-
lities (5) for I < nx and obtain

m r~ e-t-*-tdf.
J
-m
As however
2.r OD
^
e-t-"-tdt - e-1'd11 = p-b
(28) J -i
.- m ,r-3;2

Theorem 4 is proved.
The following more general questions can be asked: Consider the ran-
dom graph r,,. N(,~) with n possible vertices and N(n) edges. What is the
distribution of the number of vertices of the greatest connected component
of r,?, x(,~)and the distribution of the number of its components? What is the
typical structure of r,?, A-(11j(in the sense in which, according to our Lemma,
the typical structure of r?I,NP is that it belongs to type A)? We have solved
these problems in the present paper only in the case N(n) = + n log n fen.
We schall return to the general case in an other paper [S].

Bibliography.
[I] A. CAYLEY, A theorem on trees, Quaff. J. Pure A@, Math. 23 (1889), 376-378.
(See also The Collected Papers of A. Cuyley, Cambridge 1897, Vol. 13, pp.
26-28.)
[2] 0. DZIOBEK, Eine Formel der Substitionstheorie, S.-B. Berlin. Mafk. Ges. 17 (1917)
64-67.
[3] A. PROFER, Neuer Beweis eines Sakes fiber Permutationen, Arch. Mafh. Pkys. 27 (1918),
142-144.
[4] G. P~LYA, Kombinaforische Anzahlbestimmungen fur Gruppen, Graphen und chemische
Verbindungen, Acfa Mafk. 68 (1937), 145-255.
[5] R. I. RIDDELL, JR. and G. E. UHLENBECK, On the theory of the virial development of the
equation of state of monoatomic gases. J. Ckem. Pkys. 21 (1953), 2056-2064.
[6] E. N. GILBERT, Enumeration of labelled graphs, Canad. J. Mafk. 8 (1956), 405-411.
[7] A. R~NYI, Val6szinits&gsz~mitis, (Textbook of probability theory) Budapest, 1954 (Hun-
garian).
[S] P. ERD& and A. RBNYI, On the evolution of random graphs, Publications of tke Matke-
mafical Insfifufe of the Hungarian Academy of Sciences 5 (1960) (in print).

[Received November 19, 1958.)