ON THE EVOLUTION OF RANDOM GRAPHS

by
P . ERDŐS and A . RÉNYI
Dedicated to Professor P . Turán at
his 50th birthday.

Introduction

Our aim is to study the probable structure of a random graph rn N

which has n given labelled vertices P, P2 , . . . , Pn and N edges ; we suppose
n
that these N edges are chosen at random among the l 1 possible edges,
2

so that all ~ 2 = Cn,n possible choices are supposed to be equiprobable . Thus

1V
if G,,,,, denotes any one of the C, ,N graphs formed from n given labelled points
and having N edges, the probability that the random graph -Pn,N is identical
with G,,,N is 1 . If A is a property which a graph may or may not possess,
Cn,N
we denote by P n N (A) the probability that the random graph T. ,N possesses
the property A, i. e . we put Pn,N (A) = An ' N where A n,N denotes the
Cn N
number of those G n,N which have the property A .
An other equivalent formulation is the following : Let us suppose that
n labelled vertices P,, P 2, . . ., Pn are given . Let us choose at random an edge
among the n possible edges, so that all these edges are equiprobable . After
l2I
this let us choose an other edge among the remaining I n - 1 edges, and
continue this process so that if already k edges are fixed, any of the remaining
(n) k edges have equal probabilities to be chosen as the next one . We shall
2 -
study the "evolution" of such a random graph if N is increased . In this investi-
gation we endeavour to find what is the "typical" structure at a given stage
of evolution (i . e . if N is equal, or asymptotically equal, to a given function
N(n) of n) . By a "typical" structure we mean such a structure the probability
of which tends to 1 if n -* + - when N = N(n) . If A is such a property
that lim P n,N,(n ) ( A) = 1, we shall say that „almost all" graphs Gn,N(n)
n---
possess this property .

17

2 A Matematikai Kutató Intézet Közleménye! V. A/1-2.

18 ERDŐS-RÉNYi

The study of the evolution of graphs leads to rather surprising results .

For a number of fundamental structural properties A there exists a function
A(n) tending monotonically to + - for n -i- - such that

X(v)
if lim = 0
i
n- A(n)
(1) lim Pn,N(n) (A) _
n-+- N
if lim (n) __ +co .
n- A(n)

If such a function A(n) exists we shall call it a "threshold function" of the

property A.
In many cases besides (1) it is also true that there exists a probability
distribution function F(x) so that if 0 < x < + - and x is a point of conti-
nuity of F(x) then

(2) "Mm Pn,N(n)(A)- F(x)

n-•+
if limA(n) = x .
n->-
n)

If (2) holds we shall say that A(n) is a „regular threshold function" for the
property A and call the function F(x) the threshold distribution function of the
property A .
For certain properties A there exist two functions A,(n) and A2(n)
Á2(n)
both tending monotonically to +- for n-->-+-, and satisfying lim = 0,
n- - Al(n)
such that
(n) - A,(n) _ -
if lim
i n-+- A2(n)
(3) lim Pn,N(n)(A)
n-.- N(n) - Al(n) _ +
if lim 00 .
n--+a A2 (n)
Clearly (3) implies that
N(n)
0 if lim sup < 1
n-a+m Al(n)
(4) lim Pn,N(n)(A) _
n-.- +(n)
~'
1 if liro inf > 1.
n - +ro A,.(n)

If (3) holds we call the pair (Al(n), A2(n)) a pair of °sharp threshold"-functions
of the property A . It follows from (4) that if (Aj(n), A2 (n)) is a pair of sharp
threshold functions for the property A then A, (n) is an (ordinary) threshold
function for the property A and the threshold distribution function figuring
in (2) is the degenerated distribution function

jr t ( x) _ 1
0 for x < 1
(1 for x>1
 ON THE EVOLUTION OF RANDOM GRAPHS 19

and convergence in (2) takes place for every x 1 . In some cases besides
(3) it is also true that there exists a probability distribution function G(y)
defined for -- < y < -)- - such that if y is a point of continuity of G(y) then

NT(n) - A,(n) =
(ő) lim Pn .N(n)(A) = G(y) if lím y
n- n-- x A (n)

if (ő) holds we shall say that we have a regular sharp threshold and shall call
G(y) the sharp-threshold distribution function of the property A .
One of our chief aims will be to determine the threshold respectively
sharp threshold functions, and the corresponding distribution functions for
the most obvious structural properties, e . g. the presence in rn N of subgraphs
of a given type (trees, cycles of given order, complete subgraphs etc .) further
for certain global properties of the graph (connectedness, total number of
connected components, etc . ).
In a previous paper [7] we have considered a special problem of this
type ; we have shown that denoting by C the property that the graph is con-
nected, the pair C,(n) - 12 n log n. C2(n) = n is a pair of strong threshold
functions for the property C, and the corresponding sharp-threshold distri-
bution function is e -2v ; thus we have proved' that putting
A(12)
AT = 1 n log n + y n+ o(n.) we have
a
(Ó) lim PnMn)(C) = e_e 2Y (- - < y < + -)
n- -

In the present paper we consider the evolution of a random graph in a

more systematicc manner and try to describe the gradual development and
step-by-step unravelling of the complex structure of the graph F. ,N when
N increases while n is a given large number .
We succeeded in revealing the emergence of certain structural properties
of -V N. However a great deal remains to be done in this field . We shall call in
§ 10 . the attention of the reader to certain unsolved problems . It seems to us
further that it would be worth while to consider besides graphs also more
complex structures from the same point of view, i. e . to investigate the laws
governing their evolution in a similar spirit . This may be interesting not only
from a purely mathematical point of view . In fact, the evolution of graphs
may be considered as a rather simplified model of the evolution of certain
communication nets (railway, road or electric network systems, etc .) of a country
or some other unit . (Of course, if one aims at describing such a real situation,
one should replace the hypothesis of equiprobability of all connections by
some more realistic hypothesis .) It seems plausible that by considering the
random growth of more complicated structures (e . g . structures consisting
of different sorts of "points" and connections of different types) one could
obtain fairly reasonable models of more complex real growth processes (e . g .

i Partial result on this problem has been obtained already in 1939 by P . ERDős
and 11 . WHITNEY but their results have not been published .

2*
20 ERDÖS-RÉNYI

the growth of a complex communication net consisting of different types of

connections, and even of organic structures of living matter, etc .) .
§ § 1-3 . contain the discussion of the presence of certain components
in a random graph, while § § 4-9 . investigate certain global properties of a
random graph . Most of our investigations deal with the case when N(n) - enn
with c > 0 . In fact our results give a clear picture of the evolution of rn N(n)
when c = N(n) (which plays in a certain sense the role of time) increases .
n
In § 10 . we make some further remarks and mention some unsolved problems .
Our investigation belongs to the combinatorical theory of graphs,
which has a fairly large literature . The first who enumerated the number
of possible graphs with a given structure was A . CaYLEY [1] . Next the impor-
tant paper [2] of G . PÓLYA has to be mentioned, the starting point of which
were some chemical problems . Among more recent results we mention the
papers of G . E . UHLENBEcK and G . W . FORD [5] and E . N . GILBERT [G] .
A fairly complete bibliography will be given in a paper of F . HaRARY [8] .
In these papers the probabilistic point of view was not explicitly emphasized .
This has been done in the paper [9] of one of the authors, but the aim of the
probabilistic treatment was there different : the existence of certain types
of graphs has been shown by proving that their probability is positive . Random
trees have been considered in [ 14] .
In a recent paper [10] T . L . AusTIN, R. E . FAGEN, W. F . PENNEY and
J . RioRDAN deal with random graphs from a point of view similar to ours .
The difference between the definition of a random graph in [10] and in the
present paper consists in that in [10] it is admitted that two points should
be connected by more than one edge ("parallel" edges) . Thus in [10] it is
supposed that after a certain number of edges have already been selected,
the next edge to be selected may be any of the possible n edges between
l 21
the n given points (including the edges already selected) . Let us denote such
a random graph by I n N . The difference between the probable properties
of rn N resp, rn,N are in most (but not in all) cases negligible . The correspond-
ing probabilities are in general (if the number N of edges is not too large)
asymptotically equal . There is a third possible point of view which is in most
cases almost equivalent with these two ; we may suppose that for each pair
of n given points it is determined by a chance process whether the edge
connecting the two points should be selected or not, the probability for select-
ing any given edge being equal to the same number p > 0, and the decisions
concerning the different edges being completely independent . In this case of
course the number of edges is a random variable, haying the expectation
(nl p
; thus if we want to obtain by this method a random graph having in
2
N
the mean N edges we have to choose the value of p equal to . We shall
f~l
denote such a random graph by F, N. In many (though not all) of the problems
treated in the present paper it does not cause any essential difference if we
consider instead of rn N the random graph Q*X, .
 ON THE EVOLUTION OF RANDOM GRAPHS 21

Comparing the method of the present paper with that of [10] it should
be pointed out that our aim is to obtain threshold functions resp . distributions,
and thus we are interested in asymptotic formulae for the probabilities con-
sidered. Exact formulae are of interest to us only so far as they help in determi-
ning the asymptotic behaviour of the probabilities considered (which is
rarely the case in this field, as the exact formulae are in most cases too compli-
cated). On the other hand in [10] the emphasis is on exact formulae resp .
on generating functions . The only exception is the average number of connected
components, for the asymptotic evaluation of which a way is indicated in
§ 5 . of [10] ; this question is however more fully discussed in the present paper
and our results go beyond that of [10] . -Moreover, we consider not only the
number but also the character of the components . Thus for instance we
point out the remarkable change occuring at N - n . If L-' - ne with c < 1/2
2
then with probability tending to 1 for n + - all points except a bounded
number of points of rn,N belong to components which are trees, while for
N tint with c > 1
this is no longer the case . Further for a fixed value of
2
n the average number of components of rn,N decreases asymptotically in a
linear manner with N, when N :< n,
while for N > n the formula giving
2 2
the average number of components is not linear in N .
In what follows we shall make use of the symbols O and o . As usually
a(n) = a (b(n)) (where b(n) > 0 for n = 1, 2, . . .~ means that Mim lI b(n)l - 0,
n-+-
la(n)j .
while a(n) = 0 (b(n)) means that is bounded . The parameters on
b(n)
ja(n) 1

which the bound of may depend will be indicated if it is necessary ;

b(n)
sometimes we will indicate it by an index . Thus a(n) = 0, (b(n)) means that
~a(n)I <
K(E) where K(E) is a positive constant depending on e . We write
b(n)
a(n)
a(n) -b(n) to denote that Mim = 1.
n + . b(n)
We shall use the following definitions from the theory of graphs . (For
the general theory see [3] and [4] .)
A finite non-empty set V of labelled points P l, P2i . . . , Pn and a set
E of different unordered pairs (P;, Pj) with P1 E V, Pi E V, i j is called
a graph ; we denote it sometimes by G = {V, E} ; the number n is called
the order (or size) of the graph ; the points Pl, P2, . . . , Pn are called the vertices
and the pairs (Pi, Pj) the edges of the graph . Thus we consider non-oriented
finite graphs without parallel edges and without slings . The set E may be empty,
thus a collection of points (especially a single point) is also a graph .
A graph G2 - {V2, E2} is called a subgraph of a graph Gl = {Vl, El }
if the set of vertices V 2 of G 2 Ms a subset of the set of vertices V, of G l and the
set EZ of edges of G2 is a subset of the set El of edges of G I .
22 ERDŐS-RÉNYi

A sequence of k edges of a graph such that every two consecutive edges

and only these have a vertex in common is called a path of order k .
A cyclic sequence of k edges of a graph such that every two
consecutive edges and only these have a common vertex is called a cycle of
order k .
A graph G is called connected if any two of its points belong to a path
which is a subgraph of G .
A graph is called a tree of order (or size) k if it has k° vertices, is connected
and if none of its subgraphs is a cycle . A tree of order k has evidently k - 1
edges .
A graph is called a complete graph of order ~, if it has k vertices and

2 edges . Thus in a complete graph of order k any two points are connected
by an edge .
A subgraph G' of a graph G will be called an isolated subgraph if all
edges of G one or both endpoints of which belong to G', belong to G' . A con-
nected isolated subgraph G' of a graph G is called a component of G . The
number of points belonging to a component G' of a graph G will be called the
size of G' .
Two graphs shall be called isomorphic, if there exists a one-to-one mapp-
ing of the vertices carrying over these graphs into another .
The graph G shall be called complementary graph of G if G consists
of the same vertices P l , P2, . . . . Pn as G and of those and only those edges
(P;, Pj ) which do not occur in G .
The number of edges starting from the point P of a graph G will be called
the degree of P in G .
A graph G is called a saturated even graph of type (a, b) if it consists of
a + b points and its points can be split in two subsets V, and V2 consisting
of a resp . b points, such that G contains any edge (P, Q) with P E V, and
Q E V2 and no other edge .
A graph is called planar, if it can be drawn on the plane so that no two
of its edges intersect .
We introduce further the following definitions : If a graph G has n
2 N
vertices and N edges, we call the number the "degree" of the graph.
V
21ti
(As a matter of fact is the average degree of the vertices of G .) If a graph
n
G has the property that G has no subgraph having a larger degree than G
itself, we call G a balanced graph .
We denote by P ( . . . ) the probability- of the event in the brackets, by
M(~) resp . D 2 (á) the mean value resp . variance of the random variable s .
In cases when it is not clear from the context in which probability space the
probabilities or respectively the mean values and variances are to be under-
stood, this will be explicitly indicated . Especially M n,N resp. Dn ,N will denote
the mean value resp . variance calculated with respect to the probabilities
Pn .N •
 ON THE EVOLUTION OF RANDOM GRAPHS 23

We shall often use the following elementary asymptotic formula :

k'- 0
in! nk e 2n 6n=
(7) valid for k = o(n'l=) .
k
Our thanks are due to T . GALLAI for his valuable remarks .

§ 1. Thresholds for subgraphs of given type

If N is very small compared with n, namely if N - o (Vn) then it is
very probable that rn .N is a collection of isolated points and isolated edges,
i. e. that no two edges of rn,N have a point in common . As a matter of fact
the probability that at least two edges of rn , N shall have a point in common
is by (7) clearly
n
2N)
(1T2
.
1-~ 2 ( =0
n lnj

1'
Y1

If however íl c I/Inn where c > 0 is a constant not depending on n., then the
appearance of trees of order 3 will have a probability which tends to a posi-
tive limit for n -* + -, but the appearance of a connected component consist-
ing of more than 3 points will be still very improbable . If N is increased while n
is fixed, the situation will change only if N reaches the order of magnitude
of n 2j3 . Then trees of order 4 (but not of higher order) will appear with a pro-
bability not tending to 0 . In general, the threshold function for the presence
k-2
of trees of order k is nk -1 (k = 3, 4, . . . ) . This result is contained in the
following
Theorem 1 . Let k >_
_ 2 and l (k - 1 < Z < 9 I be positive integers . Let
l lk
k„ denote an arbitrary not empty class of connected balanced graphs consisting
of k points and l edges . The threshold function for the property that the random
graph considered should contain at least one subgraph isomorphic with some ele-
2- k

ment of y k,, is n t .

The following special cases are worth mentioning

Corollary 1 . The threshold function for the property that the random graph
k-2
contains a subgraph which is a tree of order k is nk - (k: = 3, 4 . . . . ) .
I

Corollary 2. The threshold function for the property that a graph contains
a connected subgraph consisting of k >_ 3 points and k edges (i . e . containing
exactly one cycle) is n, for each value of k .
Corollary 3. The threshold function for the property that a graph contains
a cycle of order k is n, for each value of k _>_. 3 .
24 ERDŐS-RÉNYI

Corollary 4. The threshold function for the property that a graph contains
- 1 )
a complete subgraph of order k _> 3 is n z (i k-1, .
Corollary 5. The threshold function for the property that a graph contains
a saturated even subgraph of type (a, b) (i . e . a subgraph consisting of a + b
a+b
points P1 , . . . , Pa , Q I , . . . Q b and of the ab edges (P1 , Qj ) is n2 ab .
To deduce these Corollaries one has only to verify that all 5 types of
graphs figuring in Corollaries 1-5 . are balanced, which is easily seen .
Proof of Theorem 1 . Let Bk I >_ 1 denote the number of graphs belong-
ing to the class which can be formed from k given labelled points . Clearly
if Pn,N (&k,l) denotes the probability that the random graph rn,N contains
at least one subgraph isomorphic with some element of the class ^Vjk,l, then

nk 7-
Pn,N(&k,l) < Bk,1 l =o
(n n`
21-kJ
I
12
N

As a matter of fact if we select k points (which can be done in different

171
ways) and form from them a graph isomorphic with some element of the class
66k,í (which can be done in B k,l different ways) then the number of graphs
Gn ,N which contain the selected graph as a subgraph is equal to the number
of ways the remaining N-1 edges can be selected from the I other
2 n - l
possible edges . (Of course those graphs, which contain more subgraphs iso-
morphic with some element of k,1 are counted more than once .)
k
Now clearly if N = o(n2_ 1) then by

Pn,N( k,l) = O(1)

which proves the first part of the assertion of Theorem 1 . To prove the second
part of the theorem let -Vki denote the set of all subgraphs of the complete
graph consisting of n points, isomorphic with some element of 56 k 1 . To any
SE& let us associate a random variable e(S) such that E(S) = 1 or e(S) = 0
according to whether S is a subgraph of Pn N or not . Then clearly (we write
in what follows for the sake of brevity W instead of Mn,N)
n
l

(n) Bk l N-l, Bk,l (2 N)' .

(1 .2) M (~ E(S)1 = M(E(S)) _
n n21-k
k k I
SE, ik(n)1 (n)
S E.9ik,1 ~9
 ON THE EVOLUTION OF RANDOM GRAPHS 25

On the other hand if SI and S2 are two elements of Z01)1 and if Sl and
82 do not contain a common edge then

f2J
N-21
-2l

M(E(SI) E(S2» = n,~

~2
N,

If S,_ and S2 contain exactly s common points and r common edges (1 < r < l -1)
we have

-21+r
j 2nl
N - 21 + r - N21-r
M(s(Sl ) E(S2)) - .
n
0
f n41-2r
(l2
N
On the other hand the intersection of S I and S2 being a subgraph of S I (and S2 )
by our supposition that each S is balanced, we obtain r < l i . e . s z rk
S - k l
and thus the number of such pairs of subgraphs S I and S2 does not exceed

n-k 2k _rk
k n k
0 n
B2 `k kf k-~~ 1 .
Jz i
Thus we obtain

M (~E E(S)~~)
)
k,1

n
- 21

n
M(e(S)) +
n!Bk,1
k!2( n - 2 k)!
~2'
N
n
21
+OI
n
N1 kJ2 (n 2- 1 r
N
k

~
SE i k, 1
l2
;}T
1
Now clearly
In 2
1
2~
- 21 -l
lnJ
n! N -21 n 2 N-1
<
k! 2 (n - 2 k)! n' ~ k n 2
~2I ~21
1-
 ERDŐS--RÉNYI
26

If we suppose that

n
it follows that we have
M(8 (8»2

(Z
r SE&(k,11
(1 .4) D2 1
e(S}) =0
I\ SEM
\ (n) ( CU

It follows by the inequality of Chebysheff that

Pn,N ÍZ
ffi
e(S) - Z M(£(S)) > 1 .2 M(e(S»
SE k,1 SEM k,1 sETik ,1

and thus

(1 .5) Pn,N
L SEM (r' ),
e(S) < 1 2 M(E(S »)
2 SE Y~knj
=O
GJ
l1
I.
As clearly by (1 .2) if w -~ + then n M(e(S)) + - it follows not only
k,1
that the probability that 1'„ contains at least one subgraph isomorphic
with an element of k,I tends to 1, but also that with probability tending
to 1 the number of subgraphs of Fn N isomorphic to some element of ti 1; 1
will tend to +O° with the same order of magnitude as GJ 1 .
Thus Theorem 1 is proved .
It is interesting to compare the thresholds for the appearance of a sub-
graph of a certain type in the above sense with probability near to 1, with
the number of edges which is needed in order that the graph should have
necessarily a subgraph of the given type . Such -compulsory" thresholds
have been considered by P . TURÁN [11] (see also [12]) and later by P . ERDŐS
and A . 11 . STONE [ 17 ]) . For instance for a tree of order k clearly the compulsory
n(k - 2) ]
threshold is + 1 ; for the presence of at least one cycle the com-
I 2
pulsory threshold is n while according to a theorem of P . TURÁN [11] for
(k?)
complete subgraphs of order k the compulsory threshold is (n2 - r2) +
2(k - 1)
-v
where r=-n- (k-- 1) ~- In the paper [13] of T . KŐVÁ RI,
~21 Is - 11'
V. T . Sós and P . TuRÁN it has been shown that the compulsory threshold
for the presence of a saturated even subgraph of type (a, a) is of order of magni-
tude not greater than n2 a . In all cases the "compulsory" thresholds in
TURÁN's sense are of greater order of magnitude as our "probable" thresholds .
 O\ THE EVOLUTION OF RANDOM GRAPHS 27

§ 2 . Trees
Now let us turn to the determination of threshold distribution functions
for trees of a given order . We shall prove somewhat more, namely that if
k-2
N e nk -1 where o > 0, then the number of trees of order k contained
in rn,N has in the limit for n -y + oo a Poisson distribution with mean value
_ (2 P )k 1 kk-z .
This implies that the threshold distribution function for
V
trees of order k is 1 - e - ~ .
In proving this we shall count only isolated trees of order k in Fn N , i- e.
trees of order k which are isolated subgraphs of I'n N' According to Theorem 1 .
this makes no essential difference, because if there would be a tree of order
k which is a subgraph but not an isolated subgraph of P n N , then rr,,N would
have a connected subgraph consisting of k + 1 points and the probability-
(
of this is tending to 0 if N = o l -n. k which condition is fulfilled in our
k-2
case as we suppose N - g .n k-I .
Thus we prove
N(
Theorem 2a . If lim k-2 = e > 0 and -r k denotes the number of isolated
n -~ n k-1

trees of order k in Fr, N(n)

then
A1
(2 .1) "M
p n„ti'(n)(t1, -- Í ) -- e- ~
n--, j'.

or j = 0, 1, . where
- (2 Q )k-1 kk-2
A
k!

For the proof « e need the following

Lemma 1 . Let E n1' eng- ' Enr,z be sets of random variables onn some pro-
bability space ; suppose that En (1 < i <. l n ) takes on only the values 1 and 0 . If

hm M (-'ni, Eni, . . . Eni,) - ,

r!

uniformly in . r for r = 1, 2, . . , , where A > 0 and the summation is extended

over all combinations (i l , i2 , . . . , ir ) of order r of the integers 1, 2, . . . , l n , then

r„ dJe - x
(2 .4) lim p ~~ -'n i = jl= (j = 0, 1, . . . )
n -. + . i=1 j!

i. e . the distribution of the sum Poisson-

ER , tends for n-~ + co to the

distri-bution with mean value ~ .

28 E RDŐS-RÉN YI

Proof of Lemma 1 . Let us put

(2 .5)
P11(j) = P ~Z
i-t
1„

Ene = ! . I
Clearly
(2:6) M
1Si,<iz< . . .<i .<1n j°r
thus it follows from (2 .3) that
- ~r
(2 .7) lim Z p, (j) (r = 1, 2, )
n>+mj=r r
l71 ri

uniformly in r .
It follows that for any z with I z < 1
(nz _~
(2 .8) hm 'Y Pn(9) II j ~ z r = ~, ()r = e' - 1
nyT' r=1 j-r lr •Í r. !
But

(2.9) Zr
,Z ~Pn(j) r1) -_ .G Pn(7) (1 + z)j - 1 .
r-1 (j=r l j=o

Thus choosing z = x - 1 with 0 < x < 1 it follows that

('2 .10) lim > Pn(?) j = e~ (x - ' ) for 0<x< 1 .

n - --m j=O
It follows easily that (2 .10) holds for x = 0 too . As a matter of fact
putting Gn (x) xi, we have for 0 < x < 1
j=o
P,(O) - e -1 I < I G,(x) - I + I G,,( :x) - P'(O) I + I e'(x-1) - e-d I .

_1s however

Gn (x) - P'(0) I- x Pn(l ) x

j --t
and similarly
ef.(x-1)
- e- I < x
it follows that
I Pn(O) - e- . I < I G'(x) - e4x-1) I - 2 x .
Thus we have
lim sup I P,(O) -- e - ' I< 2 x ;

as however x > 0 may be chosen arbitrarily small it follows that

lim P, (0) = e -2
 ON THE EVOLUTION OF RANDOM GRAPHS 29

i . e . that (2 .10) holds for x = 0 too . It follows by a well-known argument

that

(2 .11) lira P,(j) _ - , (j = 0 1 1 1 . . . ) -

n--+- f

As a matter of fact, as (2 .10) is valid for x = 0, (2 .11) holds for j = 0 . If

(2 .11) is already proved for j < s - 1 then it follows from (2 .10) that

(2 .12) lim P,(j) xi - S = xi - S for 0<x< 1 .

n-+° i=5 Í=S ~~
By the same argument as used in connection with (2 .10) we obtain that
(2 .12) holds for x = 0 too . Substituting x = 0 into (2 .12) we obtain that (2 .11)
holds for j = s too . Thus (2 .11) is proved by induction and the assertion of
Lemma 1 follows .
Proof of Theorem 2a. Let Tkn) denote the set of all trees of order k which
are subgraphs of the complete graph having the vertices P i , P2 , . . . 5 P,, .
If SETk") let the random variable e(S) be equal to 1 if S is an isolated subgraph
r
of n , N ; otherwise e(S) shall be equal to 0 . We shall show that the conditions
of Lemma 1 are satisfied for the sum _Y e(S) provided that N= N(n) -
SETkn)
k-2
enk -1 and A is defined by (2 .2) . As a matter of fact we have for any
SETk)
(n - k~
I
1 2
lk-1 e _2N'
N +1 = (2
(2 .13) M(E(S)) _ 1 +0
n It n2 J 1n
12 1
21
N

More generally if S 1 , 82 , . . . , Sr (Si E T(Rn) ) have pairwise no point in common

then clearly we have for each fixed k >_ - 1 and r -> 1 provided that n a+oo,
N- ;+O°
/n rk
I -U I

j2 ,(k-1)r - 2Nrk
Sr )~ = -IV-r(k - 1) r2 l
(2 .14) M("(8002) . . . e( e n ( 1 O( n 2, 1
n ~r
(l2!
N

where the bound of the 0 term depends only on k . If however the Si ( =

= 1, 2, . . ., r) are not pairwise disjoint, we have

(2 .15) M (e ( S i) e(S2) . . . e(S,» = 0 .

30 ERDŐS--RÉNYI

Taking into account that according to a classical formula Of CAYLEY (l]

the number of different trees which can be formed from k labelled points is
equal to kk -2 , it follows that
_ 2Nrk y,2
kk-2 rnkr ~2NII-k-'
(2 .16) M(E(Sl) E(S2 ) . . . E(S,)) _ -- e n (1 +0 ~~
k! r! n2 n2

where the summation on the left hand side is extended over all r-tuples of
trees belonging to the set T(k) and the bound of the 0-term depends only on k .
Note that (2 .16) is valid independently of how N is tending to +- . This
will be needed in the proof of Theorem 3 .
Thus we have, uniformly in r
7!r
(2 .17) lim _ M(E(S1 ) e(S2) . . . E(S,,» = for r = 1, 2, . . .
-iP
k-s
nk-i

where A is defined by (2 .2) .

Thus our Lemma 1 can be applied; as z k = e(S) Theorem 2 is
proved . SET kn)
We add some remarks on the formula, resulting from (2 .16) for r = 1

(2N _ 2N,k
ja2
I e n k;k -2
(2 .18) M(Tk) _
2N
n
k!
i+o ~n2 1~ l

kk-2 tk-1 C -kf

Let us investigate the functions mk (t) _ (k= 1, 2, . . . ) . accord-
k!
i 2-
ing to (2 .18) nmk I is asymptotically equal to the average number of trees of
1n
order k in rn,N . For a fixed value of k, considered as a function of t, the value
of mk (t)
L
increases for t < - 1 and decreases for t > k - 1 ; thus for a fixed
k k
value of n the average number of trees of order k reaches its maximum for
N -n 1 - 1 ; the value of this maximum is
2 k
1 ,k-1
'k-2
(1 - e-(k-I)
n ` -~
k.
k!
For large values of k we have evidently
n
M,*~ r
V2 7c k512
 ON THE EVOLUTION OF RANDOM GRAPHS 31

It is easy to see that for any t > 0 we have

Mk(t) >_ mk+>_(t) (k = 1, 2, . . . ) .
The functions y = m k(t) are shown on Fig . 1 .
It is natural to ask what will happen with the number r k of isolated
trees of order k contained in I'„ N if N(n) --*+ - . As the Poisson distribution
n k-1
(Ij ~~
e-f l~ is approaching the normal distribution if ~ . + -, one can guess
that r k will be approximately normally distributed . This is in fact true, and
is expressed by

Z
Fi .ure la.

i
o x (c) t 2c
F gure 1b .
32 ERDŐS-RÉNYI

Theorem 2h . If

N(n)
(2 .19) k-2~ + 00

n k-1

but at the same time

N(n) - 1 n log n -- k- l n loglog n

2k 2k
(2 .20) lim o
n-+- n
then denoting by "'k the number of disjoint trees of order k contained as subgraphs
in I'n N( n ) (k = 1, 2, . . . ), we have for - - < x < + -

{2 .21) hm P .M .) ~ Z k --3In,N(n) <X l = O (X)

n-+- Vmn,N(n) l
where
kk -2 (2N~k-1 _ 2kN
(2 .22) n e
In,N - n
k! nn
and

(2 .23) <P(x) = 1 x e - 2 du .
V2 n

Proof of Theorem 2h . Note first that the two conditions (2 .19)and

(2 .20) are equivalent to the single condition lim Mn N(n)= + and as
n-y+ m
M ( c k ) ti 1VIn,N
this means that the assertion of Theorem 2b can be expressed
by saying that the number of isolated trees of order k is asymptotically nor-
mally distributed always if n and N tend to +- so, that the average number
of such trees is also tending to +- . Let us consider
,r Z
M( k) = N1(s (n) E( S)) r )
k

Now we have evidently, using (2.16)

r ~r2JV) -r r! -1n,N
M (~k) _ ~ 1-F- 0 f1
1 n2 l hl ! h2 ! . . . hj !

where Mn,N is defined by (2 .22) . Now as well known (see [16], p . 176)

1 r!
(2 .24) - ag)
r
h l ! h 2 ! . . . h]-!
i~lRg=r, hi>1
 ON THE EVOLUTION OF RANDOM GRAPHS 33

where v(Í) are the Stirling numbers of the second kind (see e . g . [16], p . 168)
defined by
r
(2 .25) xr = 60) x(x - 1) . . . (x -
Í=i
Thus we obtain

(2 .26) M (Tk) _ (1 + 0 Ir2~ 09) 1In,N .

n2 Í=1

Now as well known (see e . g . [16], p . 202)

e a(ex 1) - 1=
+ +
orQ)
xr Aj_ _ W xr r
- d•(1)Ai
(2 .27)
Í=1 r=Í r~ r=1 91 .J=1
Thus it follows that
r dr l + a ~k e-
x
6rÍ)1 =
(2,28) , -- e .(e -1) I =
Í=1 dxr x=0 k=0k

We obtain therefrom

T, V
-~M""" ~ ~ - 1 11In,N r r2N
e Mn ry (k - MroN) ~ (1 + 0 ~- 2 ~ ~ .
(2 .29) M ~ (

n,N f 11T n,N

í 2 k=0
S k•I n

+ •. 2k
Now evidently ki e -1 (k - ~t}r is the r-th central moment of the Poisson
distribution with mean value 2 . It can be however easily verified that the
moments of the Poisson distribution appropriately normalized tend to the
corresponding moments of the normal distribution, i . e . we have for r = 1, 2, . . .
+ ~ xa
Ak e
(2 .30) lim
).y+ . f
2 k i k.
( . - a )rl = - 1 -
V2 n 1
xr e 2 dx .

In view of (2 .29) this implies the assertion of Theorem 2b .

In the case N (n) =1 n log n + k- 1 n loglog n + yn + o(n) when

2k 2k
the average number of isolated trees of order k in I'n,N(n) is again finite, the
following theorem is valid .
Theorem 2c. Let rk denote the number of isolated trees of order k in l'n,N
(k=1, 2, , . . ) . Then if

k.
(2 .31) í17 (n)= Inlogn+ I nloglogn+yn+o(n)
2k

where - - < y < + -, we have

~Í ~
(2 .32) "ln Pn,N(n) (""k = i) _ (j = 0, 1, . . . )
n +- j.
where
e -2ky
(2 .33) A=
k •k '

3 A Matematikai hutató Intézet Közleményei V . A(1--2 .

 34 ERDÖS---RÉNYI

Proof of Theorem 2c .Itisealynthuder conitsfTherm2c

hm Mn .N(n) (Tk) -
n-
Similarly from (2 .16) it follows that for r = 1, 2, . . .
~r
.N(n) ~~(sl) E(~2) . . . E(~r)) _
nhm ~ Mn r'
SIET k(n)
and the proof of Theorem 2c is completed by the use of our Lemma 1 exactly
as in the proof of Theorem 2a .
Note that Theorem 2c generalizes the results of the paper [7], where
only the case 1c = 1 is considered .

§ 3 . Cycles
Let us consider now the threshold function of cycles of a given order .
The situation is described by the following
Theorem 3a . Suppose that
(3 .1) N(n) -c n where e > 0 .
Let yk denote the number of cycles of order k contained in F, ,N (k= 3, 4, . . . ) .
Then we have
A1 e -a
(3 .2) "m Pn .N(n) (Yk 0, 1, . . .)
nom . ?
where

(3 .3) ` (2 c)k .
2k
Thus the threshold distribution corresponding to the threshold function A(n) = n

for the property that the graph contains a cycle of order k is 1 - e- 2k (2 1)k
It is interesting to compare Theorem 3a with the following two theorems :
Theorem 3b . Suppose again that (3 .1) holds . Let yk denote the number of
isolated cycles of order k contained in rn N (k = 3, 4, . . . ) . Then we have

(3 .4) lira Pn,N(n) (Yk* = j) _ UJ e u (Í = 0, 1, . . ))

n-+ ?
where
(2 r e-2c) k
(3 .5) ,u =
2k
Remark. Note that according to Theorem 3b for isolated cycles there
does not exist a threshold in the ordinary sense, as 1 - e-11 reaches its maxi-
mum 1 - e 2kek for c = 2 i . e . for N(n) -~ 2 and then again decreases ;
 ON THE EVOLUTION OF RANDOM GRAPHS 35

thus the probability that rn N contains an isolated cycle of order k never

approaches 1 .
k
Theorem 3c . Let 8 denote the number of components of -P. ,N consisting
of k > 3 points and k edges . If (3 .1) holds then we have

w1
(3 .6) "M Pn,N(n) (ak = Í) _ (9=0,1, . . .)

where
_ (2 C e-2c)k k-3
(3 .7) w (1 + k + -
Z 2 + . . . -} ~ .
2k t 2! (k-3)!

Proof of Theorems 3a ., 3h . and 3e . As from k given points one can form

(k - 1) ! cycles of order k we have evidently for fixed k and for N- 0(n)
2
~ 2n) -k 2N ,1
1 (n , N-k . L n~
M (Yk) = 1 (k - 1)!
, k
while

2ív
e-2N
nn l k
M(fh)= 1 ~ ~( k - -1)!
I n
2k

As regards Theorem 3c it is known (see [10] and [15]) that the number
of connected graphs Gk.k (i" e . the number of connected graphs consisting
of k labelled vertices and k edges) is exactly

(3 .10) ()k = 2 (k - 1)! I + k + 22 + . . . + .

( k kk 3) j
Now we have clearly

2 k}
I
1 ;~ ~, 2N k
8 n

(3 .11) M (S k)
= I
lti
- k' ~ -------
2k
- - I +k +
k2
2!
+ . . . +
kk-3

(k - 3)!,

3*
 -RÉNYI
36

For large values of k we have (see [15])

(3 .12) ~? kk
8
and thus
2N i-Zn ~k
~-
nee
(3 .13) M (k) -,• -- .
4k

For íl' 2 we obtain by some elementary computation using (7) that

n
for large values of k (such that k = o (n3-4) .
k3

e n
(3 .14) M(bk) .
4k
Using (3 .8), (3 .9) and (3 .11) the proofs of Theorems 3a, 3b and 3c follow
the same lines as that of Theorem 2a, using Lemma 1 . The details may be
left to the reader .
Similar results can be proved for other types of subgraphs, e . g. complete
subgraphs of a given order . As however these results and their proofs have
the same pattern as those given above we do not dwell on the subject any
longer and pass to investigate global properties of the random graph T,,,.

§ 4. The total number of points belonging to trees

We begin by proving
Theorem 4a. If N = o(n) the graph rn N is, with probability tending to
1 for n->- +-, the union of disjoint trees .
Proof of Theorem 4a . A graph consists of disjoint trees if and only if
there are no cycles in the graph . The number of graphs Gn,N which contain
at least one cycle can be enumerated as was shown in § 1 for each value k
o£ the length of this cycle . In this way, denoting by T the property that the
graph is a union of disjoint trees, and by T the opposite of this property,
i. e. that the graph contains at least one cycle, we have
(n)
- k
n .N-k,. _ O ~Nj .
(4 .1) P n,N (T) Ilk n
in2 )
N)
It follows that if N = o(n) we have lim P n N (T) = 1 which proves Theorem 4a .
n - •-
If N is of the same order of magnitude as n i, e . il' - c22 with c > 0,
then the assertion of Theorem 4a is no longer true . Nevertheless if c < 1/ 2,
 ON THE EVOLUTION OF RANDOM GRAPHS 37

still almost all points (in fact n - 0(1) points) of I'n N belong to isolated
trees . There is however a surprisingly abrupt change in the structure of "n,N
with 1V - en, when c surpasses the value . If c > 1/2 in the average only a
2
positive fraction of all points Of rn,N belong to isolated trees, and the value
of this fraction tends to 0 for c + .
Thus we shall prove
Theorem 4b . Let ti' n, ,, denote the number of those points of Tn N which
belong to an isolated tree contained in h n,N . Let us suppose that
N(v)
(4 .2) hill = c > 0 .
n--- n
Then we have
(/ 1 for c < 1/2
M ( y ~n .N(n)) _ 1
(4 .3) lim - x(c) 1
n- . ;_ T n for c > -
2c 2
where x = x(c) is the only root satisfying 0 < x < 1 of the equation
(4 .4) x e - X= tee-2c,
which can also be obtained as the sum of a series as follows :
kk
(4 .5) x(c) _ 1 (2 c e 2c) k
k=1 V
Proof of Theorem 4b . 'e shall need the well known fact that the inverse
function of the function
(4 .6) y = x e-X (0 :< X :< 1)

has the power series expansion, convergent for 0 < y < 1

e
.
+ kk-I y k
(4 .7) x = `'
k=1 kf

Let T k denote the number of isolated trees of order k contained in -r. , N . Then
clearly
n
(4 .8) V n dN = N k -r k
k=1
and thus
n
(4 .9) M(i'n,N) _ ` k M(T I ) .
k=1

By (2 .18), if (4 .2) holds, we have

lim
1 M _ 1 k-2
V1
(2 c e -2c)k
(4 .10) (~k)
n-+ . n 2c kI
 ERDŐS-RÉNYI
38

Thus we obtain from (4 .10) that for c < 1/2

kk-1(2 c e-2c)k
(4 .11) lim inf M(Vn,N(n)) > I for any s>
.
.1
nJ-+~ n - 2c k _ 1 k!
As (4 .11) holds for any s >_ 1 we obtain
kk -1 (2 c e -2c)k
(4 .12) lim inf M(V n,N(n)) > 1
n-+- n 2C k!

But according to (4 .7) for c < 1j2 we have

kk- 1 (2 c e-2c) k
=2c .
k=1 k!
Thus it follows from (4 .12) that for c <_ 11 2

(4 .13) lim inf M(Vn ,N(n)) >_ 1 .

n- . n

As however Vn,N(n) n and thus lim sup ~I(yrn,N(n)) < 1 it follows that
n -~ n
if (4 .2) holds and c < 1/2 we have
M (Vn,N(n)) = I .
(4 .14) lim
n->+ . n

Now let us consider the case c > 1. It follows from (2 .18) that if (4 .2)

holds with c > 1/2 we obtain

n 2Nnn)~ k
(4 .15) M (V n N(n)) =
n2 kk -1 2 N(n)
e
-
0(1)
2N k! n
where the bound of the term 0(1) depends only on c . As however for N(n) N
- nc with c > 1/2
kk -1 J2 AT(n) - 2 N(n), k
k=,,+1 e
k! n n
it follows that
)1
(4 .16) M(Vn,N(n)) = 2
2_
x
(ln +0(1)

(N(n) 1
where x = x I I is the only solution with o < x < 1 of the equation
` n J
2N(n)
2 N(n) -
xe - x = e n Thus it follows that if (4 .2) holds with c > 1/2
n
we have
M(Vn,N(n)) _ x(c)
(4 .17) lim
n-.+m n 2c
where x(c) is defined by (4.5) .
 ON THE EVOLUTION OF RANDOM GRAPHS 39

The graph of the function x(c) is shown on Fig . la ; its meaning is shown
by Fig. 11) . The function

for c < 11 2

for c > 11 2

is shown on Fig . 2a .

Figure 2a.

Y Gfc1- 1- 2c)

Figure 2b.
 ERDŐS-RÉNYI
40

Thus the proof of Theorem 41) is complete. Let us remark that in the
same way as we obtained (4 .16) we get that if (4 .2) holds with c < 1/2 we have
(4 .18) M( V n,N(n)) = n - 0(1)

where the bound of the 0(1) term depends only on c . (However (4 .18) is not,
true for c =1 as will be shown below .)
2
It follows by the well known inequality of Markov

(4 .19) P(~ > a) _<- 1 M(s)

a,
valid for any nonnegative random variable ~ and any a > M(~), that the
following theorem holds :
Theorem 4c . Let "7 n ,N denote the number of those points of rn,N which
belong to isolated trees contained in Fn v . Then if w n tends arbitrarily slowly
to +- for n --)- +- and if (4 .2) holds with c < 112 we have

(4 .20) lim P(VnMn)

> n - 0-0 = 1 .
n-• -

The case c > 1 f 2 is somewhat more involved . We prove

Theorem 4d . Let ",,,N denote the number of those points of rn,N which
belong to an isolated tree contained in . Let us suppose that (4 .2) holds with
c > 1/ 2 . It follows that if w n tends arbitrarily slowly to +we have

n2
(4 .21) "M P 1, n,N(n) -
9 N(n) x
'rl
(1 n(n)
I ~ V
n a) n

where x = x ~ N(n)-~ is the only solution with 0 < x < 1 of the equation
l n
2N(n)
Te -x - 2 ~4T(n}
_e
n
n.
kk
Proof. We have clearly, as the series } (2 ce -2c)k is convergent,
h`_, k t

D 2 (Vn,N(n)) = 0(n) . Thus (4 .21) follows by the inequality of Chebyshev .

Remark . It follows from (4 .21) that we have for any c > 112 and any
e > o

(4 .22) hm P iVmN(n) _ x(c) < e = 1

n-+- n 2c
where x(c) is defined by (4 .5) .
As regards the case c ij2 we formulate the theorem which will be
needed latter .
 ON THE EVOLUTION OF RANDOM GRAPHS 41

Theorem 4e . Let V, N (r) denote the number of those points of F, ,N which

belong to isolated trees of order >_ r and r n v (r) the number of isolated trees of
order >_ r contained in F ?i N . If N(n) n we have for any 6 > o
n,N(n)(r) 1 ~ kk 1
(4 .23) lim P e-k < = 1
n--x n. - k!
and
q kk-2
Zn,N(n)(7)
(4 .24) hm Pl - ~ 6 -h"
n-- . n k/r k!

The proof follows the same lines as those of the preceding theorems .

§ 5 . The total number of points belonging to cycles

Let us determine first the average number of all cycles in rn N . W e

prove that this number remains bounded if N(n) cn and c < 1'2 but not
if c =1f 2.
Theorem 5a . Let Hn,N denote the number of all cycles contained in F,, ,N .
Then we have if N(n) -V en holds with c <1
2
12
(5 .1)
n
'I'll M ( Hn, N (n))
Ta
= 2 log 1 c
- c - c2

while we have for c - 1

2
4
(5 .2) M(H,,N(n)) - log n .

Proof. Clearly- i£ y k is the number of all cycles of order I contained in

-r,,,N we have
n
Hn,N = G yk-
k=I

Now (5 .1) follows easily, taking into account that (see (3 .8))

k 2N k
l2l
1
k2
M (Yk) 0
2 (k~ (k ) n 2k +
~2
N
42 ERDŐS--RÉNYI

If c = 1f2 we have by (3 .8)

30
e 2n
M (Yk)
(5 .1)
2Ic
n 1 _ 3k' 1
As } 2n -loge, it follows that (5 .2) holds . Thus Theorem 5a
e
k=3 2k 4
is proved .
Let us remark that it follows from (5.2) that (4.18) is not true for c = 1 / 2 .
Similarly as before we can prove corresponding results concerning
the random variable Hn N itself .
We have for instance in the case c = 1/ 2 for any s > o

(5.5) lim P ( H n .N(n) - 1 ' < a =1 .

n-.+~ logn 4
This can be proved by the sane method as used above : estimating the variance
and using the inequality of Chebyshev .
An other related result, throwing more light on the appearance of cycles
in P n,N runs as follows .
Theorem 5h . Let K denote the property that a . graph contains at least one
cycle . Then we have if .
N(n) no holds with c 1/2

(5 .6) lira P n,N(n)( K ) = 1 - 1- 2 c e+'~Y .

n--+ z

Thus for e = 1 it is ,almost sure" that r n N(n) contains at least one cycle, while
2
for c < 1 the limit for n-> + co of the probability of this is less than 1 .
2
Proof. Let us suppose first c < . By an obvious sieve (taking into

account that according to Theorem 1 the probability that there will be in I'n,N(n)
with N(n) ~- no (c < 1/2) two circles having a point in common is negligibly
small) we obtain
_ lim M(H„,~v(„i)
(5 .7) lim Pn,N(n)(K ) = e - 1- 2 e eC+~~ .

Thus (5 .6) follows for c < 1/2 . As for c 1 /2 the function on the right of (5.6)
tends to 1, it follows that (5.6) holds for c = 1/2 too . The function y =
- 1 - V1 - 2c eC+C = is shown on Fig. 3 .
We prove now the following
Theorem 5c. Let Hn, N denote the total number of points of rnN which
belong to some cycle . Then we have for N = N(n) on with 0 e c <' 1/2
4 ca
(5 .8) lim M (H n,N(n)) -
n .+~ 1-2c
 ON THE EVOLUTION OF RANDOM GRAPHS 43

,c
tJ=1 - v1-2r, .ec

0
1
2
Figure 3 .

Proof of Theorem 5c . As according to Theorem 1 the probability that

two cycles should have a point in common is negligibly small, we have by (5 .3)

(2c) 3 4c 3
M ( Hn ,N(n)) n k y4_
Z
k= t 2(1 -2c) 1-2c

The size of that part of I'n,N which does not consist of trees is still more
clearly shown by the following
Theorem 5d. Let ~0n,N denote the number of those points of rn N which
belong to components containing exactly one cycle . Then we have for N = N(n)
- en in case c * 1 12
k 0 -3
(5 .9) lien M(On,N(n)) =1 (2 ee-2c) k ~1 + + + .. . +
n -~ 2 g__g 1! 21 (k -3)!
while for c =1/2 we have
~1~
h
(5 .10) M 3 n 2,r3
( n,N(n)
12

where I(x) denotes the gamma-function 1'(x) _ f tx- 1 e - t dt for x > 0 .

0
 ERDŐS-RÉNYI
44

Proof of Theorem 5d . (5 .9) follows immediately from (3 .11) ; for c- 1 ;2

we have by (3 .14)

k'
~I 1 )
3
M ('9i ;,N(n)) e n~ .ti b2 3 .
12

Remark . Note that for c -~ 1/2

kk -3 1
1 ce -2c ) k j + 1k, + . . .
2 (J
(k - 3)11 4(1 - 2c) 2

Thus the average number of points belonging to components containing

exactly one cycle tends to + - as 1 for c -* 1 1`2 .
4(1 - 2 c) 2
We now prove
Theorem 5e . For í1(n) -v cn with 0 < c < '/2 all camponents of rn .lY(n)
are with probability tending to 1 for n either trees or components contain-
ing exactly one cycle .
Proof. Let ip n, ,N, denote the number of points of F, , v; belonging to com-
ponents which contain more edges than vertices and the' number of vertices
of which is less than V log n. We have clearly for X(n) cn with c < 1'2

2
[ log n ( kl r
-k - 1 - 0(nlog2 1~
M (Vn,N(n)) <_
1
k I
n 2 21 -~-~- 2 -
k=4 lk n

(12

Thus

P(Vn,N(n) > 1) - 0( 1 log2)

22

On the other hand by Theorem 4c the probability that a component con-

sisting of more than V log n points should not be a tree tends to 0 . Thus the
assertion of Theorem 5e follows .

§ 6. The number of components

Let us turn now to the investigation of the average number of compo-

nents of It will be seen that the above discussion contains a fairly com-
plete solution of this question. We prove the following
 ON THE EVOLUTION' OF RANDOM GRAPHS 45

Theorem 6 . If ~n N denotes the number of components of r,, N then we have

if N(n) - on holds with 0 < c < 1
2
(6 .1) M (sn,N(n))
= 9t - N(n) -}- 0(1)
n
where the bound of the 0-term depends only on c . If N(n) we have
2
(6 .2) M(n,N(n)) = n - N(n) + O(log n) .

I f N(n) -on holds with c > 1 we have

2

X2(C)
M(~rz,N(n)) - }
lim _ - I x(e)

n--+m 9t 2C 22

where x =x(c) is the only solution satisfying 0 < x < 1 of the equation xe - z =
= Zee -2c , i . e .
kk-1 (2 CE 2c)is
x(C) _
k=1 k!
Proof of Theorem 6 . Let us consider first the case C < 1 . Clearly if we
2
add a new edge to a graph, then either this edge connects two points belong-
ing to different components, in which case the number of components is
decreased by 1, or it connects two points belonging to the same component
in which case the number of components does not change but at least one
new cycle is created. Thus 2
(6 .5) Sn,N - (n - N) < Hn,N

where Hn,N is the total number of cycles in '.,N . Thus by Theorem 5a it

follows that (6 .1) holds .
Similarly (6 .2) follows also from Theorem 5a . Now we consider the case
1
C > - .
2
It is easy to see that for o < y < 1e we have (see e . g . [14])
I,-2 .y k x2
(6 .6) - =x - -
k=1 k! 2
where
+ b k k-1 y k
k=1 k!

2 In fact according to a well known theorem of the theory of graphs (see [4], p. 29)
being a generalization of Euler's theorem on polyhedra we have N - n + ~n,N =
= xn,N, where xn,N - the ,cyclomatic number" of the graph rn,N - is equal to
the maximal number of independent cycles, in I'n ,N (For a definition of independent
cycles see [4] p . 28) .
 ERDŐS-RÉNYI
46

x can be characterized also as the only solution satisfying 0 < x < 1 of the
equation xe_t = y . -
It follows that if 1V(i?) nc holds with c <'/2 we have
n2 2 N( n ) 4 2 (n) ~
(6 .8) MN(n)) _ - X - -{- 0(1) = n - 1' (n) -f- 0(1)
2 N(n) n 2n 2
which leads to a second proof of the first part of Theorem 6 .
To prove the second part, let us remark first that the number of compo-
nents of order greater than 4 is clearly < . Thus if n ,ti.(A) denotes the

number of components of order < 4 of Fn ,N We have clearly

(6 .9) M(' n,N) = M(~n,N( 4 )) -I- O

fns
A
The average number of components of fixed order k which contain
?~ I k
at least k edges will be clearly- according to Theorem 1 of order f . e.
III

bounded for each fixed value of lc . As 4 can be chosen arbitrarily large we

obtain from (6 .9) that
n
(6 .10) MGl?,n,) - o, M( Tk) -
k=1

According to (2 .18) it follows that

n2 - `n kk-2 2 N - 2n k

(6 .11) M( ;n,N) e
.2k 1 k! n

and thus, according to (6 .6) if N(n) - cn holds with c > 1,12 We have

1.
(6 .12) lim MGn,N(n1) = x(c) - xu(c)
n -- n 2 e 2

where x(c) is defined by (6 .4) . Thus Theorem 6 is completely proved .

Let us add some remarks . Theorem 6 illustrates also the fundamental
n
change in the structure of which takes place if IV passes . While the

average number of components of -r,,,N (as a function of N with n fixed)

decreases linearly if N <
n this is no longer true for
N > n ; the average
2 2
number of components decreases from this point onward more and more
slowly. The graph of

1-c for 0 _
<c<
_ -1-
2
M( 5 n,N(n))
(6 .13) w(c) = lint
N(n) n 1 x2(c) ' 1
„ fx(c) - for c > / 2
2C -2 1
 ON THE EVOLUTION OF RANDOM GRAPHS 47

as a function of c is shown by Fig . 4 .

From Theorem 6 one can deduce easily that in case N(n) - cn with
c < 1/2 we have for any sequence w n tending arbitrarily slowly to infinity

(6.14) "M P(I Sn,N(n) - n -{- N(n) I < Con) = 1

(6.14) follows easily by remarking that clearly `ten >_ n: - N.

I-Cfor0 :cs%2
z=zfc)=
(c)- x2cl) for c > %2
2c (x

Ő 1
-c
1
2
Figure 4 .

For the case N(n) - cn. with c >_ 1/2 one obtains by estimating the
variance of ~n,N(n) and using the inequality of Chebyshev that for any e > 0

Sn .N(n) _ x_ (e
(6.15) lim P I f x(C - < el - 1 .
n-, n 2c 2 ~ J

The proof is similar to that of (4 .21) and therefore we do not go into details .

§ 7 . The size of the greatest tree

If N - cn with c < 1 / 2 then as we have seen in § 6 all but a finite num-

ber of points of F, N belong to components which are trees . Thus in this case
the problem of determining the size of the largest component of F,, N reduces
to the easier question of determining the greatest tree in ",,N . This question
is answered by the following .
Theorem 7a . Let do ,~ denote the number of points of the greatest tree which
is a component of F. , N* Suppose N = N(n) - cn with c + 1/2 . Let w n be a sequence
48 ERDŐS-RÉ NYI

tending arbitrarily slowly to + ~ . Then we have

(7 .1) s lim P 4 n,N(n) > f log n - 2 loglog nI + (o , r = 0

and

(7 .2) s lim P >_

I log n - 2 loglog n ~ - wn
An ,Mn) = 1

where
(7 .3) e - a = 2 ee 1-2 c (i . e . a= 2 c- 1 -log 2 c
and thus a > 0 .)
Proof of Theorem 7a . We have clearly
(7 .4) P(d nMO >- z) = P f Zk > I < M(tk)
`k>x k?z
and thus by (2 .18)
ne ° x1
(7 .5)
POn,N(n) > z) - O - Z 5, 2

It follows that if z 1 =1 log 71 - 5 loglog nl + co n

a 2
we have
(7 .6) P(ln,N(n) z zi) = O(e °n)
This proves (7 .1) . To prove (7 .2) we have to estimate the mean and variance
of r,, where z2 =
I (log n - 2 loglog n~ - wn . We have by (2 .18)
1

M(TZ2) a
5 ~`2
eaw -
(7 .7)
2cV27r
and
(7 .8) D 2 (r,,) = O(M (Zzs » .

Clearly
P(4 n,N(n) Z2) > P('rz E > 1) = 1 - P(txg = 0)

and it follows from (7 .7) and ('7 .8) by the inequality of Chebyshev that

(7 .9) P( -r x_ = 0) = O(e °n) .

Thus we obtain
(7 .10) P(4 n,N(n) >- z2) I - O(e -aw n) .
Thus (7 .2) is also proved .
Remark. If c < 1 the greatest tree which is a component of rn,N with
2
N r cn is - as mentioned above - at the same time the greatest component
 ON THE EVOLUTION OF RANDOM GRAPHS 49

of I'n N , as Tn,N contains with probability tending to 1 besides trees only com-
ponents containing a single circle and being of moderate size . This follows
evidently from Theorem 4c . As will be seen in what follows (see § 9) for
c > 1 the situation is completely different, as in this case rn N contains
2
a very large component (in fact of size G(c)n with G(c) > 0) which is not a
tree . Note that if we put c = 1 log nn we have a = 1 log n and 1 log n k
2k k a
in conformity- with Theorem 2c .
We can prove also the following

Theorem 7b. If N- en, where c -/- 12 and e = 2ce 1-2 c then the number
5
of isolated trees of order h = flog n - loglog n, + l resp . of order >_ h (where
I is an arbitrary real -number such that h is a positive -integer) contained in
Fn ,N has for large it approximately a Poisson distribution with the mean value
_ a5!'2 e -al
a s 2 6-al
I = res 2
p. u
2c V2z 2cV2n(I -e - a)
1
Corollary . The probability that rn,N(n) with N(n) -nc where c
2

does not contain a tree of order >_ 1 (log n - 5 loglog n + l tends to

aI 2
a 5/2 e -al
exp - for n. +-, where a = 2c - 1 - log 2c .
2 c ~ / 2 .-,c (1 - e -°)
The size of the greatest tree which is a component of rn N is fairly large
if N - . This could be guessed from the fact that the constant factor in the
2
n

expression 1 log n - 5 loglog n of the .probable size" of the greatest compo-

nent of Fn ,N figuring in Theorem 7a becomes infinitely large if c = 1.

2
n
For the size of the greatest tree in F„, N with N - the following
2
result is valid :
Theorem 7c. If N - n and 4,,,N denotes again the number of points
2
of the greatest tree contained in I' n N, we have for any sequence co n tending to
~-~ for n +-
(7 .11) 11111 P(4 n N > 11213Wn)
. = 0
n-- T
and
f 2 rr3
(7 .12) lim P IdroN > n = 1 .
n--+ - ,n

4 A Matematikai Kutató Intézet Közleményei V. A11-2 .

 ERDŐS-RÉNYI
50

Proof of Theorem 7c . We have by some simple computation using (7)

-kl
~n kk_2 ( ín 22
ti N-k+ nk k-2 e-k
(7.13) M(z k) _ e 6n' .
k1
f21
Thus it follows that

(7 .14) P dn ( N >
= n2-3 w<
n) =
k~~mn
M (t k) - O í 1
V ~0 n
which proves (7 .11) .
On the other hand, considering the mean and variance of z* _ Tk'
n z"
k ~, -

it follows that
M(a*) z A (03'2 where A > 0 and Dz(t*) = 0( 0)3!2)
and (7 .12) follows by using again the inequality of Chebyshev . Thus Theorem
7c is proved .
The following theorem can be proved by developing further the above
argument and using Lemma 1.
Theorem Let r(y) denote the number of trees of order > yn2 3 contained
7d. 1

in rn,N(n) where 0 < y < +- and N(n) tin . Then we have

2
,h e -'
(7 .15) "M Pn,N(n)(T(P) = 9) = - (j = 0, 1, . . . )
where
+m

= 1 e - x dx .
(7 .16)
V12 n x3; 2
s

§ 8. When is F,, ,N a planar graph ?

We have seen that the threshold for a subgraph containing k points
k
and k d edges is n2 k+a ; thus if N en the probability of the presence
+

of a subgraph having k points and k + d edges in Fn,N tends to 0 for n - +-,

for each particular pair of numbers k > 4, d > 1 . This however does not
imply that the probability of the presence of a graph of arbitrary order having
more edges than vertices in rn N with N -nc tends also to 0 for n -{-~ .

In fact this is not true for >_ If2 as is shown by the following
c
 ON THE EVOLUTION OF RANDOM GRAPHS 51

Theorem 8a. Let Y. N (d) denote the number of cycles of Gn,N of arbitrary
order which are such that exactly d diagonals of the cycle belong also to pn,N .
n
Then if N(n) = + ~n + o( V-n) where -~ < A < +-, we have
2
e-Q
(8 .1) lim P(Yn,N(n)(d) = y) _ eJ (j = 0, 1, . . . )
n-+w j1
where
AY 0
2 d
y 2a-iel'3 •e
(8 .2) 1
y•
P= 2 .6a d!
o

Proof of Theorem 8a . We have clearly as the number of diagonals of a

k(k - 3)
k - gon is equal to
2

k(k-3) l2~ ~2,

1 ~
n 2 N k
-- d,
M(xn,N(d))=~ n l (k - 1)f
d n
t JI

l 21
N
Vn
and thus if N (n) = n+ R + o(Vn)
2
n f k
3 k'
2n
M(Xn,N(n)(d) }^' k2d-1 e
(8 .4)
2 a +a ld! nd k=4
I 1
+ Vn
It follows from (8 .4) that
W a~ ~
1 yea-1 e Y3 2 dy .
ra lim M(Xn,N(n)(d))= 2
6a d
0

The proof can be finished by the same method as used in proving Theorem 2a .
Remark. Note that Theorem 8a implies that if N(n) = n + con V
2
with wn +°O then the probability that F,,,N(n) contains cycles with any
prescribed number of diagonals tends to 1, while if N(n) =n - w n Vn
2
the same probability tends to 0 . This shows again the fundamental difference
in the structure of ',,N between the cases N < n2 and N > 2n . This differ-
ence can be expressed also in the form of the following

Theorem 8h . Let us suppose that N(n) -nc. If c < 1 the probability

2

4*
 ERDŐS-RÉNYI
52

that the graph rn,N(,) is planar is tending to 1 while for c > 1 this probability tends

to 0 .
Proof of Theorem 8b . As well known trees and connected graphs contain-
ing exactly one cycle are planar . Thus the first part of Theorem Sb follows
from Theorem 5e . On the other hand if a graph contains a cycle with 3 dia-
gonals such that if these diagonals connect the pairs of points (P ;, P i ) (i =
= 1, 2, 3) the cyclic order of these points in the cycle is such that each pair
(P;, P,) dissects the cycle into two paths which both contain two of the other
points then the graph is not planar . Now it is easy to see that among the
k(k - 3)
2 k.
triples of 3 diameters of a given cycle of order k there are at least 6
3 (l
triples which have the mentioned property and thus for large values of k
approximately one out of 15 choices of the 3 diagonals will have the mentioned
property . It follows that if N(n) _ n + CJ', I/rn with wn +-, the proba-
bility that I'n
N(n) is not planar tends to 1 for n -~ +- . This proves Theorem
n
8b. We can show that for N(n) = + 7. with any real A the probability
2
of rn N(n) not being planar has a positive lower limit, but we cannot calculate
is value . It may even be 1, though this seems unlikely .

§ 9. On the growth of the greatest component

We prove in this § (see Theorem 9b) that the size of the greatest com-
ponent of rn N(n) is for N(n) - en with c > I/2 with probability tending to 1
approximately G(c)n where
x(c)
(9 .1) G(c) = 1 -
2e
and x(c) is defined by (6 .4). (The curve y = G(c) is shown on Fig . 2b) .
Thus by Theorem 6 for N(n) en with c > 1/2 almost all points of
rn N(n) (i . e. all but o(n) points) belong either to some small component which
is a tree (of size at most 1/a (logn - 2 loglogn) + 0(1) where a = 2c -1 -log 2c
by Theorem 7a) or to the single "giant" component of the size -G(C)n .
Thus the situation can be summarized as follows : the largest component
1
of rn N(n) is of order loge for N(n) _ c < 1/ 2 , of order n2,'3 for (n) and
n n 2
INT
of order n for (n) N c > 1/2 . This double "jump" of the size of the largest
n
component when N(n) passes the value 1/ 2 is one of the most striking facts
n
concerning random graphs . We prove first the following
 0\ THE EVOLUTION OF RANDOM GRAPHS 53

Theorem 9a . Let,-r, , ,N (A) denote the set of those points of rn,N which belong
to components of size >A, and let ff . ,N (A) denote the number of elements of
the set n,N (A) . If N1 (n) (c - s) n where e > 0, c - s > 1/2 and N2 (n) en
then with probability tending to 1 for n--->- +- from the H,,N,(n)(A) points
belonging to .-Tn,Nl(n)(A) more than (1 - b) Hn,N,(n)(A) points will be contained
in the same component of ~n,N2(n) for any b with 0 < b < 1 provided that
A > 50
(9 .2)
E2 62

Proof of Theorem 9a . According to Theorem 2b the number of points

belonging to trees of order :< A is with probability tending to 1 for n -~ +
equal to
/ q kk-1
n - [2(c, - E)]k -1 eo(n) .
~.k=1
kl

On the other hand, the number of points of rn N,(n) belonging to components

of size _< A and containing exactly one cycle is according to Theorem 3c
o(n) for c-e > 112 (with probability tending to 1), while it is easy to see, that
the number of points of I'n,Nl(n) belonging to components of size < A and
containing more than one cycle is also bounded with probability tending to 1.)
Our last statement follows by using the inequality (4 .19) from the fact
that the average number of components of the mentioned type is, as a simple.I
calculation similar to those carried out in previous §§, shows, of order 0 1
n
Let En1 ) denote the event that
(9 .3) ~Hn,Nl(n)(A)-nf(A,c-E)I <znf(A,c-E)

where r > 0 is an arbitrary small positive number which will be chosen later
and

f(A, c) = 1 - c kk-1 (2 ce -2 C) > 0

and let E;,1 ) denote the contrary event . It follows from what has been said
that
(9.5) lim P(E(,1 )) = 0 .
n- ;

We consider only such r n N,(,) for which (9.3) holds .

Now clearly I'n,N,(n) is obtained from -P. N I(n) by adding N2(n)-Y,_(n) ~--ns
new edges at random to rn N,(n) . The probability that such a new edge should
f Hn,Nj(n)(A)
- AT2(n)
I1
connect two points belonging to n,Ni(n)(A), is at least 2
(n~
2
and thus by (9.3) is not less than (1 - 2z) f2 (A, c - e), if n is sufficiently
large and 'r sufficiently small .
54 ERDŐs -Rr:NYI

As these edges are chosen independently from each other, it follows

by the law of large numbers that denoting by v n the number of those of the
N2 (n) - N1 (n) new edges which connect two points of n.Nl(n) and by E(2)
the event that
(9.6) vn >_ E(1 - 3r) J 2 (A, c - E) n
and by En2) the contrary event, we have

(9.7) lim P(En2)) = 0 .

n--.+
We consider now only such rn N2(n) for which En2) takes place . -Now let us
consider the subgraph rn NE(n) of I'n,N2(n) formed by the points of the set
n,N (,,)(A) and only of those edges of r. N,(,) which connect two such points .
'We shall need now the following elementary
r
Lemma 2 . Let a l , a2 , . . . , ar be positive numbers, Z aj = 1 . If max a, :<
-
a
1<j<r
j=1
then there can be found a value k (1 < k s r -- 1) such that

1- a k 1 -{- a
2 _ <- j= ] a~ ` 2
(9.8) and
1-a n 1+a
-
< )' aj <_
2 j=k+1 2

Proof of Lemma 2 . Put Sj = a (j = 1, 2, . . . , r). Let jo denote the

least integer, for which Sj > 1/l 2. In case Sjo - 1/2 > 1 /2 - Sjo_ 1 choose
k = jo - 1, while in case Sao - 1 /2 S 1 / 2 - S1o _ 1 choose k = jo. In both
cases we have I Sk - 1/2 1 < amp < a which proves our Lemma .
2 2
Let the sizes of the components of Fn N2(n) be denoted by b1 , b2, . . . , br.
Let En3) denote the event
(9 .9) max bj > Hn,NI(n)(A) (1 - S)
and E(,,3) the contrary event . Applying our Lemma with a = 1 - b to the
bj
numbers a j = it follows that if the event E (3)
n takes place, the
H n,Nl(n)(A)
Set ` n N,(n)(A) can be split in two subsets An and rn containing Hn and
Hn points such that Hn } Hn = Hn,NI(n)(A) and

(9 .10) H,,Nl(n)(A)
2 S min (Hn, H"t ) < max (Hn, Hn) < H,,,N,(n)(A) II -
2,
further no point of . °' is connected with a point of rn in F*,N2(n)
It follows that if a point P of the set '-- n Nl(n) (A) belongs to n (resp .
ten} then all other points of the component of I' n,NI(n) to which P belongs are
 ON THE EVOLUTION OF RANDOM GRAPHS 55

also contained in rn (resp . rf) . As the number of components of size > A

Hn,N )(A) the
Of rn,Ni(n) is clearly < number of such divisions of the set
1
xR ~. ~>(A)
-°n,N,(n)(A) does not exceed 2A
If further En3 ) takes place then every one of the v„ new edges connect-
ing points of n.N,(n)(A) connects either two points of n or two points
of fin . The possible number of such choices of these edges is clearly
(2n l
2 + I22 rZ )

wn
As by (9 .10)
~ Hn
1,1<
2 2
.-
b
+ ( 11á+b
=--<1a
2 2

H, 4 l 21 2 2
1

it follows that
1 H, .v,1n7(A) á ~e(1-3r)f2(A,e-e)n
(9.12) P(E(n3 ) < 2A ( 1-
2
and thus by (9.3) and (9 .6)
~(1+-r) log 2 e(1 - 3 i) f(Á, e s) á
(9 .13) P(En3 » < exp 1nf(A, c - e)
A 2 1
Thus if
(9 .14) A sá(1 - 3 z) f(A, c - s) > (1 + z) log o
then
(9.15) 1im P(EW) = 0 .
nom+

As however in case c - e > i/2 we have I(A, c - e) > G (c - s) > 0

for any A, while in case c - s -. 1 /2
kk-1 kk-1 >_ 1
(9.15x) f 1 =I- = if A z A o
lA, 2 . k=1 k 1 Ck k= A+, k! ek 2 VA

the inequality (9 .13) will be satisfied provided that r < 1 and A > 50
10 E2á2
Thus Theorem 9a is proved .
Clearly the "giant" component of rn,N2(n) the existence of which (with
probability tending to 1) has been now proved, contains more than
(1 -T)(1 -á)nf (A,c-s)
56 ERDŐS-E (;N l I

points. By choosing e, t and 6 sufficiently small and A sufficiently large,

(1 - z) (1 - S) f(A, c - E) can be brought as near to G(c) as we want . Thus
we have incidentally proved also the following
Theorem 9b . Let en ,,,, denote the size of the greatest component of rn,N'
If V(n) - en where c > lf2 we have for any 7) > 0

(9 .16) lim P Qn,N(n) -G(c)

n-- ~,' n
x(c) k 1
where G(c) = 1 - and x(c) _ - (2c e-2c)k
is the solution satisfying
2c k-1 k .
0 < x(c) < 1 of the equation x(c) e - Y(c) = 2ce -2c

Remark . As G(c) --->- 1 for c -* + - it follows as a corollary- from Theorem

9b that the size of the largest component will exceed (1 -a)n if c is suffi-
ciently large where a > 0 is arbitrarily small . This of course could be proved
directly. As a matter of fact, if the greatest component of r,,,N(,,) with N(n) -vne
would not exceed (1 - a) n (we - denote this event by Bn (a, c)) one could by
Lemma 2 divide the set V of the nn points P l, . . ., P n in two subsets P resp.
V" consisting of n' resp n"- points so that no two points belonging to different
subsets are connected and

(9 .17)
an
2
< min (n', n') <
- max W, W) - < I1 - .~ n .

But the number of such divisions does not exceed 2 1 , and if the n points
are divided in this way, the number of ways N edges can be chosen so that
only points belonging to the same subset V' resp . V" are connected, is
n'~ + ~n"~)
2 2

(n.'` + (n"` < n 2 a

As I ~ _ 9 1 - --~ . it follows
~

l JI L 2
Cl N (n) N(n)a
(9 .18) P(B,(a, c)) < 2' ~1 - < 21 e 2
-I
2;

Thus if a e > log4, then

(9 .19) lain P(B,(a, c)) = 0

which implies that for c > log 4 and AT(n) cn we have

a

(9 .20) lim P(On,N(n) >_ (1 - 60n) = 1 .

 ON THE EVOLUTION OF RANDOM GRAPHS 57

'e have seen that for N(n) - en with c > 112 l the random graph rn,N(n)
consists with probability tending to 1, neglecting o(n) points, only of isolated
n kk -2
trees (there being approximately (2c e-2c)k trees of order k) and of
2c k!
a single giant component of size G(c)n .
Clearly the isolated trees melt one after another into the giant compo-
nent, the "danger" of being absorbed by the "giant" being greater for larger
components . As shown by Theorem 2e for N(n) 1 log n only isolated
2k n
N(n) - 1/2 n log
trees of order < k survive, while for n- ~ +- the whole
n
graph will with probability tending to 1 be connected .
An interesting question is : what is the "life-time" distribution of an
isolated tree of order k which is present for AT(n) -cn ? This question is
answered by the following
Theorem 9r, . The probability that an isolated tree of order k which is present
in. 1,,N,( .) where N,(n) -~- cn and c > 1/2 should still remain an isolated tree
in rn M(n) where í1'2 (n) (c + t) n (t > 0) is approximately e -2kt ; thus the
„Life-time" of a tree of order k has approximately an exponential distribution
with mean value nand is independent of the "age" of the tree .
2 k-
Proof. The probability that no point of the tree in question will be
connected with any other point is
(n-k` -
N,(n) I 9 + k,
2
e -2kt
1
l=N,(n)+I
n,
~~ -

This proves Theorem 9e .

§ 10. Remarks and some unsolved problems

We studied in detail the evolution of "n N only till N reaches the order
of magnitude n log n . (Only Theorem 1 embraces some problems concerning
the range N(n) - n, with 1 < a < 2 .) We want to deal with the structure
of rnN(n) for N(n) - cn° with a > 1 in greater detail in a fortcoming paper ;
here we make in this direction only a few remarks .
First it is easy to see that -P, ,(r) N(n) is really nothing else, than the
2
complementary graph of Fn,N(n) • Thus each of our results can be reformulated
to give a result on the probable structure of ",,N with N being not much
less than For instance, the structure of r n N will have a second abrupt
In2
~nl
change when ű' passes the value n; if N < -en with C>1/2
2 2 22
then the complementary graph of -F, ,N will contain a connected graph of order
f(c)n, while for c < i/2 this (missing) "giant" will disappear .
 ERDŐS-RÉNYI
58

To show a less obvious example of this principle of getting result

for N near to , let us consider the maximal number of pairwise independent
~22
points in rn,N • (The vertices P and Q of the graph F are called independent
if they are not connected by an edge) .
Evidently if a set of k points is independent in rn,N(n) then the same
points form a complete subgraph in the complementary graph rn,N(n) • As
however rn,N(n) has the same structure as F n (Z) _ N(n) it follows by Theorem

1, that there will be in rn,N(n) almost surely no k independent points if -

2
rn,

-- N(n) = o In2(1 k 11) iL e . if N(n) _ - o n2(1- i .) but there will be in

k 1
2
rn,
i
rn,N(n) almost surely k independent points if N(n) _ (~ - co,, n2 0 - k-i- where
con tends arbitrarily slowly to +- . An other interesting question is : what
can be said about the degrees of the vertices of rn,N • We prove in this direction
the following
Theorem 10. Let Dn,N(n)(Pk) denote the degree of the point P k in rn,N(n)
(i . e . the number of points of rn,N(n) which are connected with P k by an edge) . Put

Dn = min D,,,N(n) (P k ) and Dn = max Dn,N(n) (Pk)

I<k :s~n t<kSn

Suppose that
N(n)
(10 .1) lim = T o0
n + m n log n
Then we have for any e > o

(10 .2) lim e D n-1

n +~ f Dn

We have further for N(n) ~-ca for any k

p (2 c)J e-2c
(10 .3) lim P(Dn,N(n) ( k) _%) _ ~ (i - 0, 1, . . . )•
n--+ - ?

Proof. The probability that a given vertex P k shall be connected by

exactly r others in rn,N is
n-1
n- 1) 2 h~2N~r - 2N
e n
i r N-r, N n .
(~ n r!
2,
N
 OA THE EVOLUTION OF RANDOM GRAPHS 59

thus if N(n) -onn the degree of a given point has approximately a Poisson
distribution with mean value 2c . The number of points having the degree r
is thus in this case approximately
(2 C)r e -2c
n - (r = 0,1, . . )
r?
If N(n) _ (n log n) wn with co n -* +- then the probability that the degree of
2 N(n) (1 - 2 N(n)
a point will be outside the interval e) and (1 + E) is ap-

proximately
(2 o)„ •l ogn' ke - 2w„ log n -O ( 1
,k-21ogn .w,2,>e .2logn •w ,i k. neYwj

and thus this probability is o 1 , for any E > 0 .

[n 1
Thus the probability that the degrees of not all n points will be between
the limit (1 ± E) 2w n log n will be tending to 0 . Thus the assertion of Theorem
10 follows .
An interesting question is : what will be the chromatic number of rn,lY '?
(The chromatic number Ch(P) of a graph r is the least positive integer h such
that the vertices of the graph can be coloured by h colours so that no two
vertices which are connected by an edge should have the same colour .)
Clearly every tree can be coloured by 2 colours, and thus by Theorem
4a almost surely Ch (.Tn N) - 2 if N = o(n) . As however the chromatic
number of a graph having an equal number of vertices and edges is equal
to 2 or 3 according to whether the only cycle contained in such a graph is
of even or odd order, it follows from Theorem 5e that almost surely Ch (I'n p,) < 3
for N(n) - no with c < IJa •
For N(n) ti
2 we have almost surely Ch tvCn>) > 3 . (Pn

As a matter of fact, in the same way, as we proved Theorem 5b, one

can prove that r contains for N(n) - 2 almost surely a cycle of odd
n N(n)

order . It is an open problem how large Ch (Pn N(n)) is for N(a) -on withc> 1/2 ?
A further result on the chromatic number can be deduced from our
above remark on independent vertices . If a graph T has the chromatic number
h, then its points can be divided into h classes, so that no two points of the
n
same class are connected by an edge ; as the largest class has at least points,
h
it follows that if f is the maximal number of independent vertices of r we have
n k) I
f >_ . IVOw we have seen that for N(n) = n _+ 2 (1 almost surely
h 2 J
j .2(1
f < k; it follows that for N(n) = - o( k~l almost surely Ch (rn,lY(n)) >
t2
6 0 ERDŐS-RÉNYI

Other open problems are the following : for what order of magnitude
r
of N(n) has n ,~~ n1 with probability tending to 1 a Hamilton-line (i .e . a path
which passes through all vertices) resp . in case n is even a factor of degree 1
(i .e . a set of disjoint edges which contain all vertices) .
An other interesting question is : what is the threshold for the appear-
ance of a "topological complete graph of order k" i .e . of k points such that
any two of them can be connected by a path and these paths do not inter-
sect . For k > 4 we do not know the solution of this question . For k = 4
it follows from Theorem 8a that the threshold is . It is interesting to
n2
compare this with an (unpublished) result of G . DIRAC according to which
if N >_ 2n: - 2 then GTn,N contains certainly a topological complete graph
of order 4 .
We hope to return to the above mentioned unsolved questions in an other
paper .
Remark added on May 16, 1960 . It should be mentioned that N . V .
SMIRNOV (see e. g . Marne~uamugecmwi C6opuux 6(1939) p . 6) has proved a
lemma which is similar to our Lemma 1 .

(Received December 28, 1959 .)

REFERENCES

CaYLEY, A . : Collected Mathematical Papers . Cambridge, 1889-1897 .

PóLYa, G . : "Kombinatorische Anzahlbestimmungen für Gruppen, Graphen and
chemisebe Vcrbindungen" . Acta Mathematics 68 (H37) 145-254 .
KÖNIG, D.: Theorie der endlichen anti unendlichen Graphen. Leipzig, 1936.
BERGE, C . : Théorie des graphes et ses applications. Paris, Dunod, 1958 .
FORD, G . W .-UHLENBECK, G . E . : "Combinatorial problems in the theory of
graphs, I." Proc . Nat . Acad . Sci . 42 (1956) USA 122-1.28 .
GILBERT, E . N . : "Enumeration of labelled graphs" . Canadian Journal o f Math .
8 (1957) 405-411 .
ERDős, P .-RÉNYI, A . : "On random graphs, I" . Publicationes Mathematicae (Debre-
cen) 6 (1959) 290-297 .
HARaRY, F . : "Unsolved problems in the enumeration of graphs" In this issue, p . 63 .
ERDős, P . : „Graph theory and probability ." Canadian Journal of Math . 11 (1959)
34-38 .
AUSTIN, T . L .-FAGEN, R . E .-PENNEY, `V . F .-RIORDAN, J . : "The number of
components in random linear graphs" . Annals of Math . Statistics 30 (1959) 747- 754.
TURÁN P.: "Egy gráfelméleti szélsőérték feladatról". Matematikai és Fizikai Lapok
48 (1941) 436-452 .
TURÁN, P. : "On the theory of graphs", Colloquium Mathematicum 3 (1954) 19-30.
KŐVÁRY, T.-Sós, V. T.-TURÁN. P.: "On a problem of K. Zarenkiewicz". Col-
loquium Mathematicum 3 (1954) 50-57 .
RÉNYI, A . : "Some remarks on the theory of trees" . Publications of the Math . Inst .
of the Hung. Acad. of Sci. 4 (1959) 73-85.
RÉNYI, A . : "On connected graphs, I ." .Publications of the Math . Inst . of the Hung .
Acad . of Sci . 4 (1959) 385-387 .
JORDAN, CH . : Calculus of finite differences . Budapest, 1939 .
ERDŐS, P .-STONE, A . H . : "On the structure of linear graphs" . Bull . Amer . Math .
Soc. 52 (1946) 1087-1091.
 61

0 РАЗВЁРТы ВАНИЕ СЛУцАЙНы Х ГРАФОВ

P . ERDŐS и А . RÉNYI
Резюме

Пусть дaны n точки Р 1 , Р 2 , . . . , Рn, и выбираем случайно друг за другом

N из возможны х ~2 ребер (Р~, Р1 ) тaк что после того что вы брани k ребра

каждый из других - k ребер имеет одинаковую вероятность быть вы-

2
lп
бранным как следующий . Работа занимается вероятной структурой так
получаемого слуцайного графa Г„, N при условии, что N = N(п) известнaя
функция от n и n очень большое число . Особенно исслeдуется изьченение
этой структуры если N нарастает при данном очень большом n . Случайно
развёртывающий граф может быть pассмотрен как упрощенный модель pостa
реальны x сетей (нaпример сетей связи) .