ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS

By
P. ERDOS (Budapest), corresponding member of the Academy,
and T. GALLA[~Budapest)

Introduction

In 1940 TUR]~N raised the following question: if the number of nodes,

n, of a graph s is prescribed and if l is an integer ~ n, what is the number
of edges which the graph has to contain in order to ensure that it necessa-
rily contains a complete /-graph? TURAN gave a precise answer to this ques-
tion by determining the smallest number depending on n and l, with the
property that a graph with n nodes and with more edges than this number
necessarily contains a complete /-graph ([9], [10]). More generally, the
question can be posed, as was done by TURAN: given a graph with a
prescribed number of nodes, what is the minimum number of edges which
ensures that the graph necessarily contains a "sufficiently large" subgraph
of a certain prescribed type? An alternative formulation of this question is as
follows: the number of nodes being fixed, we seek the maximum value of ,u,
being such that there exists a graph with ~ edges which does not contain
a subgraph of the type in question with more than a certain given number
of nodes. In our paper we are concerned with this problem for the case in
which the types of graphs considered are paths, circuits and independent
edges. (These terms are defined in w 1.)
Our results are not exhaustive, because, in general, we only give an
estimate of the extremal values, only in isolated cases - - for certain special
values of the number of the n o d e s - do we succeed in determining the
extreme values and the "extreme" graphs completely. Here are some of our
results capable of simple formulation:
Every graph with n nodes and more than (n--1)l/2 edges (l ~ 2) con-
tains a circuit with more than l edges. The value (n--1)l/2 is exact if and
only if n ~ q ( l - - 1 ) ~ l , then there exists a graph having n nodes and

The graphs considered in this paper are all finite, every edge has two distinct end-
nodes, and any two nodes are joined by at most one edge.
1. c. letters always denote non-negative integers, n always denotes an integer ~ 1.
A complete l-graph is a graph with l nodes, every pair of distinct nodes joined by an edge.
A graph is said to contain its subgraphs. (See w 1 of this paper and [6], pp. 1--3.)
338 P. ERDOS AND T. GALLAI

(n 1)l/2 edges which contains no circuit with more than l edges. (Theorem
(2. 7).)
For all n > ( k + 1)3/2, k ~ 1, every graph with n nodes and more than
n k k ( g + 1)/2 edges contains a path or a circuit with more than 2k edges.
The value n k - - k ( k + 1)/2 is exact. (Theorem (3. 6).)
For our proofs we need a group of theorems different from the above,
which are of interest in their own right. In these it is not the number of
edges, but the fact that every node has a sufficiently high degree (the degree
of a node is the number of edges incident with the node), which implies
the existence of a "sufficiently large" subgraph of a prescribed type. This
class of-problems was first considered by ZARANKIEWICZ[11] and Dll~AC [3].
From among these older results we require two theorems due to DII~AC ([3],
Theorems 3 and 4), for which we give new simple proofs in w 1. (A simple
proof of Theorem 3 can be found in [8].) We also prove some new theorems
of the type just now discussed.
In w 1 we present the necessary preliminary notions and some lemmas,
and we prove the theorems pertaining to the ZARANKIEWICZ--DIRAC field of
problems. In w 2 we carry out the estimations connected with problems of
the TURAN type. In w 3 we determine two extremal values exactly for a
sufficiently large number of nodes. In w 4 we determine the maximum num-
ber of edges in a graph of n nodes and at most k independent edges. We dis-
tinguish our more important results from the less interesting assertions leading
to them by the appellation "Theorem".

w
(1.1) Let M - - { P 1 . . . . . P,,} be a finite non-empty set and let the set
of all unordered pairs of distinct elements PIt~.--P~P~ (i:#j) of M be
denoted by N. (If n - - 1 , then N is the empty set.) The elements of M are
called nodes, the elements of N are called edges, and the edge PiPj is said
to be incident with the nodes Pi and Pj. Let N1 be an arbitrary subset of
N, M and N~ are said to define a graph F - - ( M , N1). The elements of M
and N1, respectively, are the nodes and edges of F. If PgPj E N1, then P.~
and P~ are said to be joined (in F), or we say that the edge PiP~ exists (in F).
The graph F ~- (M, N) is called complete, more exactly a complete
n-graph. The graph I ' - - ( M , N N~) is the complement of the graph
F - - (M, N0.
Let N' denote the set of pairs of elements of the finite set M' and let
N I ~ N ' . The graph F ' - - ( M ' , N I ) is called a sabgraph of the graph
F - ~ (M, N1) if M' c M and N; ~ N~. We also say that F contains F' and
 ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS 339

that F ' i s in F. If F ' is a subgraph of F, then the graph [F'] ~-(M', N ' n N1)
is called the subgraph (of F) spanned by F' or by M' in F. If P is a node
of the graph F, then F - - P denotes the graph obtained by deleting P and
all edges incident with P from I \
9 (1.2) If a (1--1) correspondence can be established between the nodes
of the graphs /'1 and F2 so that nodes joined in one graph correspond to
nodes which are joined in the other and, conversely, then the two 'graphs
are regarded as identical, and this is expressed in symbols by I ~ F , _ , .
The number of nodes and edges in the graph F is denoted by :~(F)
and r ( f ' ) , respectively.
The number of edges incident with the node P in the graph F is called
the degree of P in F. If there is no room for misunderstanding, then we
speak of the degree of P for short, and denote it by 0(P). If Q(P)-~-0, then
we call P an isolated node of F, if 0 ( P ) ~ 1, we call P a terminal node of IL
A graph L is called a loop, more accurately a P-loop, if a series 2
P1 .... ,P,,,P,~+I (n>= 1) can be constructed from the nodes of L so that every
node of L appears in the series, P ~ P~, the nodes P1,...,P,~ are all dis-
tinct, P~-~+I:#P~, and if n > l , then P,,+~:/=P~-I, and the set of edges of L
consists of P~P~+I ( i = 1,..., n). It is easy to see that we can form in at
most two ways from the nodes of a R-loop a sequence of the required pro-
perties and that P,+~ is uniquely determined. P is the initial node of the
loop and P,~+I the final node. The loop is also said to start from P and to
lead to P,,+~. We call the P-loop directed if one of the above-mentioned
sequences is made to correspond to it.
If P,,+~ is different from P~ . . . . , P~, then the loop is called a path,
more accurately a PiP**+~-path, i f P,~+~PI, then it is called a circuit.
Paths and circuits will be designated by the common term arc. If P , , + ~ P i
(1 _ - < j ~ n - - 2 ) , then the nodes P~,Pj+~. . . . . P,+I and the edges P~P~+I
( i - ~ j . . . . , n) together form a circuit which is called the circuit of the loop L.
The number of edges of L is called the length of L. Paths, circuits and arcs
of length l are called l-paths, l-gons and l-arcs, respectively. The equation
L=(P~ ..... P,~,P,~+I~-Pj) (1 ~ j ~ n - - 2 )
states that the graph L is a P v l o o p wich is composed of the nodes P~ . . . . , P,~
and the edges P~P~+~( i : 1, ..., n). The equation
W ~ ( P l . . . . , P,~+I)

2 If j and g are natural numbers and j < g or j > K , then P;, . . . , P v denotes the set
of nodes P~ where i runs through the natural numbers from j to g. If j - - g , then P ; , . . . , P:~
means Pj by itself.
340 P. ERDOS AND T. GALLAI

states that W is a P1Pn+l-path composed of the nodes P l , . . , ~ + 1 and the

edges PIP,+1 (i = 1. . . . . n).
A set of edges of F is called independent if no two of them have a
node in common. We shall say that the maximum number of independent
edges is k, if there exists a set of k such edges and there does not exist
a set of k + 1 such edges.
(1.3) The graph F is connected if it consists of a single node or if
corresponding to any two distinct nodes P and Q F contains a PQ,path.
The "maximal" connected subgraphs of F are called its components. The
subgraph F' of the graph F is maximal with respect to some property if F
contains no subgraph with this property of which F' is a proper subgraph.
If F' is a component of I', then F - - F ' denotes the graph obtained from F
by deleting F'.
The nodes P1 . . . . , Pj (j--> 1) are said to separate the two (distinct)
nodes A and B in the connected graph F if P i 4 : A , P i - ~ B ( i = 1 . . . . , j )
and every AB-path in F contains at least one of the nodes Pl . . . . ,P~.
The nodes P 1 , . . . , Pi divide the connected graph F if I" contains two nodes
which they separate.
The graph F is n-foM connected (n ~ 2) if it is connected and if no
set of fewer than n nodes divides it. A complete /-graph is said to be n'fold
connected for all n.
The maximal twofold connected subgraphs of the connected graph F
are called the members of F. Every edge of F is an edge of some member
and every member, except for the graph consisting of a single node, contains
more than one node. If F is connected but not twofold connected, then it
has more than one member, and it may be verified ([6], pp. 224--231) that
two of its members have at most one node in common and that such a node
divides F. 3 Furthermore, it may easily be verified that F has at least two
members containing only one cut-node each. Such members are called ter-
minal members of F. If F' is a terminal member of F, then F - - F ' denotes
the graph obtained from F by deleting all edges of F' and all nodes of
F' except its cut-node.
(1.4) I f in the graph F every node except the single node A has
degree >--k (k >--2) and :r(I') ~ 2k, then, if 0 (A) ~ 2, F is twofold connected
and if e(A) = 1, lhen F - - A is twofoM connected and F is connected.
PROOF. Suppose first that q ( A ) ~ 2 . Then it follows from our assump-
tions that every component of F and every terminal member of F (if any)
contains at least 3 nodes, and at least two of these have degree >=k. If F
.~ Such a node is called a cut-node of the graph.
 ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS 341

is not twofold connected, then it h a s two Components or two terminal mem-

bers. At least one of these components or terminal members has not more
than k nodes. The nodes in it with at most one exception have degree <k,
which is a contradiction.
if • ( A ) - - 1 , then let the edge incident with A be denoted by A A'.
If k ~ 3 , then the degree of A' in F - - A is clearly at least 2. If k = 2 , then
only ~ ( F ) = 4 is possible, in which case the degree of A' in F - - A is 2.
In both cases the theorem with Q(A)_>--2 can be applied to F - - A . F is
obviously connected.
(1.5) A circuit of the graph I" which contains all nodes of F is called
a Hamiltonian line of F, H-line for short. A path of I" which contains all
nodes of F is called an open H-line of I'. Two distinct nodes P and Q of
F are said to be H-independent in F if F contains no open H-line starting
:in P and ending in Q.
Our later reasoning is based on the following lemma:
LEMMA (1. 6) I f the circuit C - = (P~,..., P,,, P,~+I P,) is an H'line o f
lhe graph I" a n d i f P~ and Pj (i,j--[=n ~ 1) are H-independent nodes of F, then

e (p;+i) + e (&+i) = n;
~(P) denotes the degree of the node P in F.
PROOF. It may be assumed that i ~ 1 and 1 < j ~ n . Because two
neighbouring nodes of C are not H-independent, 3 ~ j ~ n - - 1 . Accordingly,
n ~ 4 and the nodes PL, P,, Py, P~,, are distinct.
If F contains the edge P"P:s ( 3 < g ~ j ) , then it does not contain the
edge P:;~IP~,-~. For if this edge belonged to I . then the path
W = (P~, P.,,. . . . , Py+~, P,a, . . . . , P.,, P , , , . . . , Pj)
would be an open H-line of F starting in Pl and ending in Pa.
It follows that P~Pj+, does not exist in F, since P2Pa does.
If the edge P..,Pt ( j § exists in 1, then the edge P~§162
cannot exist in F. For if this edge existed, then the path
W' ~---(P~, . . ., P,, P,, . . ., Py.,, P,+, . . . . . P,,+,)
would be an open H-line of F starting in P~ and ending in Pn:~l= Pa.
Accordingly, with every node, other than P1, joined to P2 in F there
can be associated a node not joined to Py+~ in such a way that these asso-
ciated nodes are all distinct. It follows that the number of nodes not joined
to P~-+~ is at least o(P~)--I, so that ~,(Pj+i) ~ ( n - - 1 ) - - ( e ( P , 2 ) - - l ) ~ n--e(P,,).
342 P, ERD()S AND T. OALLAI

NOTE. It follows from Lemma (1.6) and its proof that, under the hypo-
theses of the lemma, 9(Pi_l)+e(Pj_l)<n and the edges P~+~Pj+~ and P~_.~Pj_~
do not exist in F (Po - - P ~ ) .
Let the loop L - - (Pi . . . . . Pn, P,~+i =- P~) (1 < m < n 2) be a subgraph
of the graph F. A node Pj belonging to the circuit C - - (P ........ P., P,~+I--P,.)
of the loop L and different from the terminal node P,~ of L is called H-node
of the loop L (with respect to F ) if F contains a ~ P . , - p a t h containing all
the nodes of C and no other nodes, or, otherwise expressed, if the graph [C]
spanned by C in F contains an open H-line starting in P; and ending in P.,.
(1.7) Let L - - (P1 . . . . . P~, P,~+I-- P,,~) (1 < m < n 2) be a loop o f
the graph F and let the circuit of L be denoted by C. I f every H-node
of L (with respect to F) has degree > k ( k ~ 2 ) in [C] and if .w( C) <_2k 1,
then every node of C different from Pm is an H-node of L.
PROOF. P,,+I and P,, are clearly H-nodes. Our theorem is established
if we prove the following assertion: [f Pj+I (m < j < n) is an /-/-node, then so
is Ps. To see that this is so, let it be assumed that Pj+l (m < j < n) is an
H-node and Ps is not. Then P,,~ and P~. are H-independent in [C], and, since
C is an H-line of the graph [C], it follows from Lemma (1.6) that .o'(P,,~+i)+
+ O (, sp+ ~ ) < ~ r ( C ) , where o'(Pi) den6tes the degree of the node P~ in [C].
But it was assumed that ,o'(Pj+x)>--k and O'(Pm+~)>k. We have a contra-
diction l
Let A denote a node of degree > 1 of the graph F and let the degree
of all nodes of F o t h e r than A be > 2 . If W - - ( P l .... ,P,~) ( P ~ - - A ) is
any path of F which starts from A, then, the degree of P,~ being > 2 , F
contains an edge P,,P,+~ incident with P,, and different from P,,P,_~. This
edge and W together form an A-loop which is longer than IV. Thus to
every path W starting in A there exists an A-loop longer than W, hence the
longest A-loops of F are not path (i. e. they contain circuits).
These longest A-loops of F which possess the longest circuits will be
called maximal A-loops.
(1.8) Let A be a node of the graph F with degree > 1 and suppos e
that the degree of every node of F other than A is > k where k ->-2.
Further, let L be a maximal A-loop and let the terminal node of L be denot-
ed by B and its circuit by C. Then if ~ r ( C ) < 2 k 1, it follows that [C] is
either a component or a terminal member of F and in the latter case the cut-
node of [C] is B. In both cases every node of [C] distinct from B is connect-
ed to B by an open H-line of [C] and ~ ( C ) > k + l .
PROOF. It f011ows from the assumptions made that F contains a maxi-
mal A-loop. Let L - - (P~, ..., P~, P~+a - - P,~) (1 ~ m ~ n 2, PI - - A,/~,,~ - - B)
 ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS 343

be the maximal A-loop concerned. If Pj ( m < j ~ n ) is any H-node of L,

then Pj is not joined by an edge in I" to any node which is not in C. For
if PiP is an edge of F and if W denotes an open H-line of [C] leading
from Pj to Pro, then if we add the edge P~.P and the path (P1 .... , P,,) to
W the result is an A-loop which is as long as L and which is a path if
P--/=Pr (i ~ l, ..., n) and which has a 10nger circuit than C if P = P g
(1 <=g<m). A contradiction with the maximal nature of L is therefore avoided
only if P = P , j ( m ~ g ~ n ) . It follows that the degree of every H-node of
L in [C] is ~ k , therefore if : ~ ( C ) ~ 2 k - - 1 , then according to (1.7) every
node of C other than B is an H-node of L. There follows firstly the exist-
ence of the open H-lines asserted in the theorem and secondly that every
node of C other than B is joined exclusively to nodes of C. The latter fact
implies that [C] is a component or a terminal member with cut-node B.
([C] is obviously twofold connected because of C.) ~ ( C ) >=k -[-1, because
Q ( P , ) ~ k and P,, is joined exclusively to nodes of C.
If the graph F of Theorem (1.8) has at most 2k--1 nodes, then it
follows from Theorem (1.4) that [C] = F or [C] = F - - A according as ~ ( A ) ~ 2
or q ( A ) ~ l and that, if : w ( F ) = 2 k and q(A)>=2, then y ~ ( C ) = 2 k .
The following two theorems can be deduced:
THnORF.M (1.9) If the node A of the graph F is not isolated and the
degree of every node of F distinct from A is ~ k (k >=2) and if n~(F) <=2 k - - 1,
then A is connected by open H-lines to every node of F.
THEOREM (1. 10) (DIRAC) If every node of the graph F has degree >=k
(k >= 2), and if ~(1")<~ 2k, then F has an H-line.
(1.9) obviously implies the following theorem:
THEOREiVl (1. 1 1) If every node of the graph F has degree >=k (k ~ 2 ) ,
and if : ~ ( F ) ~ 2 k - - 1 , then any two distinct nodes are connected by an open
H-line.
If the graph F of Theorem (1.8) is twofold connected and if ~(F)-->_
~ 2 k , then :r(C)>--2k. From this it follows that the node A of Theorem
(1.8) is the initial node of a path of length ~ 2 k ~ l .
The following two theorems can be deduced:
THEOREM (1. 12) If F is a twofold connected graph and every node
with the exception of one single node A has degree >=k (k > -- 1) and if in
addition ~r(F) >=2k, then F contains a path with at least 2 k - - 1 edges which
starts from A.
(This theorem is trivial for k = 1.)
344 P. ERDOS AND T. GALLAI

THEOREM (1. 13) (DIRAC) If the degree of every node of the twofold
connected graph F is > k ( k > 2 ) and if n:(F)=> 2k, then F contains a
circuit with at least 2k edges.
REMARK. It follows from the above considerations that the assertions in
Theorems (1.10) and (1.13) remain true if the degree of every node except
one node A is at least k ( k > 2 ) and o ( A ) > 2 . If two nodes have degree
<k, then these theorems are not generally true.
THEOREM (1. 14) If F is a connected graph and the degree of every
one of its nodes is > k (k > 1) and if ~ ( F ) > 2 k - k 1, then F contains a path
with 21: or more edges?
PROOF. If k ~ 1, the theorem is trivial. In what follows it will be assumed
that k > 2.
If F is twofold connected, then, by (1. 13), F contains a circuit
C ~ (t:)1. . . . . . t~,y~, P , , . I ~ P 1 ) (m > 2 k ) . If m > 2 k , W - - ( P 1 . . . . . P,,) is a path
of the required kind. If m - - 2 k , then, by our assumptions, I" contains a
node P which is not in C and which is joined to a node of C, say P1.
Then W - - ( P , P1 . . . . . P.~) is a path of the required kind.
If F is not twofold connected, then let F~ and I1_~ denote two terminal
-members of F, and A1 and A,2 their cut-nodes. 1"1 and F2 are twofold con-
nected and apart from A1 and A,) their nodes have degree > k in F1 and F,),
respectively. This is possible only if n : ( F ~ ) > k @ 1 and ~ ( I ~ ) > k + 1. From
this and from (1.9) and (1.12) it follows that F1 contains a path of length
> k which starts in A1 and F~ contains a path of length > k which starts
in A~. If A i = A~, then these two paths together constitute a path with at
least 2k edges, and if AI--/:A.~, then these two paths together with an
A1Ae-path of F /'1 F,2 constitute a path with more than 2k edges.
REMARK. Theorem (1.14) can be proved easily whithout using the preced-
ing theorems.
(1.15) The "accuracy" of Theorems (1.13) and (1. 14) is demonstrated
by the following graph F :
F consists of the nodes P~ . . . . , P ~ , Q i , . . . , Q , ~ - k ( 2 < k < n k) and
of all edges PiQj ( i - - 1 , . . . , k ; j 1. . . . , n k). F is an even graph ([6],
p. 170). The degree of every node of F is > k and it may easily be veri-
fied that F is k-fold connected, further that the longest circuits and paths
in F have 2k edges.
It is seen from the example of this graph that in the theorems in ques-

4 This result was obtained independently by O.A. DmAc.

 ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS 345

tion it is not possible to assert the existence of longer paths and circuits
than those proved to exist, even if the connectivity is assumed to be higher.
(1.16) Using an altered form of Theorem (!. 8) and MF~NGEI,~'S well-
known "n-path theorem" the following theorems can be established:
If the degree of every node of the twofold connected graph I" is ~ k
( k ~ 2) and if : z ( F ) ~ 2k, then through each node of F there passes a circuit
having at least 2k edges.
If the degree of every node of the twofold connected graph I" with the
exception of the two nodes A and B is >~k (k ~ 2), then every node of F
lies on an A B-path having at least k edges.
These theorems are not proved in this paper.

w
(2. 1) Let the classes of all graphs containing exactly n nodes and
containing, respectively, no paths, circuits, arcs with more than l edges ( l ~ 1 )
be denoted by F(n, l), G(n,l), H(n, l). The graphs in each class which
contain the most edges are called the extreme graphs of the class concerned,
and the number of edges in these graphs will be denoted by f(n, l), g(n, l)
and h(n, l), respectively. So if the graph l" is a member of, respectively,
F(n, l), G(n, l), H(n, l), then the following inequalities h o l d "
(,) r(l') ~ f ( n , 1), r ( l ' ) ~ g ( n , l), .r(I') ~ h(n, 1),
and if 1' is an extreme graph of the class concerned, then equality holds
under (,).
We wish to estimate or determine f, g and h and to find the extreme
graphs.
Clearly, if n ~ l ( l ~ 2), the only extreme graphs of F(n, 1--1), G(n, l),
H(n,l) are the complete n-graphs.
(2.2) Our method of estimating the values in question from below is
to construct graphs belonging to the classes F, G, H and containing as
many edges as possible.
In this paper F~1' (I ~ k < n ) denotes the graph which consists of the
nodes P~,..., Pk, Qt, ..., Q,~-I~ together with all the edges PiPj ( i , j = 1,..., k;
i--.-.-~j) and all the edges P~Q~ (i=: 1, . . . , k ; j = - 1 , . . . , n--k).
If l<~k < n - - l , then r27~+~
.~ denotes the graph obtained from the graph
F~k by the addition of the edge Q~Q2. Accordingly, the graph F;z~ is defined
only for 1 2 2 and n > [ ( l + 1 ) / 2 ] . In what follows the graphs Fi~ will be
called stars and the graphs F,3~ will be called A-stars for all values of n.
7 Acta Mathematica X/3--4
340 P. ERDOS AND T. OALLAI

With the notation

2k+ 1)= 20+ 1

we have J~(/2,,):~(n,l) for l ~ 2 and n>[(l-[-1)/2].
It may easily be verified that the graph F;z~ has the following proper-
ties: it is [l/2]-fold connected, it contains no path and no circuit with more
than 1 edges, if n = > l ~ 3 , then it contains an l-gon, and if n > l , then it
contains an/-path. Hence F;7~ is a member of each of the classes F(n, 1),
G(n, I) and H(n, l), and so - - having regard to the remarks concerning the
case n E l in (2. 1) - -

(2.3) f(n, l)>=~(n, l), -qo(n,

g(n, l ) > l), h(n, l)>=~(n, l) (1>=1).
(2.4) The graph F,,,j ( j ~ l ) is defined as follows: Let n-~qjq-r
where r<j. F,,,i has exactly n nodes, and if r = O , it consists of q compo-
nents, while if r > 0 , it consists of q-[-I components, if r = 0 , all its com-
ponents are complete j-graphs, and if r>O, then q of its components are
complete j-graphs and the remaining component is a complete r-graph. ( I"~,~
consists of a single node.)
G* (n, j ) (j >=1) denotes the following class of graphs : Let n : q ( j - - l ) q- r
where 1--<r <'_ =j--l. G*(n,j) is the class of connected graphs containing
exactly n nodes which have q members if n > 1 and r - : 1, and q-}-1 mem-
bers if n > l and r>l, every member is a complete j-graph if n > 1 and
r = 1, and q members are complete j-graphs and one member is a complete
r-graph if n > 1 and r > 1. G*(1,j) {El, l} for all j ~ 1.
-~-

F*.,,,~ is to denote that element of G*(n, j) which contains at most one

cut-node.
In the notation
ga(n,j,r) -1~ n j - - - ~1r ( j q - l - - r )
it is found that
v(F~,5+l)-~p(n,j,r ) where n:q(jq-1)q-r (r<j-[-l)
and that

if FE G*(n,j), then v(F):ga(n,j, r) where n ~-q(j--1)-kr (1 <=r<=j--1).

The following statements may easily be verified:

F,,,,16F(n,I), G*(n,l)cG(n,l), 17,,,zEH(n,I ) (1~1).

 ON MAXIMAL PATHS A N D CIR C U I T S O F G R A P H S 347

Hence
i /(n,l)~v(n,l,r) (n=q(l+l)+r;r<l+l),
(2.5) g(n, l) >_ap(n, l, r) (n:q(I--1)+r; 1 =<r~/--l), (l>= 1).
h(n, l) :> V~(n, l--1, r) (n ql+r; r<l)
The statements below follow from a simple calculation:
(1) If l-----2k (k--
>_ 1), then ~(n, 2 k ) ~ 0 ( n , 2k, r) and equality holds only
if r - - k or r-~--k+ 1. So (2. 5) gives a better estimate of f and g" unless
r-~-~k or r~---k+ 1, in which cases (2. 3) and (2.5) are equally good.
(2) If l : 2 k (k>--l) and n > k ( k + l ) , then ~(n, 2k)>~(n, 2k--l,r).
Here (2.3) gives a better estimate of h.
(3) If l : 2k + 1 (k => 1) and n > k + 3, then ~ (n, 2k + 1)<~v(n, 2 k + 1, r).
For this case (2.5) gives a better estimate of f and g.
(4) If l = 2 k + l , then, according as (a) r = k or r = k + l , or
(b) r = k - - 1 or r : k + 2 , or (c) r < k - - I or r > k + 2 , ~o(n,2 k + l ) > , : ,
or < ~v(n, 2k, r) and, accordingly, (2. 3) gives a better or equally good or
worse estimate of h than (2. 5).
In order to estimate f,g and h from above we need the theorems of w 1.
THEOREM (2. 6)
:- 1nl
f(n, l)<

equality holds only if n ~ q ( l + l ) , in which case F,~,z+l is the only extreme

graph of the class F(n, l)
PROOF. If n=-< l + l , then a graph F having exactly n nodes cannot
contain a path with more than l edges, so
v(F)< n ( n - - l ) < nl
= 2 =2
and equality holds here only if F is a complete ( l + 1)-graph. So the theorem
is true if n < = l + 1.
Now let n ' > l + l and suppose that the theorem is true for all n such
that n<n'. We prove that in that case the theorem is also true for n'.
Let F ~ F(n', !).
(1) If F is not connected, then let its components be Fx .... ,1~ (p-->_2),
yr(F~) : nl (i~- 1,..., p). Then n l + . . . + n p ~---n', n~< n' and F~ E F(ni, l)
(i---~ 1 , . . . , p ) . So by hypothesis
1 ,l
v ( F ) : v(F1)+... + v ( F p ) ~ nl I + . . . + n ~ = n -2,

7*
348 r,. ERDOS AND T. OALLAI

l
and equality holds only if e(l;.)=-n~-~ ( i ~ 1. . . . ,p), i.e. if El, . . . , F ~ are
~11 complete ( l § 1)-graphs. In this case our theorem is therefore true for n'.
(2) Suppose that I" is connected. We show that there exists a node P'
whose degree in F, c.)(P'), is at most l/2. For if no such node P' exists,
then, whether l - - 2 k or l - = 2 k + 1, q ( P ) > / : @ l for every node P. If l - - 2 k ,
then ~ r ( F ) > 2 k - F 1 , and if r r ( F ) > 2 k + 2 , then, by (1.14), F contains a
path having at least 2t"-t-2 edges, while if ~ ( / ' ) 2 k + 2 , then, by Theorem
(1.10), F contains a path having 2 k + 1 edges. If l-= 2 k + 1, then .-.z(F)>
>2k+2 and so, by (1. I4), [' contains a path having 2 k + 2 edges. In
,every case we have a contradiction.
Suppose therefore that 9 (P') < 1/2 and let F ' - - I ~ P ' . Then F ' ~ F ( n ' - - 1, l).
F' cannot contain a complete ( l + 1)-graph because if it did, then I ' would
contain an (l-t- 1)-path. So by our induction hypothesis ,r(F')<(n' 1)l/2,
and therefore

v ( F ) == o(P') -t- ~(F') < ~l + (n' 1) /__ = n , -2-"

l

THEOREM (2.7)
g(n, l) ~ 2 ( n - - l ) / (l ~ 2),

equality holds only if n - - q ( l 1 ) + I, in which case the extreme graphs of

the class G(n, l) are the elements of the class G*(n, l)
PROOF. The theorem is trivially true for n - - 1 . If l < n < l and
F ~ O(n, l), then
n(n 1) (n--l)l
,,(F) < 2 < 2

.and equality here can hold only if F is a complete l-graph. The theorem is
therefore true for u < l.
Suppose that n ' > l and suppose that the theorem is true for all n if
n<n'. We show that it is then true also for n'. Let F ( G ( n ' , l ) .
(1) If F is not connected, let its components be /'1 . . . . . 1~, ( p > 2 ) and
let ~r(F~) - - n~. Then nl _t_ ... -4- n~:.-= n', n~< n' and F~ E G(n,, l) (i == 1, ..., p).
By our hypothesis therefore
~'(I')--~,(FO+... +r(Fp)<(nl 1)I/24...~-
-~-(n,~ 1 ) l / 2 - - ( n ' - - p ) l / 2 < ( n ' 1)I/2.
(2) If F is connected but not twofold connected, then let F~ denote a
~erminal member of F and let F e - - F - - I ~ . Jz(I~)-~--nl, :z(lb)--=n2. Then
 ON MAXIMAL PATHS AND CIRCUITS OF GRAPHS 349

n ' ~ nj-~ n2--1, nl < n', n2 < n', FI ~ G(&, 1), 1~ ~ O(n~, l), and so, according
to our hypothesis,
r (F) ~ r (1 ~1)q- u (F=,)~ (n~ -- 1) l/2 -[- (n2 -- 1) l/2 ~ (n'-- 1) I/2,
and equality can hold only if F~ ~ G*(n~,l) and F,,~ G*(n2,/). But then
1" ~ G*(n' l).
(3) Let F be twofold connected. We show that F then has a node P~
of degree~=l/2. For if no such node exists, then Q ( P ) ~ k + 1 for every
node P b o t h if l = 2 k and if l--~2k+1. If l ~ 2 k , then ,~t(F)>2k, in
which case if . ~ ( F ) ~ 2 k + 2 , then by (1.13) 1 ~ contains an m-gon wit~
m ~ 2 k + 2 while if : - t ( F ) ~ 2 k + l , then by (1.10) 1" has an H-line and
therefore contains a (2k~-1)-gon. If l = - 2 k + 1 , then : t ( F ) > 2 k @ l and so
F contains an m-gon with m ~ 2 k @ 2 by (1.13). We have obtained a contra-
diction in every case.
So assume that q ( P ' ) ~ l / 2 and let I~'---I~---P '. Then F'~ G(n'--l, l).:
F' is connected and q ( P ' ) ~ 2 . F' is not an element of G*(n'--l,l). For if
this were the case, then P' would lie on a circuit with more than I edges.
From our induction hypothesis it follows that
r ( v ) = ,,,(p') + < l/2 + ( n ' - 2) I/2 = (n'-- 1) I/2.
This proves the theorem.
For the investigation of h(n, l) it is useful to consider the cases l - 2 k
and l ~- 2k @ 1 separately.
THEOREM (2.8)
h(n, 2k) ~ ( n - - 1 ) k ( k ~ 1),
if n - - 1 , then equality holds for all k and 1"~,~ is the extreme graph of
H(1,2k), if n > l and k = = l , then equality holds for all n and the star
with n nodes 1"i~ is the only extreme graph of the class H(n, 2), finally if
n >1 and k > l , then equality holds only if n ~ 2 k , and the complete (2k)-
graph is the only extreme graph of H(2k, 2k).
PROOF. Because H ( n , l ) ~ G ( n , l ) , we have that h(n, 2 k ) ~ g ( n , 2k).
By (2. 7), g(n, 2k) ~ (n--1)k, so
h(n, 2k)>(n--1)k.
Equality can hold only if n~=q(2k~l)-[-1 and if H(n, 2k) contains an
element of G*(n, 2k). But an element of G*(n, 2k) belongs to H(n, 2k) only
.if it contains no path with more than 2k edges. This holds only for the
graphs described in Theorem (2, 8), whether k ~ - 1 or k > 1.
~50 p. ERD6S AND T. GALLAI

If l - - 2 k + l , then, since H ( n , l ) = F ( n , 1), we need only consider

~he case k ~> 1.
THEOREM (2.9)
h(n, 2 k + l ) ~ n k ( k ~ I),

and here equality does not hold if k ~ 1 and n - - I or n - - 2, and it does

hold if k ~ 1 and n > 2, in which case the extreme graphs are those of which
all the components are A-stars; if k> 1, then equality holds only if
n ~ q ( 2 k + l) (q>O), and the only extreme graph is F,,.~+~.
The proof which is similar to the proof of (2.6) will be left to the reader.

w
In this paragraph we determine the graphs with the most edges among
the connected graphs of the class F(n, 2k) and the extreme graphs of the
class H(n, 2k) for sufficiently large values of n.
We denote the classes of the connected graphs in F(n, l), G(n, l), H(n, l),
respectively, by F(n,l), G(n, l), I4(n, l), and the number of edges in the
extreme graphs of these classes (the graphs with the most edges) by
f(n,l), ~(n, 1), h(n, l), respectively. From the fact that F,I is connected it
follows that each of these maximal numbers of edges is ~=~(n,l),
(3. 1) If F is an extreme graph of any one of the classes l~(n, 2k),
G(n, 2k), /~(n, 2k) ( k ~ 2 ) and if n > k ~ - - k + l, then F contains a 2k-gon.
PROOF. If I ' contains no 2k'gon, then 1"~ G(n, 2 k - - l ) , and so b y
(')
(2. 7) ~'(I') H ( n - - l ) k - - - - ~ . OI1 the other hand, it follows from the extremal
property of F and fi'om the remark on the graphs F/~ made above, that
T(F) >=~(F~ k) -= tl k - - k(k + 1)/'2. Accordingly,

from which it follows that n H k 2 - - k + l . This contradicts our hypothesis.

From Lemma (1.6) we deduce the following lemma:
LEMMA (3, 2) Let the graph F have an H-line and let ~ ( F ) ~ m >=4. If
F contains p mtttually H-independent nodes S~,..., Sp, then p <=m/2 and

2 J'
 ON MAXIMAL PATHS AND CIRCUITS OF GI~APHS 351

and if p = m/2, then equality in (1) holds only if F:-I';i and the nodes
$1,..., Sj) span a complete p-graph in F.
PROOF. The theorem is trivially true for m = 4. In what follows it will be
assumed that m > 4. Let C = (Pl, ..., P,,, P,,+I ~ P~) be an H-line of F and let
the nodes & - - P ~ i ( i = 1. . . . ,p) be mutually H-independent. It may be
assumed that 1= ol < . . . < o~ ~ m. Further let R~.= P~i-1, Ti=- P~r e(T 0--:- e;
(i - - 1, ..., p; P0 = P,,) and let the elements of the sets {R1,..., R~,}, {$1, ..., S~,}
and {'1"1,..., T~} be named R-, S-, T-nodes, respectively.
Because the S-nodes are H-independent, two S-nodes cannot be neigh-
bouts on C, and therefore this is true of the R- and T-nodes also, so that
an S-node cannot coincide with a T-node or with an R-node. It follows
that p ~ m/2.
By (1.6)
(2) e;+ej~m ( i , j - - 1 , . . . , p ; i=[=j).
Adding these inequalities together
(3) el + . . " + e~,~ pm/2.
Let F denote the complement of the graph F and let the degree of
T; in P be denoted by e; (i:--1 . . . . ,p). Then
(4) e; -[- e,: = m -- 1 (i = 1,..., p).
By the remark after Lemma (1.6) neither two R-nodes nor two T-nodes
can be joined in F. Consequently, all the edges TiT~. ( i , j = 1. . . . ,p;i=[=j)
are in I', and so the number of edges of F at least one end-node of which

Accordingly,

and so, having regard to (4) and (3),

r(T') >~p(m-- l)--(el + . . . . ~-ep)--(P)>=p(m--1)--

Hence
(6) = 2 + "
Equality holds here only if it holds in (2) and in (5) throughout. If
p = m/2, then this is the case only if e~ . . . . .
ep = m/2 and every edge
352 p. ERD()S AND T. GALLAI

of /= is incident with some T-node, that is to say any two nodes which are
not T-nodes are joined by an edge in I'. In this case, however, every
/?-node is also a T-node; and these are all joined to every S-node, further
any two S-nodes are joined to each other. This implies that F ~ , . ,14-~p
~ , and
that the S-nodes span a complete p-graph in F.
(3.3) If the connected graph I" contains a 2k-gon ( k > 2 ) b u t does
not contain any path having more than 2k edges and if n - - ~ ( F ) >
01.
~3k-1-2, then v(F)<cf(n, 2k), and equality holds only for F - - I ~ T .
PROOF. Let C ~ (P1, ..., Pe~, P2f~+l- - P~) be a 2k-gon of F. The nodes
of C will be called P-nodes and the remaining nodes of F will be
called Q-nodes. Let the Q-nodes be denoted by Q 1 , . . . , Q(~ where q - - n
2k>k@2.
Not two Q-nodes are joined in F. For if the edge QtQ2 exists, for
example, then, since F is connected, there is a path W in F which starts
in Q~ or Q2 and ends in a P-node, say P~, and does nol contain any
P-node other than P~, and contains only one of the nodes Q~, Q~ say Q1.
Then the edge QIQ, and the paths W and (P~. . . . . P~) together constitute
a path with at least 2 k @ 1 edges. This contradicts our hypotheses.
From this and from the fact that I" is connected it follows that every
Q-node is joined to some P-node.
If P; and Py are distinct P-nodes and if there are two distinct nodes
Q, and Q~ such that the edges P~Q, and PjQ~, exist, then P; and Pj are
H-independent in [C]. For such an open H-line leading from P~ to P3 in
[C] would, together with the edges P~Qg and PjQh, constitute a ( 2 k + 1)-path.
We divide the P-nodes into three classes. A P-node will be called an
c~-node if it is joined to at least two Q-nodes, it will be called a ~-node if
it is joined to exactly one Q-node and it will be called a y-node if it is not
joined to any Q-node. The number of ~-, 3- and y-nodes will be denoted
by p~, p~ and p~, respectively, p~ -t-p,~+ P v == 2k.
According to the above, any pair of cr are H-independent in [C],
and so are any ~r and any /?-node. Since two neighbouring nodes of
C are not H-independent in [C], the neighbours on C of an e-node can
only be y-nodes. It follows from this that p~ <p,~, and so p~ < k and
(1) p~ < 2/r 2p~.

C is an H-line of the graph [C], so Lemma (3.2) applies to the

e-nodes. Accordingly,

2 '
 ON Ik{AXIMAL P A T H S AND C I R C U I T S OF G R A P H S 353

The number of edges which join a P-node to a Q-node is

(3) "v~,~,,=<qp~, q-p~,
and so it follows from (1), (2) and (3) that

v(F)== v ( [ C ] ) + ~ p ~ ( 2 k ) + 2 k + ( q - - k - - 2 ) p ~ + ( p~+ I )
2 ~ "

Because q = > k + 2 , the expression on the right attains its maximum

value in the range O<=p~<~k only if p~=k, and a simple calculation shows
that this maximum is equal to nk__{k~r~ l l ~2 ( n , k 2 k )). S o ~ :, ( F ) ~ ( n ,
/

and equality holds only if p ~ = k and equality holds in (1), (2) and (3).
But if p ~ k , then p~-~-0 and equality holds in (1), and further, by (3. 2),
equality then holds in (2) only if [ C ] - - F ~ and the e-nodes span a com-
plete k-graph in [C]. Finally, because p~---0, equality holds in (3) when
every Q-node is joined to every a-node. From the properties which have
been enumerated it follows that F ~ F ~ j~.
THEOREM (3.4) /f n>k-'--k+6 (ki~ 1), then f(n, 2 k ) - - ~ ( n , 2k), and
the only extreme graph of the class tY(n, 2k) is l'~~'~.
PROOF (1) First assume that k = 1. By (2.6) f(n, 2)=<n with equality
holding only if n = 3 q , and then 1",,,~ is the only extreme graph. From this
and from f(n, 2 ) ~ f ( n , 2) it is seen that ](n, 2 ) =< n--1 ~q,(n, 2) except if
n-=3. But if l" is connected and :a(F)--n, v ( l ' ) = ~ n - - 1 , then F is a tree
([6], p. 47), and therefore does not contain a path with more than 2 edges
only if F = F ~ . The theorem is therefore true if k-- 1 and n >3.
(2) Assume that k ~ 2 . Then, because n > k 2 - - k + 6 , by (3.1) any
extreme graph F of /~(n, 2k) contains a 2k-gon, and since k~--k+6>=
3 k + 2, it follows from Theorem (,3. 3) that ~a(F) ~ Cp(n, 2k), equality hold-
ins only if .l 9 --,.,~.
l ~21';
Our theorem follows from this and from l%'f ~ F(n, 2k).
(3. 5). If I" is an extreme graph of the class H(n, 2k), then F has
not more than (k+ 1)/2 components.
PROOF. Suppose that the components of F are l'j, ...,/~, and :~(I~) - n~
( i ~ 1 . . . . . p). Then n ~ + . . . + n j , :n and 1",: is an extreme graph of the
class H(n~,k) for i--- 1..... p. Then, according to Theorem (2.8), v(l~)
=<(n~--l)k ( i = 1 . . . . ,p), and so
On the other hand, ~v(F)~ 't (n, 2k) by (2.3), therefore
(n--p)k ~z nk--k(k + 1)/2,
and hence p ~ (k-k 1)/2.
354 p. ERDOS AND T. OALLAI

THEOREM (3.6) If n > l - - ( k - [ - 1 ) 3, then h(n, 2k)=cf(n, 2k) and F~'~

is the only extreme graph of the class H(n, 2k).
PROOF. For k 1 the assertion of our theorem is contained in (2.8).
1
Suppose that k > 2 , n>)-(k-kl) ~, and F is an extreme graph of the class
H(n, 21"). According to (3.5), F then has a component F' such that
n'--~r(F')>(1"-b 1)~. F ' is an extreme graph of the class / - ) ( n ' , 2 k ) a n d
therefore contains a 2k-gon by (3.1). ( k - l - 1 ) 2 > 3 k § because k > 2 , so
by (3.3) v(F')<q~(n'_2l:), equality holding only if F ~ - I;~:. "~ But
,,E{l(n',21r and so /7~,,~ and v(F')--q:(n',2k). We show that
F ' - - F . For suppose that I"-J~F and let F " - - F - - F ' . n " - - ~ , ( F " ) ~ n - - n "
and F ' ~ H ( n " , 2 k ) . By (2.8) r(F")<-(n '' I)1., and so J'(F)--~'(F')-4-
[ _ _ \

+'v(I"')~n'k [k+21 } +(n" l)k--q~(n, 2k)--k. Butthiscontradicts(2.3).

CONIECTURES. We conjecture from the above that all extreme graphs

of the classes occurring in w 2 and w 3 can be found among the graphs
I~,, F,,,z and the members of the class G*(n, l). Among the twofold connect-
ed graphs IZ~, is probably in every case the only extreme graph if n >l--t- 1.

w
We are going to prove the following
THEOREM (4. 1) Let ~ ( F ) - - n . Assume further the maximum number
of independent edges is k (k > 1). Then

v(F) ~ max 2 '

Equality can occur only if 1 " - - F f , or if one component of F is a

complete ( 2 k + l)-graph and the other components are isolated nodes.
PROOF. We can clearly assume n>2k. Choose k independent edges
and call these e'-edges and the remaining edges ~-edges. The n - - 2 k nodes
of the graph which are not incident with a'-edges we call unsaturated.
Following BERGE ([2], p. 176) we add a node U to F and connect U with
every unsaturated node by an edge. The new graph we call F' and the new
edges and the old r162 we call r (U is incident with n - - 2 k
a-edges, every other node is incident with exactly one e-edge).
Let L be a directed U-loop. We call L alternating if its edges are
alternatingly e-edges and ~-edges in the ordering determined by L. A node
 ON MAXIMAL PATHS AND C I R C U I T S O F G R A P H S 355

P of F' is called e,accessible if there exists an alternatixlg &"-loop of final

node P whose last edge is an a-edge and it is called ~-accessible if there
exists an alternating U-lcop of final node P whose last edge is a g-edge.
Further, by definition, U is called g-accessible. The nodes which are g-acces-
sible but not e-accessible are called ~-nodes. Denote the number of fl-nodes
by , u + l .
It is easy t o see that U is not e-accessible, thus U is a g-node ([1],
[21, p. 176; [5], p. 140).
Denote the components of the graph obtained from F" by omitting the
~-nodes and the edges incident to it by Ih . . . . ,I~,, ( m >~ 1). I ) is called
odd or even if .~z(F;) is odd or even. If I~ is odd, put n ( F O ~ - 2 a i + l ,
if /~ is even, put ~ ( F ~ ) - - 2 a ; . An e-edge one node of which is a g-node
and the other node of which belongs to one of the I ) we call an entering
edge of ~ .
The following facts are well known ([I], [2], pp. 169--170; [5], pp.
141--142).
Every e-eclge incident to a ~'-node is an entering edge of some odd Fi
and every odd l~i has exactly one entering edge.
From this it follows that every a-edge is either an entering edge (of
some odd F~) or is an edge of some F~. Further 1"~ contains exactly
a~ e-edges.
r
The a-edges in F, (1 ~ i s m ) are clearly ~-edge~. We obtain their
number by subtracting from k the number of a'-edges incident to the
g-nodes, i.e. their number is k--,-. Thus

ai== k--!l.
i=1

If F~ is even, then v ( F i ) ~ ( 2 2 i ) . If 1"~ is odd, then

- 2 -- 2 -t-2a~.
Thus

~ t , ( l ~ ) <= ~ + ~ _ . ~ 2 a < 2k 2!, + 2 k - - 2 ~ , = 2/c-- , - + I .

"i=l i=1 \ i=l

Equality is only possible if every I'~ is odd and is a complete graph and
a~0 for all i ~ 2 .
The number of edges in I ' incident to the g-hades is less than or
equal to ( n - - , O l , + ( ~ ) .
356 P. E R D O S AND T, GALLAI: ON MAXIMAL P A T H S AND C I R C U I T S O F G R A P H S

Since every edge of F either belongs to one of the 17; or is incident

to one of the f~'-nodes, we obtain that

-- 2 T (n ,,),u-t- - - f(,u).

S i n c e f ( , , ) is a c o n v e x f u n c t i o n of , a n d 0 ~ , - < k, w e o b t a i n

Equality is only possible if u - - 0 or , ~ - - k , if , u ~ 0 , one of the /7; must

be a complete (2k-F1)-graph and the other F; must be isolated nodes.
If , , - - k , all the F~ must be isolated nodes and every 3-node must be con-
nected with all the nodes of F, i.e. I"---=1 "e~

(Received 24 June 1959)

Bibliography

[1] H. B. BZLCK, Regul~ire Faktoren yon Graphen, Journal f. reine u. angew. Math.. 188
(1950), pp. 228--252.
121 c. BERGZ, Theorie des graphes et ses applications (Paris, 1958).
[31 G. A. DIRAC, Some theorems on abstract graphs, Proc. London Math. Soc. (3), 2 (1952),
pp. 69--81.
141 P. ERD6S and A. H. STONE, On the structure of linear graphs, Bull, Amer. Math. Soc.,
52 (1946), pp. 1087--1091.
151 T. GaLLA~,On factorisation of graphs. Aeta Math. Acad. ScL Hung., ! (1950), pp. 133 153.
161 D. KGm6, Theorie der endlichen trod unendlichen Graphen (Leipzig, 1936).
I71 T. K6vAm, Vzna T. S6s and P. TuRAr~ On a problem of K. Zarankiewicz, Colloquium
Math., 3 (1954), pp. 50--57.
[81 D. ]. NEWMAN, A problem in graph theory, Amer. Math. Monthly, 65 (1958), p. 611.
[91 P. TurcAN, Egy grMelmGleti szGls6&tGkfeladatr61. Mat. ds Fiz. Lapok, 48 (1941), pp.
436 452.
11ol P. TURAN. On the theory of graphs, Colloquium Math., 3 (1954), pp. 19 30.
1111 K. ZARANKIEWmZ.Sur les relations symGtriques darts l'ensemble fini, Colloquium Math.,
1 (1947), pp. 10--14.