Counting independent sets in triangle-free graphs

Je Cooper

, Dhruv Mubayi

November 15, 2011

Abstract
Ajtai, Komlos, and Szemeredi proved that for suciently large t every triangle-
free graph with n vertices and average degree t has an independent set of size at
least
n
100t
log t. We extend this by proving that the number of independent sets in
such a graph is at least
2
1
2400
n
t
log
2
t
.
This result is sharp for innitely many t, n apart from the constant. An easy
consequence of our result is that there exists c

> 0 such that every n-vertex

triangle-free graph has at least
2
c

nlog n
independent sets. We conjecture that the exponent above can be improved to

n(log n)
3/2
. This would be sharp by the celebrated result of Kim which shows
that the Ramsey number R(3, k) has order of magnitude k
2
/ log k.
1 Introduction
An independent set in a graph G = (V, E) is a set I V of vertices such that no two
vertices in I are adjacent. The independence number of G, denoted (G), is the size of
the largest independent set in G. Determining the independence number of a graph is one
of the most pervasive and fundamental problems in graph theory. The independence
number naturally arises when studying other fundamental graph parameters like the
chromatic number (minimum size of a partition of V into independent sets), clique
number (independence number of the complementary graph), minimum vertex cover

Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago,

IL 60607; email: [email protected]

Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago,

IL 60607; email: [email protected]; research supported in part by NSF grant DMS 0969092.
1
(complement of a maximum independent set), matching number (independence number
in the line graph) and many others.
Throughout this paper, we suppose that G is a graph with n vertices and average
degree t. Turans [11] basic theorem of extremal graph theory, in complementary form,
states that (G) n/(t+1) for any graph G. This bound is tight, as demonstrated by
the complement of Turans graph G = T(n, r) which, in the case n = kr is the disjoint
union of r cliques, each with k vertices (then (G) = r and t = k 1). Since G contains
large cliques it is natural to ask whether Turans bound on (G) can be improved if we
prohibit cliques of a prescribed (small) size in G.
In [1], Ajtai, Koml os, and Szemeredi showed that if G contains no K
3
, then this is
indeed the case, by improving Turans bound by a factor that is logarithmic in t. More
precisely, they proved that if G is triangle-free, then
(G)
n
100t
log t.
Shortly after, Shearer [10] improved this to (G) (1 o(1))
n
t
log t (assume for con-
venience throughout this paper that log = log
2
). Random graphs [10] show that for
innitely many t and n with t = t(n) as n , there are n-vertex triangle-free
graphs with average degree t and independence number (2 o(1))((n/t) log t). Conse-
quently, the results of [1, 10] cannot be improved apart from the multiplicative constant.
There is a tight connection between the problem of determining (G) and questions in
Ramsey theory. More precisely, determining the minimum possible (G) for a triangle-
free G is equivalent to determining the Ramsey number R(3, k), which is the minimum
n so that every graph on n vertices contains a triangle or an independent set of size
k. Moreover, the above lower bounds for (G) are equivalent to the upper bound
R(3, t) = O(t
2
/ log t). It was a major open problem, dating back to the 1940s, to
determine the order of magnitude of R(3, t), and this was achieved by Kim [7] who
showed that for every n suciently large, there exists an n-vertex triangle-free graph G
with (G) < 9

nlog n. As a consequence, the upper bound R(3, t) = O(t

2
/ log t) from
[1] is of the correct order of magnitude.
In this paper, our goal is to take the result of Ajtai, Koml os, and Szemeredi [1] further
by not only nding an independent set of the size guaranteed by their result, but by
showing that many of the vertex subsets of approximately that size are independent
sets.
Denition 1. Given a graph G, let i(G) denote the number of independent sets in G.
Upper bounds for i(G) have been motivated by combinatorial group theory. In [2],
Alon showed that if G is a d-regular graph, then i(G) 2
(1/2+o(d))n
; he also conjectured
2
that
i(G) (2
d+1
1)
n/2d
.
Kahn [6] proved this conjecture for d-regular bipartite graphs. Galvin [5] obtained a
similar bound for any d-regular graph G, namely
i(G) 2
n/2(1+1/d+c/d

log d/d)
for some constant c . Finally, Zhao [12] recently resolved Alons conjecture.
In this paper, we consider lower bounds for i(G). This problem is fundamental in
extremal graph theory, indeed, the Erd os-Stone theorem [4] gives a lower bound for
i(G) that is the correct order of magnitude provided n/t is a constant. More recently,
the problem in the range t = (n) has been investigated by Razborov [9], Nikiforov [8],
and Reiher. For example, the results of Razborov and Nikiforov determine g(, 3), the
minimum triangle density of an n-vertex graph with edge density
1
2
< < 1. Looking at
the complementary graph, this gives tight lower bounds on the number of independent
sets of size three in a graph with density 1
t
n
.
Lower bounds for i(G) appear not to have been studied with the same intensity when
t is much smaller than n, in particular, when t and t/n 0. Let us make some
easy observations that are relevant for our work here. We assume that := (G) n/4.
Since every subset of an independent set is also independent, Turans theorem implies
i(G) 2

2
n/(t+1)
.
In Section 2, we will improve this to
i(G) 2
1
250
n
t
log t
. (1)
Our proof uses the standard probabilistic argument which establishes the order of mag-
nitude given by Turans bound on (G). This result is certainly not new, and we present
it only to serve as a warm-up for our main result in Section 3. Let us observe below
that the result is essentially tight.
As no subset of size more than (G) is independent, an easy upper bound on i(G)
(using n/4) is
i(G)

i=0

n
i

. (2)
Since (T(kr, r)) = r = n/(t +1) (recall that n = kr and t = k 1), this bound implies
that as n
i(T(kr, r)) 2

kr
r

2(ek)
r
= 2
1+r log ek
= 2
(1+o(1))
n
t
log t
.
3
Thus, apart from the constant, the exponent in (1) cannot be improved.
Our main result addresses the case where G contains no triangles. As in the case of
the independence number, prohibiting triangles improves the bound in (1).
Theorem 2. (Main Result) Suppose that G is a triangle-free graph on n vertices with
average degree t, where t is suciently large. Then
i(G) 2
n
2400t
log
2
t
. (3)
Suitable modications of Random graphs provide constructions of n-vertex triangle-free
graphs G with average degree t = t(n) as n , and (G) = O((n/t) log t).
Plugging this into (2), we see that Theorem 2 is tight (apart from the constant) for
innitely many t. However, it remains open if the theorem is sharp for all t where
t = n
1/2+o(1)
. Indeed, the open problem that remains is to obtain a sharp lower bound
on i(G) for triangle-free graphs with no restriction on degree. Since all subsets of the
neighborhood of a vertex of maximum degree are independent, i(G) > 2
t
. Combining
this with (3) we get
i(G) > max{2
t
, 2
n
2400t
log
2
t
} > 2
cn
1/2
log n
for some constant c > 0. We conjecture that this can be improved as follows.
Conjecture 3. There is an absolute positive constant c such that every n-vertex triangle-
free graph G satises
i(G) > 2
cn
1/2
(log n)
3/2
.
The conjecture, if true, is sharp (apart from the constant in the exponent) by the
graphs (due to Kim [7] and more recently Bohman [3]) which show that R(3, t) =
(t
2
/ log t). Indeed, their graphs are triangle-free and have independence number
(G) = (t) = (
n
t
log t) = (

nlog n),
so i(G) 2
O(

nlog
3/2
n)
by (2).
As mentioned before, throughout the paper, all logarithms are base 2. For a graph
G, let n(G), e(G) and t(G) denote the number of vertices, edges, and average degree of
G.
2 General case
In this section, we give the simple proof of (1). Our purpose in doing this is to familiarize
the reader with the general approach to the proof of Theorem 2 in the next section.
4
Proposition 4. If G is a graph on n vertices with average degree t, where 2 t
n
800
,
then i(G) 2
1
250
n
t
log t
.
Proof. Set k =
1
100
n
t
. Pick a k-set uniformly at random from all k-sets in V (G). Let
H be the subgraph induced by the k vertices. Then
E[e(H)] =
1
2
nt

n2
k2

n
k
=
1
2
nt
k(k 1)
n(n 1)
<
1
2
tk
2
n
.
Recall that Markovs inequality states that if X is a positive random variable and a > 0,
then Pr[X a] E[X]/a; hence Pr[e(H) 2
1
2
tk
2
n
] 1/2. So for at least half of the
choices for H, e(H)
tk
2
n
. Therefore, the number of choices of H for which e(H)
tk
2
n
is at least
1
2

n
k

1
2
(
n
k
)
k
> 2
k
2
log n/k
= 2
k
2
(log nlog k)
. (4)
Now, if e(H) <
tk
2
n
=
1
100
k, then at most
1
50
k of the vertices in H have degree at least
one. This in turn implies that H contains an independent set I of size at least
49
50
k. The
set I can be obtained from any H which contains it; the number of ways to pick the
1
50
k
vertices of H I is at most

n
k/50

(
50ne
k
)
k/50
= 2
k
50
log 50ne
k
50
log k
2
k
50
(log nlog k)+
k
50
log 100t
. (5)
Combining this with (4) and using
20
23
1
100
n
t
< k
1
100
n
t
for t
n
800
,
i(G) 2
k(
1
2

1
50
)(log nlog k)
k
50
log 100t
2
k
24
50
log 100t
k
50
log 100t
= 2
k
23
50
log 100t
> 2
1
250
n
t
log t
.
3 Triangle-free graphs
In this section we prove our main result, Theorem 2. We begin with some modications
of a lemma from [1] (see the proof of Lemma 4 in [1]).
Lemma 5. (Ajtai-Komlos-Szemeredi [1]) Suppose that G is a triangle-free graph on
n vertices with average degree t, and let k n/100t. Let H be the subgraph consisting
of k vertices chosen uniformly at random from all the k-sets contained in {v V (G) :
deg(v) 10t}. Let M be the subgraph of G consisting of vertices adjacent to no vertex
in H. Let n

and t

denote the number of vertices and average degree of M. Then the

5
random variables H and M satisfy:
E[n(M)] > n(1
k
n t
)
t+1
>
9n
10
(6)
E[e(M)] >
nt
2
(1
k
n 20t
)
20t+1
>
nt
10
(7)
E[e(H)] =
1
2
nt
k(k 1)
n(n 1)

tk
2
n
(8)
Var[n(M)] <
2nk(t + 1)(10t + 1)
n k 20t 2
< nt (9)
Var[e(M)] < 2400kt
4
< 40nt
3
(10)
Var[e(H)]
tk
2
(10k + n)
n
2
(11)
Further, if e(M) < (1 + ) E[e(M)] and n(M) > (1 ) E[n(M)], then n

/t

> n/t,
where = 800

t/n and = 1 1/t c

10

t/n for some positive constant c

10
.
Remark. Ajtai-Koml os-Szemeredi state their lemma for k = n/100t and prove each of
the rst inequalities in (6), (7), (9), and (10) for all k. They prove each of the second
inequalities for k = n/100t, but it is easily observed that they continue to hold for
k < n/100t.
The next lemma is implied by the computation in the proof of Lemma 4 from [1].
However, the last statement of Lemma 6 is crucial to our proof of Theorem 2, so we
make the computations in [1] explicit.
Lemma 6. Suppose G is a triangle-free graph on n > 2
50
vertices with average degree
t 2

nlog n and k n/100t. Then G contains a subgraph H with n(H) = k e(H)

k/50. Moreover, if M is the subgraph of G consisting of vertices adjacent to no vertex
in H, then
1. n(M) > n(G)/2 and
2. n(M)/t(M) > n/t, where = 1 1/t c
10

t/n.
Further, if the vertices in H are chosen uniformly at random from all the k-sets contained
in {v V (G) : deg(v) 10t}, then at least half of the choices for H satisfy e(H) k/50,
along with conditions 1 and 2.
6
Proof. Recall that for a random variable X and a > 0, Chebyshevs inequality states
that Pr[|X E[X]| a] Var[X]/a
2
. Thus, with a = k/50 E[e(H)], Lemma 5
implies
Pr[e(H) k/50]
tk
2
(10k + n)
n
2
(k/50
k
2
t
2n
)
2
=
t(10k + n)
n
2
(1/50
kt
2n
)
2

t(n/10t + n)
n
2
(1/50 1/200)
2
=
1/10 + t
n(3/200)
2
< 5000
t
n
5000
2 log n

n
1/1000.
So with probability at most 1/1000, the condition e(H) k/50 fails.
Set = 800

t/n. Again by Lemma 5 and Chebyshev, with a = E[n(M)],

Pr[n(M) n/2] Pr[n(M) (1 ) E[n(M)]]
nt
(
9n
10
)
2
800
2
t
n
1/1000.
Thus the probability that condition (1) fails is at most 1/1000.
With a = E[e(M)],
Pr[e(M) (1 + ) E[e(M)]]
40nt
3
nt
10
800
2
t
n
= 1/160.
Since Pr[e(M) (1+) E[e(M)] or n(M) n/2] < 1/160+1/1000, the last assertion of
Lemma 5 implies that the probability of condition (2) failing is at most 1/160 +1/1000.
Therefore, the probability that condition e(H) k/50 fails or condition (1) fails or
condition (2) fails is at most 1/1000 + 1/1000 + 1/160 + 1/1000 < 1/2.
Our proof of Theorem 2 is achieved by analyzing Algorithm 1 below. The algorithm
is a slight modication of the algorithm from [1] that yields an independent set of size
1
100
n
t
log t. Recall that c
10
is the constant that appear in Lemma 5.
Theorem 7. Suppose Algorithm 1 is run on a triangle-free graph G with n vertices and
average degree t, where 2
100
< t <

nlog n and n > (3c
10
)
12
. If Algorithm 1 terminates
at line 12, then |I| >
1
2
n
t
log
2
t. Otherwise, for each iteration i = 0, ..., R1, Algorithm
7
Algorithm 1: Independent set algorithm
Input: Triangle-free graph G with n vertices, average degree t
Output: Independent set I
1 M
o
= G;
2 R = (log t)/2;
3 for i 0 to R do
4 n
i
= number of vertices in M
i
;
5 t
i
= average degree in M
i
;
6
i
= 1 1/t
i1
c
10

t
i1
/n
i1
;
7 if i = 0 or
i
> 1 1/ log t then
8 Apply Lemma 6 with G = M
i
and k =
1
200
n
t
;
9 M
i+1
, H
i+1
= M, H from Lemma 6;
10 else
11 I = Independent set in M
i1
of size n
i1
/(t
i1
+ 1);
12 return I;
13 end
14 end
15 H = H
1
H
R
;
16 I = Independent set in H of size
48
50
kR;
17 return I;
1 successfully applies Lemma 6 to the graph M
i
to obtain a graph H
i+1
with k =
1
200
n
t

vertices. Moreover, the graph H in line 15 is the disjoint union of the H

i
, and the
independent set I in line 16 consists of
48
50
kR vertices from H.
Proof. We break our proof into two cases, depending on whether Algorithm 1 terminates
at line 17 or 12.
Line 17: We need to show that Lemma 6 can be applied at every iteration and that
the graph H in line 15 contains an independent set of size at least
48
50
kR
1
500
n
t
log t.
If i = 0, then k n/100t, t

nlog n, and n > t > 2
50
, so we may apply Lemma 6
to obtain graphs M
1
and H
1
, where |V (H
1
)| = k. Suppose that i > 0 and that Lemma
6 was successfully applied at all previous iterations. Using 1 of Lemma 6, i < R, and
R = (log t)/2 < (log n)/2,
n
i
n/2
i
> n/2
R
>

n >

t > 2
50
. (12)
By the condition in line 7,
i
> 1 1/ log t for each iteration i. So by 2 of Lemma 6,
8
n
i
t
i

n
t

2
. . .
i
>
n
t
(1 1/ log t)
R
. Thus
n
i
t
i
>
n
t
(1 1/ log t)
R
>
n
t
(1
R
log t
)
1
2
n
t
.
In particular,
k
1
200
n
t

1
100
n
i
t
i
, (13)
and also, t <

nlog n and n
i
< n yield
t
i
< 2n
i
t
n
2n
i
log n

n
2n
i
log n
i

n
i
= 2

n
i
log n
i
. (14)
The inequalities (12), (13), and (14) ensure that we may again apply Lemma 6 with
M
i
, M
i+1
, H
i+1
, and k playing the roles of G, M, H, and k, respectively. Applying Lemma
6 R times yields a collection of sparse graphs H
1
, H
2
, . . . , H
R
, each with k vertices.
Each H
i
contains at most
2
50
k vertices of degree at least one, so each H
i
contains an
independent set of size at least
48
50
k. By denition of M
i
, these independent sets may be
combined into one independent set of size at least
48
50
kR.
Line 12: Suppose that the algorithm terminates at line 12 during iteration i +1. Then
n
i
, t
i
(and n
i+1
, t
i+1
) have been dened and
i+1
= 1 1/t
i
c
10

t
i
/n
i
. If t
i
(
3
2
)
2/3
,
then 1/t
i
> 1/ log
3
t. Assume t
i
> (
3
2
)
2/3
. Then
2
3
t
1/3
i
> 1/t
i
. (15)
By (12), n
i
>

n, so for n > (3c
10
)
12
,
n
3
i
> n
3/2
> (3c
10
)
6
n > (3c
10
)
6
t
i
.
This implies
1
3
t
1/3
i
> c
10

t
i
/n
i
. (16)
Combining (15) and (16) yields
1 t
1/3
i
< 1 1/t
i
c
10

t
i
n
i
=
i+1
1 1/ log t.
Thus 1/t
i
> 1/ log
3
t. Since t 2
100
(which implies that t > (log
5
t + log
2
t)
2
) and
i < R
log t
2
,
n
i
t
i
+ 1
>
n
2
i
1
(log
3
t + 1)

t
1
(log
3
t + 1)
=
n
t

t
(log
3
t + 1)
>
n
t
log
2
t.
Turans theorem now implies that M
i
contains an independent set of size at least
n
i
t
i
+1
>
1
2
n
t
log
2
t.
9
We now complete the proof of Theorem 2 by obtaining a lower bound on the number
of outcomes given by line 17 of Algorithm 1.
Proof of Theorem 2. Recall that we are to show that if G is a triangle-free graph
on n vertices with average degree t suciently large, then i(G) 2
n
2400t
log
2
t
. Assume
t > max{(3c
10
)
12
, 2
100
}. Then n > t > (3c
10
)
12
. Also, if t >

nlog n, then G has
a vertex whose neighborhood contains at least t >
n

n
log n >
n
t
log
2
n vertices. Since
G is triangle-free, this neighborhood forms an independent set, which contains at least
2
n
t
log
2
n
> 2
n
2400t
log
2
t
subsets, which are also independent. Thus we may assume that
t

nlog n; in particular, G satises the hypotheses of Theorem 7.
If the algorithm terminates at line 12, then G contains an independent set of size at
least
n
2t
log
2
t; so G contains at least 2
n
2t
log
2
t
> 2
n
2400t
log
2
t
independent sets, and we are
done. Thus we may assume that Algorithm 1 terminates at line 17. Consequently, at
each iteration i, the algorithm applies Lemma 6 to pick a sparse graph with k =
1
2
n
100t

vertices. The vertices in this graph are chosen from

L
i
= {v V (M
i
) : deg(v) 10t
i
}.
Note that
n
i
t
i
=

vL
i
deg(v) +

vV (M
i
)L
i
deg(v)

vV (M
i
)L
i
deg(v) (n
i
|L
i
|)10t
i
.
This, together with 1 in Lemma 6, implies |L
i
|
9
10
n
i
>
9
10
n
i1
/2
9
10
n/2
i
. At least
half of the k-sets in L
i
satify the conditions of Lemma 6, so the number of choices for
H
i
is at least
1
2

|L
i
|
k

1
2

.9n/2
i
k

.
Therefore, the number of choices for the sequence H
1
, . . . , H
R
is at least
R1

i=0
1
2

|L
i
|
k

1
2
R
R1

i=0

.9n/2
i
k

1
2
R
(
.9n
k
)
kR
2
kR
2
/2
= 2
kRlog .9nkRlog kkR
2
/2R
= 2
kR(log nlog k)kR
2
/2+kRlog .9R
> 2
kR(log nlog k)
kR
2
log t
2
kRR
> 2
kR(log nlog k)
kR
4
log t2kR
. (17)
Recall that Algorithm 1 obtains an independent set I of size
48
50
kR from the graph
H = H
1
H
R
. For a xed I, the number of graphs H that yield I is at most the
10
number of possibilities for H I. This is at most

n
|V (HI)|

, which is at most

n
2
50
kR

(
50ne
2kR
)
2
50
kR
= 2
2
50
kRlog n+
2
50
kRlog 50e
2
50
kRlog 2kR
< 2
2
50
kR(log nlog k)+
2
50
kRlog 50e
< 2
2
50
kR(log nlog k)+kR
. (18)
For a xed H, the number of partitions H
1
H
R
= H is at most the number of
partitions of kR elements into R sets of size k, which is less than

kR
k

R
(Re)
kR
= 2
kRlog Re
< 2
kRlog R+2kR
< 2
kR
1
4
R+2kR
2
kR
1
8
log t+2kR
. (19)
Since each H yields an independent set, the total number of independent sets that can
be returned at line 17 of the algorithm is at least
# of ways to obtain H
(# of H that yield a xed I)(# of partitions that yield H)
.
Since
n
268t
k
n
200t
and R > (log t)/3, (17), (18), and (19) imply that this is at least
2
48
50
kR(log nlog k)
5
8
kRlog t5kR
2
48
50
kRlog 200t
5
8
kRlog t5kR
> 2
134
400
kRlog t
> 2
134
1200
k log
2
t
> 2
1
2400
n
t
log
2
t
.
References
[1] Miklos Ajtai, Janos Komlos, and Endre Szemeredi, A dense innite Sidon sequence,
European J. Combin. 2 (1981), no. 1, 111. MR 611925 (83f:10056)
[2] Noga Alon, Independent sets in regular graphs and sum-free subsets of nite groups,
Israel J. Math. 73 (1991), no. 2, 247256. MR 1135215 (92k:11024)
[3] Tom Bohman, The triangle-free process, Adv. Math. 221 (2009), no. 5, 16531677.
MR 2522430 (2010h:05271)
[4] Paul Erdos and Arthur H. Stone, On the structure of linear graphs, Bull. Amer.
Math. Soc. 52 (1946), 10871091. MR 0018807 (8,333b)
11
[5] David Galvin, An upper bound for the number of independent sets in regular graphs,
Discrete Math. 309 (2009), no. 23-24, 66356640. MR 2558628 (2010m:05217)
[6] Je Kahn, An entropy approach to the hard-core model on bipartite graphs, Combin.
Probab. Comput. 10 (2001), no. 3, 219237. MR 1841642 (2003a:05111)
[7] Jeong Han Kim, The Ramsey number R(3, t) has order of magnitude t
2
/ log t, Ran-
dom Structures Algorithms 7 (1995), no. 3, 173207. MR 1369063 (96m:05140)
[8] Vladimir Nikiforov, The number of cliques in graphs of given order and size, Trans.
Amer. Math. Soc. 363 (2011), no. 3, 15991618. MR 2737279
[9] Alexander A. Razborov, On the minimal density of triangles in graphs, Combin.
Probab. Comput. 17 (2008), no. 4, 603618. MR 2433944 (2009i:05118)
[10] James B. Shearer, A note on the independence number of triangle-free graphs, Dis-
crete Math. 46 (1983), no. 1, 8387. MR 708165 (85b:05158)
[11] Paul Tur an, On an extremal problem in graph theory (in hungarian), Math. Fiz.
Lapok 48 (1941), 436452.
[12] Yufei Zhao, The number of independent sets in a regular graph, Combin. Probab.
Comput. 19 (2010), no. 2, 315320. MR 2593625 (2011e:05200)
12