a

r
X
i
v
:
1
4
0
4
.
3
3
0
6
v
1


[
m
a
t
h
.
C
O
]


1
2

A
p
r

2
0
1
4
Decomposing random graphs into few cycles and edges
Daniel Korandi

Michael Krivelevich

Benny Sudakov

Abstract
Over 50 years ago, Erd os and Gallai conjectured that the edges of every graph on n vertices
can be decomposed into O(n) cycles and edges. Among other results, Conlon, Fox and Sudakov
recently proved that this holds for the random graph G(n, p) with probability approaching 1. In
this paper we show that for most edge probabilities G(n, p) can be decomposed into a union of
n
4
+
np
2
+o(n) cycles and edges whp. This result is asymptotically tight.
1 Introduction
Problems about packing and covering the edge set of a graph using cycles and paths have been
intensively studied since the 1960s. One of the oldest questions in this area was asked by Erdos and
Gallai [7, 8]. They conjectured that the edge set of any graph G on n vertices can be covered by
n 1 cycles and edges, and can be partitioned into a union of O(n) cycles and edges. The covering
part was proved by Pyber [11] in the 1980s, but the partitioning part is still open. As noted in [8],
it is not hard to show that O(nlog n) cycles end edges suce. This bound was recently improved to
O(nlog log n) by Conlon, Fox and Sudakov in [6], where they also proved that the conjecture holds
for random graphs and graphs of linear minimum degree. The present paper treats the problem in
the case of random graphs in more detail.
Let 0 p(n) 1 and dene G(n, p) to be the random graph on n vertices, where the edges are
included independently with probability p. We hope to nd a close to optimal partition of the edges
of a random graph into cycles and edges. Observe that in any such partition each odd-degree vertex
needs to be incident to at least one edge, so if G(n, p) has s odd-degree vertices, then we certainly
need at least s/2 edges. Also, a typical random graph has about
_
n
2
_
p edges, whereas a cycle may
contain no more than n edges, so we need at least about
np
2
cycles (or edges) to cover the remaining
edges. This simple argument gives a lower bound of
np
2
+
s
2
on the optimum.
In this paper we show that this lower bound is asymptotically tight. Let odd(G) denote the
number of odd-degree vertices in the graph G. We say that G(n, p) (with p = p(n)) satises some
property P with high probability or whp if the probability that P holds tends to 1 as n approaches
innity. Our main result is the following theorem.

Department of Mathematics, ETH, 8092 Zurich. Email: [email protected].

School of Mathematical Sciences, Raymond and Beverly Sackler Faculty of Exact Sciences, Tel Aviv University,
Tel Aviv 6997801, Israel. Email: [email protected]. Research supported in part by a USA-Israel BSF grant and
by a grant from the Israel Science Foundation.

Department of Mathematics, ETH, 8092 Zurich. Email: [email protected]. Research supported in

part by SNSF grant 200021-149111 and by a USA-Israel BSF grant.
1
Theorem 1.1. Let the edge probability p(n) satisfy p =
_
log log n
n
_
. Then whp G(n, p) can be split
into
odd(G(n,p))
2
+
np
2
+o(n) cycles and edges.
In fact, as we show in Lemma 2.2, for most of the ps in the range, odd(G(n, p))
n
2
. This
immediately implies the following, perhaps more tangible, corollary.
Corollary 1.2. Let p = p(n) be in the range [
_
log log n
n
_
, 1
_
1
n
_
]. Then whp G(n, p) can be split
into
n
4
+
np
2
+o(n) cycles and edges.
Here we use the standard notation of (f) for any function that is asymptotically greater than
the function f(n), i.e., g(n) = (f(n)) if lim
n
g(n)
f(n)
= . In this paper log stands for the natural
logarithm, and for the sake of clarity we omit the oor and ceiling signs whenever they are not
essential. We call G an Euler graph if all the vertices of G have even degree (not requiring that G
be connected).
1.1 Proof outline
We will break the probability range into three parts (the sparse, the intermediate and the dense
ranges), and prove Theorem 1.1 separately for each part. The proofs of the denser cases build on the
sparser cases, but all the proofs have the following pattern: we start with deleting
odd(G(n,p))
2
+o(n)
edges so that the remaining graph is Euler, and then we extract relatively few long cycles to reduce
the problem to a sparser case.
In the sparse case we will check the Tutte condition to show that there is a large matching on
the odd-degree vertices, and then use expansion properties to iteratively nd cycles that are much
longer than the average degree. In the end, we are left with a sparse Euler subgraph, which breaks
into o(n) cycles.
The denser cases are somewhat more complicated. We will need to break G(n, p) into several
random subgraphs. These graphs will not be independent, but we can remove edges from one of
them without aecting the random structure of the others.
In the intermediate case we break G(n, p) into three random subgraphs, G(n, p) = G
1
G
2
G
3
.
First we nd an edge set E
0
in G
2
such that G(n, p) E
0
is Euler. Then we break (G
2
G
3
) E
0
into matchings and G
1
into even sparser random graphs. Using a result by Broder, Frieze, Suen and
Upfal [4] about paths connecting a prescribed set of vertex pairs in random graphs, we connect up
the matchings into cycles using the parts of G
1
. The remaining edges are all from G
1
, and the tools
from the sparse case take care of them.
In the dense case we break into four subgraphs, G(n, p) = G
1
G
2
G
3
G
4
, where G
4
contains
the majority of the edges. Again we start by nding the edge set E
0
in G
3
. Next, we apply a recent
packing result by Knox, K uhn and Osthus [9] to nd many edge-disjoint Hamilton cycles in G
4
.
Then we break the remaining edges from G
3
G
4
into matchings and use G
2
to connect them into
cycles. At this point we have a still intact random graph G
1
and some edges from G
2
left, and these
t into the intermediate setting, hence the previous results complete the proof.
The paper is organized as follows: in Section 2 we prove all the results we need about odd-degree
vertices, including the typical existence of E
0
and the fact that normally about half the vertices
2
are odd. Section 3 shows why we can iteratively remove relatively long cycles from sparse random
graphs, and proves Theorem 1.1 in the sparse range. In Section 4 we show the details of breaking into
matchings and then connecting them into few cycles, and prove the main lemma for the intermediate
range. Finally, we complete the proofs of the denser cases in Section 5.
2 Covering the odd-degree vertices
The rst step in the proof is to reduce the problem to an Euler subgraph by removing relatively few
edges. This section is devoted to the discussion of the results we need about odd-degree vertices in
random graphs.
Throughout the paper, we will use the following Cherno-type inequalities (see, e.g., [1]).
Claim 2.1. Suppose we have independent indicator random variables X
1
, . . . , X
n
, where X
i
= 1 with
probability p
i
and X
i
= 0 otherwise, for all i = 1, . . . , n. Let X = X
1
+ + X
n
be the sum of the
variables and p = (p
1
+ +p
n
)/n be the average probability so that E(X) = np. Then:
(a) P(X > np +a) e
2a
2
/n
for all a > 0,
(b) P(X > 2np) e
np/20
,
(c) P(X < np a) e
a
2
/np
and hence
(d) P(X < np/2) e
np/20
.
In particular, all the above estimates hold when X Bin(n, p) is a binomial random variable.
First we need the following estimate on the number of vertices of odd degree in G(n, p).
Claim 2.2. Suppose the probability p satises
_
1
n
_
< p < 1
_
1
n
_
. Then whp G = G(n, p)
contains
n
2
(1 +o(1)) vertices of odd degree.
Proof. Let us x a bipartition of the vertices into two nearly equal parts V = V
1
V
2
, where
n
1
= |V
1
| = n/2 and n
2
= |V
2
| = n/2. Our plan is to show that whp roughly half the vertices of
V
1
and roughly half the vertices of V
2
have odd degree in G.
So let Y
1
be the number of odd-degree vertices in V
1
. The probability that a specic vertex v
has odd degree is
1(12p)
n1
2
, so by the linearity of expectation, the expected number of odd degree
vertices in V
1
is
E(Y
1
) =
n
1
2

n
1
(1 2p)
n1
2

n
1
2
for (1/n) p 1 (1/n). We still need to prove that Y
1
is tightly concentrated around its mean.
Let us now expose the random edges spanned by V
1
. Then the degrees of the vertices in V
1
are
determined by the crossing edges between V
1
and V
2
. At this point, each vertex in V
1
has n
2
incident
edges unexposed, and these edges are all dierent. So let v be any vertex in V
1
. The probability that
v is connected to an odd number of vertices in V
2
is p
1
=
1(12p)
n
2
2

1
2
, so the probability that v
will end up having an odd degree is either p
1
or (1 p
1
) (depending on the parity of its degree in
G[V
1
]). This means that, conditioning on any collection of edges spanned by V
1
, Y
1
is the sum of
3
n
1
indicator variables with probabilities p
1
or 1 p
1
, so we can apply Lemma 2.1(a) and (c) with
average probability p
1
p
1
1 p
1
to get
P(|Y
1
E(Y
1
)| > a) e
2a
2
/n
1
+e
a
2
/n
1
p
1
.
Then taking a = n
2/3
and using that p
1
1/2 we get
P(|Y
1
E(Y
1
)| > n
2/3
) 3e
2n
1/3
= o(1).
for large enough n.
Repeating the same argument for V
2
, we see that if Y
2
is the number of odd-degree vertices in
V
2
, then
E(Y
2
) =
n
2
n
2
(1 2p)
n1
2
n
2
/2
and
P(|Y
2
E(Y
2
)| > n
2/3
) 3e
2n
1/3
.
So with probability 1 6e
2n
1/3
= 1 o(1) the number of odd vertices is
n
2
(1 +o(1)).
In order to show that typically there is a small set of edges covering the odd vertices, we will
need some properties of sparse random graphs. The following somewhat technical lemma collects
the tools we use in the proof.
Lemma 2.3. Let p = p(n) satisfy
_
log log n
n
_
p
log
10
n
n
. Then the following properties hold whp
for G = G(n, p):
1. the giant component of G covers all but at most o(n) vertices, and all other components are
trees,
2. the independence number (G) is at most
2 log(np)
p
,
3. the diameter of the giant component of G is at most
2 log n
log(np)
,
4. any set T of at most 2n
1/10
vertices spans at most 2|T| edges,
5. any two disjoint vertex sets T, T

of the same size n

1/10
t n/100 are connected by less than
tnp/6 edges, and
6. all but at most n/2 log
2
n vertices in G have at least np/5 odd-degree neighbors.
Proof. For the rst two properties we refer to Bollobas [2], while the third property was proved by
Chung and Lu [5]. To prove the fourth one, note that the probability that a xed set T of size t
spans at least 2t edges is at most
_
(
t
2
)
2t
_
p
2t
. So the probability that some T of size t 2n
1/10
spans
at least 2t edges is at most
2n
1/10

t=5
_
n
t
_

_
t
2
2t
_
p
2t

2n
1/10

t=5
(nt
4
p
2
)
t

2n
1/10

t=5
n
t/2
= o(1).
4
We use a similar argument to prove the fth property. For xed T
1
and T
2
of size t, the probability
that at least tnp/6 edges connect them in G is at most
_
t
2
tnp/6
_
p
tnp/6
. So the probability that the
property does not hold can be bounded from above by
n/100

t=n
1/10
_
n
t
_
2

_
t
2
tnp/6
_
p
tnp/6

n/100

t=n
1/10
_
en
t
_
2t

_
et
2
p
tnp/6
_
tnp/6

n/100

t=n
1/10
_
e
2
n
2
t
2

_
6et
n
_
log log n/6
_
t
.
Here
6et
n
<
1
2
, so for large enough n this probability is smaller than

n/100
t=n
1/10
(1/2)
t
< 22
n
1/10
= o(1).
The last property has a similar avor to Claim 2.2. We will prove that the probability that a
particular vertex has fewer than np/5 odd neighbors is at most 1/ log
3
n. Then the expected number
of such vertices is at most n/ log
3
n, hence the probability that there are more than n/2 log
2
n of
them is, by Markovs inequality, at most 2/ log n = o(1).
So let us pick a vertex v in G. Using Claim 2.1, we see that the probability that it has fewer than
np/2 or more than 2np neighbors is at most 2e
np/20
. Assume this is not the case, and expose the
edges spanned by the neighborhood N of v. Now the number of odd neighbors of v is determined
by the edges between N and N = V (N v). In fact, the probability that a vertex u N is
connected to an odd number of vertices in N is p
1
=
1(12p)
nd1
2

1
2
, where d = |N| < n/2 is the
degree of v, so u has an odd degree in G with probability p
1
or (1 p
1
). Thus the number of odd
neighbors of v is a sum of d indicator random variables with probabilities p
1
or (1 p
1
). Another
application of Claim 2.1 then shows that the probability that v has fewer than np/5 < dp
1
/2 odd
neighbors (conditioned on np/2 d 2np) is at most e
dp
1
/20
< e
np/50
.
Summarizing the previous paragraph, the probability that v has fewer than np/5 odd neighbors
is at most
2e
np/20
+e
np/50
3e
np/50
e
3 log log n
=
1
log
3
n
for large enough n, establishing the sixth property.
Now we are ready to prove the following lemma, which will serve as a tool to get rid of odd
degrees. Its exibility the fact that we can apply it to any set S will prove useful when tackling
the dense case.
Lemma 2.4. Let p =
_
log n
n
_
but p
log
10
n
n
. Then whp in G = G(n, p) for any vertex set S of
even cardinality, there is a collection E
0
of
|S|
2
+ o(n) edges in G such that S is exactly the set of
vertices incident to an odd number of edges in E
0
.
Proof. Take any set S. First, we nd a matching in S
0
= S greedily, by selecting one edge e
i
at a time
spanned by S
i
, and then removing the vertices of e
i
from S
i
to get S
i+1
. At the end of the process,
the remaining set of vertices S

S is independent in G. The second property from Lemma 2.3

implies that |S

|
2 log(np)
p
. Now let us pair up the vertices in S

arbitrarily, and for each of these

pairs {v
j,1
, v
j,2
} take a shortest path P
j
in G connecting them. (Recall that for p =
_
log n
n
_
the
random graph G(n, p) is whp connected.) The third property ensures that each of the P
j
contains
at most
2 log n
log(np)
edges. Note that we do not assume these paths to be edge-disjoint and they may
contain the e
i
s as well.
5
Let us dene E
0
to be the mod 2 union of the P
j
and the e
i
, i.e., we include an edge e in E
0
if
it appears an odd number of times among them. Then the set of odd-degree vertices in E
0
is indeed
S, and the number of edges is at most
|S|
2
+
2 log(np)
p

2 log n
log(np)
=
|S|
2
+o(n).
We can actually push down the probability p a bit by giving up on the above mentioned exibility.
The following lemma takes care of the odd-degree vertices and the small components in the sparse
case.
Lemma 2.5. Let p =
_
log log n
n
_
, but p
log
10
n
n
, and let S be the set of odd vertices in G = G(n, p).
Then whp there is a collection E
0
of
|S|
2
+o(n) edges in G such that GE
0
is an Euler graph.
Proof. Let us assume that G satises all the properties from Lemma 2.3. Our plan is to show that
there is a matching in the induced subgraph G
S
= G[S] covering all but at most n/ log
2
n vertices
in S, using the defect version of Tuttes theorem on G
S
. For this we need that the deletion of any
set T of t vertices from G
S
creates no more than t + n/ log
2
n odd-vertex components. In fact, we
will prove that the deletion of t vertices breaks G
S
into at most t +n/ log
2
n components.
If t
n
100
then this easily follows from the second property: if at least t components were created,
then we could nd an independent set in G of size t simply by picking a vertex from each component.
But (G)
2 log(np)
p
<
n
100
, so this is impossible.
Now suppose there is a set T of t <
n
100
vertices such that G
S
T has at least t+
n
log
2
n
components.
Here the number of components containing at least 2 log
2
n vertices is clearly at most
n
2 log
2
n
. On
the other hand, according to the sixth property, there are at most
n
2 log
2
n
components containing a
vertex of degree less than np/5 in G
S
. Thus there are t components of size at most 2 log
2
n, and
hence of average degree at most 4 by the fourth property, such that all the vertices contained in them
have at least np/5 neighbors in G
S
. Each such component then has a vertex with at least np/5 4
neighbors in T. Pick a vertex like that from t of these components to form the set T

.
We see that there are at least t(np/5 4) > tnp/6 edges going between T and T

, but this
contradicts the fourth property if t n
1/10
and the fth property if n
1/10
< t <
n
100
. So by Tuttes
theorem, we can nd some edge set M that forms a matching in S, covering all but
n
log
2
n
of its
vertices.
Now let F be the set of edges appearing in the small components of G (recall that in this
probability range, G(n, p) whp has a giant component and some smaller ones). According to the rst
property, these edges span trees on o(n) vertices, so |F| = o(n). We see that the set M F takes
care of all odd vertices outside the giant component and all but at most
n
log
2
n
of them inside. The
rest of the proof follows the idea from the previous lemma: we pair up the remaining odd vertices
arbitrarily and take shortest paths P
j
in G connecting them. Once again, the third property implies
that each path has length at most
2 log n
log(np)
. Then E
0
, the mod 2 union of the P
j
and M F satises
all requirements and uses at most
|S|
2
+o(n) +
n
log
2
n

2 log n
log np
=
|S|
2
+o(n)
6
edges.
3 Cycle decompositions in sparse random graphs
In this section we show that Euler subgraphs of sparse random graphs can be decomposed into
o(n) edge-disjoint cycles. The following statement from [6], which we use to nd long cycles, is an
immediate consequence of applying P osas rotation-extension technique [10] as described in [3].
Lemma 3.1. If a graph G does not contain any cycle of length at least 3t, then there is a set T of
at most t vertices such that |N(T)| 2|T|.
We say that a graph G is suciently sparse if any set of vertices S spans less than r|S| edges,
where r = max{
|S|
12 log
2
n
, 7}. Note that any subgraph of a suciently sparse graph is also suciently
sparse.
In what follows, we proceed by showing that on the one hand, for p small enough the graph
G(n, p) is typically suciently sparse, while on the other hand, any suciently sparse graph contains
few edge-disjoint cycles covering most of the edges.
Lemma 3.2. Let p = p(n) < n
1/6
and let G = G(n, p). Then whp G is suciently sparse.
Proof. For xed s, the probability that there is an S of size s containing at least rs edges is at most
_
n
s
_

__
s
2
_
rs
_
p
rs
n
s

_
esp
2r
_
rs
=
_
n
_
esp
2r
_
r
_
s
Put x = n
_
esp
2r
_
r
. If we further assume that r
s
12 log
2
n
and r 7 then we get that for large n and
p < n
1/6
x n (6ep log
2
n)
r
n (6e)
7
_
log
2
n
n
1/6
_
7
= o(1),
where we used the fact that 6ep log
2
n < 1 for large n. Hence the probability that for some s there
is a set S of size s which spans more than max{
s
12 log
2
n
, 7} s edges, i.e., that G is not suciently
sparse, is at most
n

i=2
x
i
<
x
2
1 x
< x = o(1).
Corollary 3.3. For p < n
1/6
, whp all subgraphs of G(n, p) are suciently sparse.
Using these lemmas we can derive one of our main tools, the fact that we can make a subgraph
of a random graph relatively sparse by iteratively removing long cycles.
Proposition 3.4. Let H be a suciently sparse n-vertex graph of average degree d > 84. Then it
contains a cycle of length at least d log
2
n.
7
Proof. Suppose not. Dene H G to be the d/2-core of G, i.e., the non-empty subgraph obtained by
repeatedly removing vertices of degree less than d/2. Since there is no cycle of length at least d log
2
n
in H, we can apply Lemma 3.1 to nd a set T of t d log
2
n/3 vertices such that |N(T)| 2|T| = 2t.
Let S = T N(T) and s = |S|, then the minimum degree condition implies that there are at least
dt/4 edges incident to T, hence the set S spans at least sd/12 edges.
Note that r = d/12 > 7. Also, since s 3t d log
2
n, we have r s/12 log
2
n. So the set S spans
at least max{
s
12 log
2
n
, 7} |S| edges, contradicting the assumption that H is suciently sparse.
Corollary 3.5. Let p < n
1/6
. Then whp G G(n, p) has the following property. If H is a subgraph
of G on n
0
vertices of average degree d, then H can be decomposed into a union of at most 2n
0
/ log n
0
cycles and a graph H

of average degree at most 84.

Proof. By Corollary 3.3 all subgraphs of G(n, p) are suciently sparse with high probability. So we
can repeatedly apply Proposition 3.4 to remove cycles from H as follows. Let H
0
= H. As long
as the graph H
i
has average degree d
i
> 84, we nd a cycle C
i+1
in it of length at least d
i
log
2
n
0
and dene H
i+1
= H
i
E(C
i+1
). After some nite number (say l) of steps we end up with a graph
H

= H
l
of average degree at most 84. We need to bound l.
Going from H
i
of average degree d
i
to H
j
, the rst graph of average degree below d
i
/2, we remove
at most d
i
n
0
/2 edges using cycles of length at least
d
i
2
log
2
n
0
. So the number of cycles we removed
is at most n
0
/ log
2
n
0
. Thus if H had average degree d then l n
0
log
2
d/ log
2
n
0
2n
0
/ log n
0
, as
needed.
To conclude this section, we prove Theorem 1.1 in the range
_
log log n
n
_
p
log
10
n
n
by observing
that whp G(n, p) contains no more than o(n) short cycles.
Lemma 3.6. Let p <
log
10
n
n
. Then with high probability G(n, p) contains no more than

n cycles
of length at most log log n.
Proof. Let X
k
be the number of cycles of length k in G(n, p). The number of cycles in K
n
of length k
is at most n
k
, and each cycle has probability p
k
of being included in G(n, p), hence E(X
k
) (np)
k

(log
10
n)
k
. So if X =

log log n
k=3
X
k
is the number of cycles of length at most log log n, then we clearly
have
E(X)
log log n

k=3
E(X
k
)
log log n

k=3
(log
10
n)
k
(log
10
n)
2 log log n
using log
10
n > 2.
Hence we can apply Markovs inequality to bound the probability that there are more than

n
short cycles:
P(X

n)
(log
10
n)
2 log log n

n
exp(20(log log n)
2
(log n)/2) = o(1).
Corollary 3.7. Let p <
log
10
n
n
. Then whp any Euler subgraph H of G = G(n, p) can be decomposed
into o(n) cycles.
8
Proof. Use Corollary 3.5 to remove o(n) cycles and end up with a graph H
1
of average degree at
most 84. Note that H
1
is still an Euler graph, so we can break the edges of G
1
into cycles arbitrarily.
We claim that the number of cycles we get is o(n). Indeed, by Lemma 3.6, whp there are at most

n = o(n) short cycles, i.e., of length at most log log n, while the number of long cycles can be
bounded by the number of edges divided by log log n. Since H
1
contains a linear number of edges,
the number of long cycles is O(
n
log log n
) = o(n) and we are done.
Our theorem for small p is then an immediate consequence of Lemma 2.5 and Corollary 3.7.
Theorem 3.8. Let
_
log log n
n
_
< p <
log
10
n
n
. Then whp G = G(n, p) can be decomposed into
odd(G)
2
+o(n) cycles and edges.
4 The main ingredients for the dense case
For larger p we use a strong theorem by Broder, Frieze, Suen and Upfal [4] about the existence
of edge-disjoint paths in random graphs connecting a prescribed set of vertex pairs. We need the
following denition to state it: Suppose S is a set of vertices in a graph G. We dene the maximum
neighborhood-ratio function r
G
(S) to be max
vV (G)
|N
G
(v)S|
|N
G
(v)|
.
Theorem 4.1 (Broder-Frieze-Suen-Upfal). Let p =
_
log n
n
_
. Then there are two constants , > 0
such that with probability at least 1
1
n
the following holds in G(n, p). For any set F = {(a
i
, b
i
)|a
i
, b
i

V, i = 1, ..., k} of at most
nlog(np)
log n
disjoint pairs in G = G(n, p) satisfying the property below, there
are vertex-disjoint paths connecting a
i
to b
i
:
There is no vertex v which has more than a -fraction of its neighborhood covered by the vertices
in F. In other words, r
G
(S) , where S is the set of vertices appearing in F.
We shall use the following statement to establish this property, so that we can apply Theorem 4.1
in our coming proofs.
Lemma 4.2. Let M be a matching covering some vertices of V , and let G G(n,
log
3
n
n
) be a random
graph on the same vertex set, where G and M are not necessarily independent. Then with probability
1n
(1)
we can break G into log n random graphs G
i
G(n,
log
2
n
n
) and M into log n submatchings
M
i
of at most
n
log n
edges each, such that r
G
i
(S
i
)
4

log n
for all i, where S
i
is the set of endvertices
of M
i
.
Proof. First, we partition the edge set of G into log n graphs G
1
, . . . , G
log n
by choosing an index i
e
for each edge e uniformly and independently at random, and placing e in G
ie
. Then each G
i
has
distribution G(n,
log
2
n
n
). Then we can apply Claim 2.1(b) and (d) to the degree of each vertex of
each of the G
i
s, and use the union bound to see that all the G
i
have minimum degree at least
log
2
n
2
and maximum degree at most 2 log
2
n with probability 1 2nlog n e
log
2
n/40
= 1 n
(1)
.
Now break the M into log n random matchings M
1
, . . . , M
log n
similarly, by placing each edge
f M independently in M
i
f
where i
f
is a random index chosen uniformly. Since there are at most
9
n/2 edges in M, another application of Claim 2.1(b) gives that with probability 1log ne
n/20 log n
=
1 n
(1)
each M
i
contains at most
n
log n
edges.
If G
i
has maximum degree at most 2 log
2
n, then the neighborhood of an arbitrary vertex v in
G
i
may meet at most 2 log
2
n edges from M. The probability that at least (log n)
3/2
of them are
selected in M
i
is at most
_
2 log
2
n
log
3/2
n
_
1
(log n)
log
3/2
n

_
2e log
2
n
log
3/2
log n
_
log
3/2
n

_
2e
log
1/2
n
_
log
3/2
n
e
log nlog log n
.
So taking the union bound over all vertices v and indices i gives that with probability 1 n
(1)
,
the neighborhood of any v in any G
i
meets at most log
3/2
n of the edges in M
i
, and thus it contains
at most 2 log
3/2
n vertices from S
i
. Since all neighborhoods have size at least
log
2
n
2
, we get that
r
G
i
(S
i
)
4

log n
for all i.
The next theorem is the main tool on our way to the proof of Theorem 1.1.
Theorem 4.3. Let
2 log
5
n
n
p n
1/6
and G = G(n, p). Suppose G is randomly split into G

and
G

by placing its edges independently into G

with probability p

=
2 log
5
n
np
and into G

otherwise.
Then whp for any subgraph H of G

such that G

H is Euler, G

H can be decomposed into o(n)

cycles.
Proof. Note that, although G

and G

are far from being independent, G

on its own has distribution

G(n, p

p) = G(n,
2 log
5
n
n
) and G

has distribution G(n, (1 p

)p) where (1 p

)p < n
1/6
. It is easy
to see from Claim 2.1 that whp G has maximum degree at most 2n
5/6
.
By Corollary 3.5, whp any H G

can be decomposed into o(n) cycles and a graph H

0
of average
degree at most 84. Therefore it is enough for us to show that whp G

satises the following: for any

H
0
containing at most 42n edges such that G

H
0
is Euler, G

H
0
is an edge-disjoint union of
o(n) cycles. Our plan is to break H
0
into few matchings and then to use Theorem 4.1 on random
subgraphs of G

to connect them into cycles.

So dene V
0
V to be the set of vertices of degree at least
log
2
n
2
in H
0
, and let V
1
= V V
0
be the rest. Note that |V
0
| = O(
n
log
2
n
). We break into matchings in two rounds: rst we take care
of the edges spanned by V
1
, then the ones crossing between V
0
and V
1
. Let us split G

into two
random graphs G
1
, G
2
G(n,
log
5
n
n
) by placing each edge of G

independently in one of them with

probability 1/2.
Now look at the subgraph of H
0
spanned by V
1
: the maximum degree is at most
log
2
n
2
1, so
we can break the edge set into log
2
n matchings M
1
, . . . , M
log
2
n
. To nd the cycles, we also split
G
1
into log
2
n parts G
1,1
, . . . , G
1,log
2
n
so that each G
1,i
has distribution G(n,
log
3
n
n
). We can use
Lemma 4.2 to further break each M
i
and G
1,i
into log n parts M
i,j
and G
1,i,j
, respectively, so that
whp each M
i,j
contains O(n/ log n) edges, and the endvertices of M
i,j
only cover a o(1)-fraction of
the neighborhood of any vertex in G
1,i,j
.
We create the cycles as follows: if M
i,j
consists of the edges v
1
v

1
, . . . , v
k
v

k
, then we choose the
corresponding set of pairs to be F
1,i,j
= {(v

1
, v
2
), (v

2
, v
3
), . . . , (v

k
, v
1
)}. Here the above properties
of M
i,j
and G
1,i,j
ensure that we can apply Theorem 4.1, and with probability at least 1
1
n
the
10
matching can be closed into a cycle. Hence with probability 1
log
3
n
n
= 1 o(1) all the M
i,j
s can
be covered by log
3
n cycles altogether.
Let us turn to the edges of H
0
between V
0
and V
1
, and dene the following auxiliary multigraph
G
a
on V
1
: for each v V
0
, pair up its neighbors in V
1
(except maybe one vertex, if the neighborhood
has odd size) and let E
v
be the set of edges a matching corresponding to this pairing. Dene
the edge set of G
a
to be the disjoint union of the E
v
for v V
0
. The idea is that an edge ww

E
v
corresponds to the path wvw

, so we want to nd cycles covering the edges in G

a
and then lead them
through the original V
0
V
1
edges.
By the denition of V
1
, the maximum degree in G
a
is at most
log
2
n
2
1, so the edge set
vV
0
E
v
can be split into log
2
n matchings N
1
, . . . , N
log
2
n
. Now it is time to break G
2
into log
2
n random
subgraphs G
2,i
of distribution G(n,
log
3
n
n
) each. Once again, we use Lemma 4.2 to prepare for the
cycle cover by splitting each of the N
i
and G
2,i
into log n parts N
i,j
and G
2,i,j
. When we dene the
set of pairs F
2,i,j
, we need to be a little bit careful: we must make sure that no cycle contains more
than one edge from any given E
v
. This way the cycles do not become self-intersecting after switching
the edges from E
v
back to the corresponding paths through v. Since the maximum degree of G, and
hence the cardinality of E
v
, is at most 2n
5/6
, we may achieve this by using at most 2n
5/6
cycles per
matching. Indeed, split N
i,j
into at most 2n
5/6
subsets N
i,j,k
so that none of the N
i,j,k
contains more
than one edge from the same E
v
(this can be done greedily). Then dene the sets of pairs F
2,i,j,k
for
k = 1, . . . , 2n
5/6
to close N
i,j,k
into a cycle the same way as before, and take F
2,i,j
=
2n
5/6
k=1
F
2,i,j,k
.
As above, all conditions of Theorem 4.1 are satised when we use G
2,i,j
to nd the paths cor-
responding to F
2,i,j
that close N
i,j
into cycles, and since the error probabilities were all O(
1
n
), with
high probability we can simultaneously do so for all i and j. We have log
3
n matchings, so in total
we get 2n
5/6
log
3
n = o(n) edge-disjoint cycles that cover all but o(n) edges of H
0
between V
0
and
V
1
(missing at most one incident edge for each v V
0
).
Finally, we apply Corollary 3.5 on the subgraph of H
0
induced by V
0
to see that the edges
spanned by V
0
can be partitioned into O(n/ log
3
n) cycles and O(n/ log
2
n) = o(n) edges (recall that
|V
0
| = O(n/ log
2
n)).
So far we have found o(n) edge-disjoint cycles in G

H. Once we remove them, we get an Euler

graph containing only o(n) edges from H. So we can nd o(n) edge-disjoint cycles covering all of them
and remove these cycles, as well, to get an Euler subgraph of G

G(n,
log
5
n
n
). Now Corollary 3.7
shows that we can partition the remaining graph into o(n) cycles, concluding our proof.
5 Cycle-edge decompositions in dense random graphs
At last, we are ready to prove Theorem 1.1 in the denser settings. The case p n
1/6
is fairly
straightforward from our previous results, we just need to be a little bit careful.
Theorem 5.1. Let
log
6
n
n
p n
1/6
. Then whp G = G(n, p) can be decomposed into
odd(G)
2
+o(n)
cycles and edges.
Proof. We split G into the union of three disjoint random graphs G
1
, G
2
and G
3
by putting each
edge e E(G) independently into one of the graphs. With probability p
1
=
2 log
5
n
np
we place e into
11
G
1
, with probability p
2
=
log
2
n
np
we place it into G
2
, and with probability 1 p
1
p
2
we place it into
G
3
. This way G
1
G(n,
2 log
5
n
n
) and G
2
G(n,
log
2
n
n
).
Now let S be the set of odd-degree vertices in G. Applying Lemma 2.4 to S in G
2
gives a set E
0
of |S|/2 +o(n) edges in G
2
such that GE
0
is Euler. Taking H to be the subgraph G
2
G
3
E
0
of
G

= G
2
G
3
and setting G

= G
1
, we can apply Theorem 4.3 to split the edge set of G E
0
into
o(n) cycles. The theorem follows.
To get a tight result for larger p, we must remove cycles containing nearly all vertices. A recent
result by Knox, K uhn and Osthus [9] helps us to nd many edge-disjoint Hamilton cycles in random
graphs.
Theorem 5.2 (Knox-K uhn-Osthus). Let
log
50
n
n
p 1 n
1/5
and G = G(n, p). Then whp G
contains (G)/2 edge-disjoint Hamilton cycles.
Let us point out that we do not actually need such a strong result: (G)/2n

disjoint Hamilton
cycles would also suce for our purposes.
Theorem 5.3. Let p n
1/6
. Then whp G = G(n, p) can be decomposed into
odd(G)
2
+
np
2
+ o(n)
edges.
Proof. Similarly to Theorem 5.1, we partition G into the union of four disjoint random graphs
G
1
, G
2
, G
3
and G
4
by assigning each edge of G independently to G
1
with probability p
1
=
2 log
5
n
np
,
to G
2
with probability p
2
=
n
4/5
log
3
n
np
, to G
3
with probability p
3
=
log
2
n
np
and to G
4
otherwise (with
probability p
4
= 1p
1
p
2
p
3
= 1o(1)). It is easy to see from the Cherno bound (Claim 2.1(a))
that with high probability the maximum degree of G
3
is at most npp
3
+ n
3/5
2n
3/5
, and the
maximum degree of G
4
is at most npp
4
+n
3/5
. Let us assume this is the case.
Let S be the set of odd-degree vertices in G. Now as before, we use Lemma 2.4 to nd a set E
0
of
|S|
2
+o(n) edges in G
3
such that GE
0
is Euler. Notice that G
4
G(n, pp
4
), where pp
4
= p(1o(1)),
so whp G
4
has minimum degree at least npp
4
n
3/5
by Claim 2.1(c). Also, pp
4
< 1n
1/5
(because
pp
2
> n
1/5
), hence we can apply Theorem 5.2 and nd
npp
4
n
3/5
2
edge-disjoint Hamilton cycles in
it. Let H
0
be the graph obtained from G
3
G
4
E
0
by removing these cycles. Then the maximum
degree of H
0
is at most 4n
3/5
, hence we can break the edge set of H
0
into n
4/5
matchings M
i
.
We want to use Theorem 4.1 to close the M
i
s into cycles, so let us split G
2
into n
4/5
random
graphs G

i
G(n,
log
3
n
n
) uniformly. By Lemma 4.2 we can further partition each M
i
and G

i
, with
probability 1 n
(1)
, into log n matchings and graphs, M
i,j
and G

i,j
in such a way that we can
apply Theorem 4.1 on G

i,j
for any pairing F
i,j
on the vertices of M
i,j
. Choose F
i,j
, as before, so
that the resulting paths together with M
i,j
form a cycle. Then with probability at least 1
1
n
the
theorem produces the required cycle, so whp all n
4/5
log n cycles exist simultaneously.
This way we nd n
4/5
log n = o(n) edge-disjoint cycles covering H
0
and some edges in G
2
. Let H
be the graph containing the unused edges of G
2
. Then G
1
H is Euler, and we can apply Theorem 4.3
with the host graph G
1
G
2
from distribution G(n, p(p
1
+p
2
)), and the partition G

= G
1
, G

= G
2
.
This gives us a decomposition of G

H into o(n) cycles whp, completing our proof.

12
6 Concluding remarks
The above proof settles the question for p =
_
log log n
n
_
, but it would be nice to have a result for the
whole probability range. The bottleneck in our proof is Lemma 2.5, where we obtain a small edge
set E
0
such that G(n, p) E
0
is Euler. We believe that similar ideas can be applied to prove this
lemma for even smaller values of p if one puts more eort into nding short paths between vertices
not covered by the matching. In any case, it seems that the asymptotics of the optimum is dened
by the smallest such E
0
for any p log n/n, so a complete solution to the problem would rst need
to describe this minimum in the whole range.
References
[1] N. Alon and J. H. Spencer, The probabilistic method, 3rd ed., Wiley, 2008.
[2] B. Bollobas, Random graphs, 2nd ed., Cambridge Stud. Adv. Math. 73, Cambridge University
Press, 2001.
[3] S. Brandt, H. Broersma, R. Diestel and M. Kriesell, Global connectivity and expansion: long
cycles and factors in f-connected graphs, Combinatorica 26 (2006), 1736.
[4] A. Z. Broder, A. M. Frieze, S. Suen and E. Upfal, An ecient algorithm for the vertex-disjoint
paths problem in random graphs, Proceedings of SODA 96, pp. 261268.
[5] F. Chung and L. Lu, The diameter of sparse random graphs, Advances in Applied Math. 26
(2001), 257279.
[6] D. Conlon, J. Fox and B. Sudakov, Cycle packing, submitted
[7] P. Erdos, On some of my conjectures in number theory and combinatorics, Proceedings of
the fourteenth Southeastern conference on combinatorics, graph theory and computing (Boca
Raton, Fla., 1983), Congr. Numer. 39 (1983), 319.
[8] P. Erdos, A. W. Goodman and L. P osa, The representation of a graph by set intersections,
Canad. J. Math. 18 (1966), 106112.
[9] F. Knox, D. K uhn and D. Osthus, Edge-disjoint Hamilton cycles in random graphs, Random
Structures and Algorithms, to appear.
[10] L. P osa, Hamiltonian circuits in random graphs, Discrete Math. 14 (1976), 359364.
[11] L. Pyber, An Erdos-Gallai conjecture, Combinatorica 5 (1985), 6779.
13