Topics in random graphs

Richard Montgomery
Lent 2018

Abstract
These notes accompanied a graduate course given in Cambridge in the Lent term of
2018. The course aimed to cover a range of different techniques relevant to random graphs.
Any undefined notation is standard and can be found in, for example, [7, 13, 16]. Any
comments or corrections are welcome.

Contents
1 Introduction 2

2 Thresholds 2

3 Sharp Thresholds 3

4 Long paths and cycles 4

4.1 Long paths in G(n, Cn ) I: Depth First Search . . . . . . . . . . . . . . . . . . . . . 4
4.2 Long paths in G(n, Cn ) II: Pósa rotation . . . . . . . . . . . . . . . . . . . . . . . 6
C
4.3 Long paths in G(n, n ) III: Revealing edges as needed . . . . . . . . . . . . . . . 8
4.4 Long paths in G(n, Cn ) IV: Pósa rotation and extension, and edge sprinkling . . . 10
4.5 Long paths in G(n, Cn ) V: Conditional existence of boosters . . . . . . . . . . . . 15
4.6 Hamiltonicity in the random graph process . . . . . . . . . . . . . . . . . . . . . 17

5 Random Directed Graphs 20

6 The chromatic number of dense random graphs 22

6.1 Diversion: Upper and lower tails . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.2 The chromatic number of G(n, p). . . . . . . . . . . . . . . . . . . . . . . . . . . 23

7 Squared Hamilton cycles in random graphs 24

7.1 Overview of the proof of Theorem 44 . . . . . . . . . . . . . . . . . . . . . . . . . 25
7.2 Finding connectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
7.3 Finding a covering connector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7.4 Diversion: Realistic ambitions for constructing absorbers . . . . . . . . . . . . . . 29
7.5 Constructing absorbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

1
1 Introduction
Consider graphs G which are finite and simple, with vertex set V (G) and edge set E(G), and
write |G| = |V (G)| and e(G) = |E(G)|. The study of random graphs began in full with the work
of Erdős and Rényi [10]. We will study the following random graph model.
Definition. The binomial random graph or Erdős-Rényi random graph, G = G(n, p), has
V (G) = [n] = {1, . . . , n} and each possible edge is included independently at random with
probability p.
Definition. A graph property P is a collection of graphs.
Example. . P1 = {G : e(G) > 0} P4 = {G : χ(G) > 3}
P2 = {G : G contains a triangle} P5 = {G : χ(G) 6= 3}
P3 = {G : e(G) is odd}

Definition. P is non-trivial if every G ∈ P has e(G) > 0 and, for sufficiently large n, ∃G ∈ P
with V (G) = [n].
Definition. P is monotone if G ∈ P, G ⊂ H and V (H) = V (G) =⇒ H ∈ P. I.e., we cannot
leave P by adding edges.
Example. P1 , P2 , P4 above are monotone, while P3 , P5 are not.
Key meta-question. Is G(n, p) typically in P?
Definition. When p = p(n), say G(n, p) ∈ P almost surely (a.s.) or with high probability
(w.h.p.) if P(G(n, p) ∈ P) → 1 as n → ∞.
Asymptotic notation: For f, g non-zero functions of n, if f (n)/g(n) → ∞ as n → ∞, then
we say f = ω(g) and g = o(f ). If ∃C > 0 such that, for all n, |f (n)| ≤ C|g(n)|, then we say
f = O(g) and g = Ω(f ). Where the implicit constant C depends on another variable, we will
indicate this in the subscript.

2 Thresholds
Definition. A function p̂ is a threshold for a property P if
• p = o(p̂) =⇒ G(n, p) ∈
/ P a.s., and
• p = ω(p̂) =⇒ G(n, p) ∈ P a.s.
Exercise. If p̂0 , p̂1 are thresholds for P then ∃c, C > 0 such that cp̂0 < p̂1 < C p̂0 .
Proposition 1. The function p̂ = 1/n2 is a threshold for P1 = {G : e(G) > 0}.
n n
Proof. If p = ω(p̂), then P(G(n, p) ∈/
P) = (1 − p)( 2 ) ≤ e−p( 2 ) = o(1). On the other hand, if
p = o(p̂), then P(G(n, p) ∈ P) ≤ p n2 = o(1).
The following lemma was first proved by Bollobás and Thomason [8].
Lemma 2. All non-trivial properties have a threshold.
Idea. We can couple probability models, i.e. ‘reveal edges in batches’.

2
Proposition 3. If G0 ∼ G(n, p) and G1 ∼ G(n, q) are independent, then G0 ∪ G1 ∼ G(n, p +
q − pq).
Proof. The events {e ∈ E(G0 ∪ G1 )}, e ∈ [n](2) , are independent and, ∀e ∈ [n](2) ,

P(e ∈ E(G0 ∪ G1 )) = 1 − P(e ∈

/ E(G0 )) · P(e ∈
/ E(G1 )) = p + q − pq.

Corollary 4. If P is monotone, then ∀p, q ∈ [0, 1],

P(G(n, p + q) ∈
/ P) ≤ P(G(n, p) ∈
/ P) · P(G(n, q) ∈
/ P).

Proof of Lemma 2. For each (sufficiently large) n, let p̂(n) be such that P(G(n, p̂) ∈ P) = 21 .
Note that, as G ∈ P =⇒ e(G) > 0, P(G(n, 0) ∈ P) = 0. As there is some G ∈ P with
V (G) = [n], and P is monotone, the complete graph on [n] is in P. Thus, P(G(n, 1) ∈ P) = 1,
and such a p̂ thus exists.
If p = ω(p̂), then, applying Corollary 4 bp/p̂c times, we have

bp/p̂c  1 bp/p̂c
P(G(n, p) ∈
/ P) ≤ P(G(n, p̂) ∈
/P = → 0 as n → ∞.
2
If p = o(p̂), then, similarly,
1
bp̂/pc
= P(G(n, p̂) ∈
/ P) ≤ P(G(n, p) ∈
/ P) .
2
Thus, as p̂ = ω(p), P(G(n, p) ∈
/ P) = 1 − o(1).
Exercise. Prove that p̂ = 1/n is a threshold for G(n, p) to contain a triangle.

3 Sharp Thresholds
Definition. A function p̂ is a sharp threshold for a property P if, ∀ε > 0, as n → ∞,
• P(G(n, (1 − ε)p̂) ∈ P) → 0, and
• P(G(n, (1 + ε)p̂) ∈ P) → 1.
Exercise. If p̂0 , p̂1 are sharp thresholds for P, then p̂0 − p̂1 → 0.
Definition. A threshold for P is coarse if P has no sharp threshold.
Example. P1 = {G : e(G) > 0} has a coarse threshold. Indeed, if p = C/n2 , with C constant,
n
then P(G(n, p) ∈ P) = 1 − (1 − p)( 2 ) → 1 − e−C/2 as n → ∞. Thus, if p̂ was a sharp threshold
for P1 , then, as G(n, 2p̂) is almost surely in P, we have p̂ = ω(1/n), and as G(n, p̂/2) is almost
surely not in P, we have p̂ = o(1/n), a contradiction.
Exercise. P = {G : K3 ⊂ G} has a coarse threshold.
Rule of thumb. Properties that are ‘locally checkable’ have a coarse threshold. E.g. that G
contains a triangle can be confirmed with only three edges, but to confirm that δ(G) > 0
requires at least n/2 edges. This notion was formalised by Friedgut [12].
Example. P = {G : δ(G) ≥ 2} has a (very!) sharp threshold, as follows.
log n+log log n+ω(1)
Lemma 5. a) If p = n , then P(δ(G(n, p)) ≥ 2) → 1 as n → ∞.

3
 log n+log log n−ω(1)
b) If p = n , then P(δ(G(n, p)) ≤ 1) → 1 as n → ∞.
p
Exercise. Show that, for each p ∈ (0, 1), e− 1−p ≤ 1 − p ≤ e−p .
Proof of Lemma 5. Note that we can assume that log n
2n ≤ p ≤
2 log n
n . Let X be the number of
vertices with degree at most 1 in G = G(n, p). Then,

EX = n((1 − p)n−1 + (n − 1)p(1 − p)n−2 ) = (1 + o(1))n2 p(1 − p)n . (1)

2 log n
a) First moment method. From (1), and as p ≤ n ,

EX ≤ (2 + o(1))n log n · e−pn = o(1).

Thus, by Markov’s inequality, P(X ≥ 1) ≤ EX = o(1).

log n
b) Second moment method. From (1), and as p ≥ 2n , we have
1  1  pn
EX ≥ + o(1) n log n · (1 − p)n ≥ + o(1) n log n · e 1−p = ω(1).
2 2
By Markov’s inequality (or Chebychev’s inequality),

E((X − EX)2 ) E(X 2 )

P(X = 0) ≤ P (X − EX)2 ≥ (EX)2 ≤ 2
= − 1.
(EX) (EX)2

Then,
X
E(X 2 ) = EX + P(d(u), d(v) ≤ 1)
u6=v

≤ EX + n2 ((1 − p)2n−3 + (2n − 3)p(1 − p)2n−4 + (n − 2)2 p2 (1 − p)2n−5 )

= EX + (1 + o(1))n4 p2 (1 − p)2n = EX + (1 + o(1)) · (EX)2 .

Thus, P(X = 0) ≤ 1/EX + (1 + o(1)) − 1 = o(1).

4 Long paths and cycles

We will study how long a path we can typically find in G(n, C n ), with C a large constant. In
order to cover different techniques, we will prove 5 increasingly good bounds on the size of a
typical longest path in G(n, C
n ). We will then show that if we start with the empty graph on n
vertices and add random edges one-by-one, then the first graph with minimum degree 2 almost
surely contains a Hamilton cycle (here, one with n vertices).
Definition. For each u ∈ V (G), N (u) = {v : uv ∈ E(G)}, and, for each U ⊂ V (G), N (U ) =
(∪u∈U N (u)) \ U .

4.1 Long paths in G(n, Cn ) I: Depth First Search

log C
Theorem 6. G(n, C


n ) almost surely has a path with length at least 1 − OC ( C ) n.

4
 The following nice analysis of the Depth First Search algorithm and its use to find long paths
(as part of a more general result) is due to Krivelevich, Lee and Sudakov [19]. The result was
originally shown by Ajtai, Komlós and Szemerédi [1] and de la Vega [9].
Depth First Search (DFS) algorithm. Run the following algorithm on a graph G, governed
by the following updating variables:
U – the set of unexplored vertices,
D – the set of dead vertices,
r – the number of active vertices, and
A = (a1 , . . . , ar ) – the active vertices, which form a path in that order.

Start: Pick some v ∈ V (G), and let

U = V (G) \ {v}, D = ∅, r = 1, and A = (a1 ) = (v).

Step 1:
If possible, pick w ∈ N (ar ) ∩ U , update

ar+1 → w, A → (a1 , . . . , ar , ar+1 ), U → U \ {w}, and r → r + 1,

and repeat Step 1.

If not possible, go to Step 2.
Step 2:
The neighbours of ar are fully explored. Update

D → D ∪ {ar }, A → (a1 , . . . , ar−1 ), and r → r − 1.

If r ≥ 1, go to Step 1.
Otherwise (r = 0), go to Step 3.
Step 3:
If U 6= ∅, pick v ∈ U , update

r → 1, a1 → v, U → U \ {v}, and A → (a1 ),

and go to Step 1.
Otherwise, stop the algorithm.

Observe: a) There are never any edges from D to U .

b) After each step, r + |U | + |D| = |G|.
c) During each step, either |U | decreases by 1 and |D| is unchanged, or |D| increases by 1
and |U | is unchanged.
d) The algorithm finishes when U = ∅ and r = 0, and thus D = V (G).
Definition. Say G is m-joined if there is an edge in G between any 2 disjoint vertex sets with
size m.
Lemma 7. If G is m-joined, then it has a path with more than |G| − 2m vertices.
Proof. Run DFS on G. By c) and d) above, after some step we must have |U | = |D| = s, for
some s. By a), and as G is m-joined, s < m. Thus, by b), at this point the active vertices form
a path with more than |G| − 2m vertices.

5
Proposition 8. For sufficiently large constant C, G(n, C
n ) is almost surely m-joined for each
m ≥ 3n log C/C.
C
Proof. Let m = 3n log C/C, p = n and G = G(n, p). Then, for sufficiently large C,
 2
n 2
P(∃ disjoint A, B ⊂ V (G) with |A|, |B| = m, eG (A, B) = 0) ≤ (1 − p)m
m
 en 2m 2
≤ e−pm
m
≤ (C 2 e−3 log C )m → 0.

Thus, Lemma 7 and Proposition 8 imply Theorem 6.

log C
Exercise. Prove G(n, C


n ) almost surely has a cycle with length at least 1 − OC ( C ) n.
Exercise. Changing the algorithm or the analysis, show that G(n, C
n ) almost surely has a path
with length at least (1 − OC ( C1 ))n.
Exercise. Let Gp be a random subgraph of G, with each edge chosen independently at random
with probability p. Show that, if pk → ∞ and δ(G) ≥ k, then Gp almost surely (as k → ∞) has
a path with length (1 − ok (1))k.

4.2 Long paths in G(n, Cn ) II: Pósa rotation

To improve Theorem 6, we will use Pósa rotation, first introduced by Pósa [24] when studying
the appearance of Hamilton cycles in the binomial random graph.
Theorem 9. G(n, C 1
n ) almost surely has a path with length at least (1 − OC ( C ))n.

Pósa rotation: Consider a path P , a longest path containing V (P ) in a graph G with v as one
endpoint. Let u be the other endpoint and note that all the neighbours of u lie in P .

P :
v u

If x ∈ N (u), and y is the neighbour of x furthest from v in the path P , then we can break xy
and rotate P with v fixed to get the new path P − xy + ux with endpoint y.

P − xy + ux :
v x y u

We find a new vertex y, whose neighbours must also all lie in P (by the maximality of P ). We
can then use edges from y to rotate again:

x
v y u

Definition. Let F (P, v) be the set of endpoints not equal to v which can be achieved by itera-
tively rotating P with v fixed.

6
Observe: a) The maximality of P implies that, for each u ∈ F (P, v), N (u) ⊂ V (P ).
b) If xy ∈ E(P ) is broken for a rotation, then x or y becomes an endpoint.
Now, label, for each z ∈ V (P ), the neighbours of z in P as

P :
v z− z z+ u

We then get the following claim.

Claim. If z ∈ N (F (P, v)), then either z − or z + ∈ F (P, v).
Proof. Let z ∈ N (F (P, v)) and assume z − , z + ∈
/ F (P, v). Pick y ∈ F (P, v) with zy ∈ E(G), and
let Py be a path with endpoint y reached by rotating P with v fixed. As z − , z, z + ∈/ F (P, v), by
b) above, both z − z and zz + are not broken in rotating P to get Py . Thus, either Py is

Py :
v z− z z+ u

and thus z + ∈ F (P, v), a contradiction, or Py is

Py :
v z+ z z− u

and thus z − ∈ F (P, v), which is also a contradiction.

Every vertex in N (F (P, v)) thus has a neighbour in F (P, v) in the graph P . As F (P, v) has
fewer than 2|F (P, v)| neighbours within P (using that u has only 1 neighbour in P ), we get the
following.
Lemma 10. (Pósa’s Lemma) If P , a longest path containing V (P ) in G has v as an endpoint,
then |N (F (P, v))| < 2|F (P, v)|.
Definition. G is a (d, m)-expander if, for all A ⊂ V (G) with |A| ≤ m, |N (A)| ≥ d|A|.
That is, Pósa’s lemma implies that if G is a (2, m)-expander then |F (P, v)| > m.
Lemma 11. Let m, n ∈ N satisfy m ≤ n/20. Suppose |G| = n and, for each A ⊂ V (G) with
|A| = m, |N (A)| ≥ 4n/5. Then, G has a path with length at least n − 2m.
Note. Compare this to Lemma 7, where G was required to be m-joined.
Proof of Lemma 11. We will find a large subgraph G0 ⊂ G which is a (2, n/5)-expander by
removing a largest set that does not expand. Pick B ⊂ V (G), with |N (B)| < 2|B| and |B| < m,
or B = ∅, to maximise |B|.
Claim. For all U ⊂ V (G) \ B with |U | ≤ n/5, |N (U ) \ B| ≥ 2|U |.
Proof of claim. For contradiction, say there is some U ⊂ V (G) \ B with 0 < |U | ≤ n/5 and
|N (U ) \ B| < 2|U |. Then,

|N (U ∪ B)| ≤ |N (U ) \ B| + |N (B)| < 2|U | + 2|B| = 2|U ∪ B|.

7
By the choice of B, then, |U ∪ B| ≥ m. Pick A ⊂ U ∪ B with |A| = m, so that
4n  n  2n
|N (U ∪ B)| ≥ |N (A)| − |U ∪ B| ≥ − +m ≥ + 2m ≥ 2|U ∪ B|,
5 5 5
a contradiction.
Let G0 = G − B = G[V (G) \ B], which, by the claim, is a (2, n/5)-expander. Pick a longest
path P in G0 , with an endpoint v, say. By Lemma 10, |FG0 (P, v)| > n/5. As P is a longest path
in G0 , there are no edges between FG0 (P, v) and V (G0 ) \ V (P ) = V (G) \ (B ∪ V (P )), so that, by
the condition in the lemma, |V (G) \ (B ∪ V (P ))| < m. Thus, |P | > n − |B| − m > n − 2m.
To find the condition for Lemma 11 almost surely in G(n, p), we use the following concentra-
tion inequality.
Lemma 12. (Chernoff ’s inequality – see, for example [16, Corollary 2.3]) If X is a binomial
variable and 0 < ε ≤ 3/2, then
 ε2 EX 
P(|X − EX| ≥ εEX) ≤ 2 exp − .
3
Proposition 13. For sufficiently large C, in G = G(n, C
n ) almost surely any A ⊂ V (G) with
|A| = m := 100n/C satisfies |N (A)| ≥ 9n/10.
Proof. Let p = C/n. For each A ⊂ V (G) with |A| = m, and v ∈ V (G) \ A,

/ N (A)) = (1 − p)m ≤ e−pm = e−100 .

P(v ∈

For large n, then, E|N (A)| ≥ 99n/100. As |N (A)| ∼ Bin(n − m, (1 − p)m ), by Lemma 12 with
ε = 1/20,
  1   n 
P(|N (A)| < 9n/10) ≤ P |N (A)| < 1 − E|N (A)| ≤ 2 exp − 4 .
20 10
n en


As |V (G)(m) | = m ≤ ( m )m ≤ C 100n/C ≤ exp(n/105 ), for large C, by a union bound, we have
 n   n 
P(∃A ⊂ V (G) with |A| = m and |N (A)| < 9n/10) ≤ 2 exp − 4 · exp = o(1).
10 105
Then, Lemma 11 and Proposition 13 imply Theorem 9.

4.3 Long paths in G(n, Cn ) III: Revealing edges as needed

We will now improve Theorem 9, using an ad hoc method, which demonstrates how an algorithm
can be combined with revealing whether each edge is present or not only when it is relevant to
the algorithm.
We have used Pósa rotation on a maximal path P with an endpoint v in a (2, n5 )-expander
to get a linear-sized set F (P, v) of possible endpoints. If we reveal more edges between a vertex
w∈ / V (P ) and V (P ) with probability C n , then w has an neighbour in F (P, v) with probability
 C |F (P,v)| C
1− 1− ≥ 1 − e− 5 ,
n
and using such a neighbour we could find a path with vertex set V (P ) ∪ {w}. By repeating
(something like) this, we can iteratively extend the path, suceeding in attaching each other
vertex to the path with probability 1 − exp(−Ω(C)). This will allow us to prove the following.

8
Theorem 14. For sufficiently large C, G(n, C/n) almost surely contains a path with length at
least (1 − e−C/100 )n.
Definition. A sequence of random variables X1 , X2 , . . . is a martingale (respectively, super-
martingale or submartingale) if, for each i, E|Xi | < ∞ and E(Xi |X1 , . . . , Xi−1 ) = Xi−1 (respec-
tively ≤ Xi−1 or ≥ Xi−1 ).
Theorem 15. (Azuma-Hoeffding inequality) Let (Xi )ni=0 be a submartingale with X0 constant,
and, for each 1 ≤ i ≤ n, |Xi − Xi−1 | ≤ Ci . Then, for all t > 0,

t2
 
P(Xn ≤ EXn − t) ≤ exp − n .
2 i=1 Ci2
P

Proof of Theorem 14. Consider independent random graphs G0 , G1 ∼ G(n, C/2n). By Proposi-
tion 3, it is sufficient to show that G0 ∪ G1 almost surely contains a path with length at least
(1 − e−C/100 )n.
Exercise. G0 almost surely contains a (4, n/10)-expander G00 ⊂ G0 with at least (1 − 1/100)n
vertices. (See proof of Lemma 11.)
We will run an algorithm governed by the following updating variables:
H - a large (2, n/10)-expander.
P - a path which increases in length throughout the algorithm.
D - the set of vertices we have failed to add to the path.
W - a utility set to ensure expansion is kept in H as we add vertices and edges.

Start: Let H = G00 and let P be a longest path in H, with an endpoint v, say. Let W = D = ∅.
Carry out the following stage for each 1 ≤ i ≤ n.
Stage i:
If V (P ) ∪ D = [n], then let Xi = 1.
Otherwise, pick w ∈ [n] \ (V (P ) ∪ D) and reveal edges in G1 between w and V (P ) - say
this gives w the neighbourhood A in V (P ).
If |A ∩ (FH (P, v) \ W )| ≥ 6, then let Xi = 1, and do the following.
/ V (G00 ), then add 6 vertices from A \ W to W .
• If w ∈
• Add w and the edges between w and A to H.
• Update P to be a maximal path in H containing V (P ) and w, with v an endpoint.
Otherwise, update D → D ∪ {w}, and let Xi = 0.

Observe: a) Each edge in G1 is revealed at most once.

Pn
b) At the end, |P | ≥ n − |D| = n − |{i : Xi = 0}| = i=1 Xi .
c) Always: for each U ⊂ V (H) \ V (G00 ), |NH (U )| ≥ |NH (U, W )| ≥ 6|U | − |U | = 5|U |.
d) Always: H is a (2, n/10)-expander, as follows.

Proof of d). Let U ⊂ V (H), with 1 ≤ |U | ≤ n/10, and set U1 = U ∩ V (G00 ) and U2 = U \ V (G00 ).
Either |U1 | ≥ |U |/2, whence |NH (U )| ≥ |NG00 (U1 )| ≥ 4|U1 | ≥ 2|U |,
or |U2 | ≥ |U |/2, whence |NH (U )| ≥ |NH (U2 )| − |U1 | ≥ 5|U2 | − |U2 | = 4|U2 | ≥ 2|U |.

9
 e) Always: |W | ≤ 6|V (G) \ V (G00 )| ≤ 3n/50, so, by Lemma 10 and d), |FH (P, v) \ W | ≥
n/10 − 3n/50 = n/25.
f) For each i, as at stage i we have, by e), |FH (P, v) \ W | ≥ n/25,

P(Xi = 0|X1 , . . . , Xi−1 ) ≤ P(Bin(n/25, C/2n) < 6)

5   j   n −j
X n j C C 25
≤ 1−
j=0
25 2n 2n

≤ 6C 5 eC/75
≤ e−C/100 /3,

for large C.
Pj
Now, let ε = e−C/100 /3 and, for each 1 ≤ j ≤ m, let Yj = i=1 (Xi − (1 − ε)n). As
E(Yj |Y1 , . . . , Yj−1 ) ≥ 0 for each j, (Yj )nj=1 is a submartingale. As |Yj − Yj−1 | ≤ 1, for each j, by
Theorem 15 we have
(εn)2
 
P(Yn < −εn) ≤ exp − = o(1).
2n
Pn
Note that Yn < −εn implies that ( i=1 Xi ) − (1 − ε)n < −εn. Thus, by a), at the end we almost
surely have that |P | > n − 2εn.

4.4 Long paths in G(n, Cn ) IV: Pósa rotation and extension, and edge
sprinkling
To improve Theorem 14, we will now use Pósa rotation and extension and edge sprinkling, a
method developed in the literature beginning with the pioneering work of Pósa [24].
Suppose G is a (2, n5 )-expander and P ⊂ G is a maximal length path. By rotating both ends
of P we can find Ω(n2 ) paths with different endvertex pairs. Adding such an endvertex pair
creates a cycle, and (often) a longer path.

Definition. Let `(G) be the length of a longest path in G.

Definition. e ∈ V (G)(2) is a booster for G if `(G + e) > `(G) or G + e is Hamiltonian.
Observe: Iteratively adding |G| boosters to G gives a Hamiltonian graph.
Exercise. Show that any (2, n/5)-expander is connected.
Lemma 16. A (2, n/5)-expander G has at least n2 /50 boosters.
Proof. Let P be a longest path in G, with v an endpoint. For each x ∈ F (P, v), let Px be a
v, x-path with V (Px ) = V (P ). Let

E = {xy : x ∈ F (P, v), y ∈ F (Px , x)}.

Claim A: |E| ≥ n2 /50.

n n
Proof of Claim A. By Lemma 10, |F (P, v)| ≥ 5 and, for each x ∈ F (P, v), |F (Px , x)| ≥ 5.
Thus, |E| ≥ ( n5 )2 /2.

10
Claim B: Each e ∈ E is a booster.
Proof of Claim B. For each e ∈ E, G + e contains a cycle with vertex set V (P ), Ce say. If
V (P ) = V (G), G + e is thus Hamiltonian. If not, then, as G is connected, there is some
w ∈ V (G) \ V (P ) and z ∈ V (P ) with wz ∈ E(G), so that Ce + wz has a path with vertex set
V (P ) ∪ {w}. Thus, `(G + e) ≥ `(Ce + wz) > `(G).
Lemma 17. If G0 is a (2, n5 )-expander with V (G0 ) = [n], and G = G(n, 1000
n ), then G0 ∪ G is
almost surely Hamiltonian.
Proof. Let m = 10n and p = 100
n2 . Let Gi ∼ G(n, p), 1 ≤ i ≤ m, be independent. By Proposi-
tion 3,
P(G0 ∪ G is Hamiltonian) ≥ P ∪m


i=0 Gi is Hamiltonian .

For each 1 ≤ i ≤ m, let

Gi contains a booster for ∪i−1


1 j=0 Gj ,
Xi =
0 otherwise.

Pm
Observe: i=1 Xi ≥ n =⇒ ∪m
i=0 Gi is Hamiltonian.

For each 1 ≤ i ≤ m, ∪i−1 n

j=0 Gj is a (2, 5 )-expander as it contains G0 . Thus, by Lemma 16,
∪i−1
j=0 Gj
2
has at least n /50 boosters, and hence

2 1
P(Xi = 0) ≤ (1 − p)n /50
≤ e−2 ≤ .
2
P m
i 1
Thus, j=1 (Xj − 2 ) is a submartingale, with differences at most 1. By Theorem 15,
i=1
n2
Pm 1
Pm
P( j=1 (Xj − 2 ) < −n) ≤ exp(− 2m ) = o(1). Thus, almost surely, j=1 Xj ≥ m/2 − n ≥ n and
hence ∪m
j=0 Gj is Hamiltonian.

To apply this we need to find large expanding subgraphs almost surely in G(n, C
n ). The
expansion will follow from a few simple principles:

• Typically, there are few vertices with small degree.

• Vertices of small degree are mostly spaced apart.
• For sets U of non-small degree vertices, U ∪ N (U ) contains many edges, so typically must
be much bigger than U .

Definition. For any A ⊂ V (G), let eG (A) = e(G[A]).

Definition. Let S ⊂ V (G). An S-path in G is a path with length at most 4 between two
vertices in S. An S-cycle is a cycle with length at most 4 containing some vertex in S.
Lemma 18. Let m, D ≥ 4 be integers. Let G be an n-vertex graph and S = {v : d(v) < D}.
Let x be the number of vertices in G with degree at most 1. Let y be the number of S-paths and
S-cycles. Suppose that the following hold.
(1) x + 7y ≤ n/5 and y ≤ m.
(2) Every set A ⊂ V (G) with |A| = m satisfies |N (A)| ≥ 4n/5.

11
 (3) There are no sets A ⊂ V (G) with |A| ≤ 10m and eG (A) ≥ D|A|/100.
Then, G contains a (2, n/5)-expander with at least n − x − 7y vertices.
Proof. Let B0 be the set of vertices with degree at most 1 in G or vertices in an S-path or
S-cycle. Note that |B0 | ≤ x + 5y and |B0 \ S| ≤ 3y. Let B1 ⊂ V (G) \ (B0 ∪ S) be a largest set
subject to either B1 = ∅ or |B1 | ≤ y and eG ((B0 ∪ B1 ) \ S) ≥ D|B1 |/2.
Claim A. |B1 | < y.
Proof of Claim A. Suppose to the contrary that |B1 | = y. Then, as |B0 \ S| ≤ 3y,

eG ((B0 ∪ B1 ) \ S) ≥ D|B1 |/2 ≥ D|(B0 ∪ B1 ) \ S|/8.

But, as |(B0 ∪ B1 ) \ S| ≤ 4y ≤ 10m, this contradicts (3).

Let B2 = N (B1 ) ∩ S. Each vertex in B1 is not in B0 , so has at most one neighbour in S.

Thus, |B2 | ≤ |B1 | ≤ y. Let H = G − (B0 ∪ B1 ∪ B2 ), and note that |H| ≥ n − x − 7y.
Claim B. For each v ∈ V (H) \ S, |NH (v) \ S| ≥ D/4.
Proof of Claim B. As v ∈ / S, dG (v) ≥ D. As v ∈/ B0 , v can have at most 1 neighbour in S. If
|NG (v) ∩ (B0 ∪ B1 )| ≥ D/2, then eG ((B0 ∪ B1 ∪ {v}) \ S) ≥ D(|B1 | + 1)/2, contradicting the
choice of B1 . Thus, |NH (v) \ S| ≥ D − 1 − D/2 ≥ D/4.

Claim C. H is a (2, n/5)-expander.

Proof of Claim C. Let U ⊂ V (H) with 0 < |U | ≤ n/5. If |U | ≥ m, then, picking a set U 0 ⊂ U
with |U 0 | = m and considering N (U 0 ), we have, using (1) and (2),

|NH (U )| ≥ 4n/5 − (|G| − |H|) − |U | ≥ 2|U | + (n/5 − x − 7y) ≥ 2|U |.

Suppose then that |U | < m. Let U1 = U ∩ S and U2 = U \ S. Each vertex in U1 ⊂ S has no

neighbours in S, else we removed it. If a vertex in U1 ⊂ S has a neighbour in G on an S-path or
S-cycle, or a neighbour in S \ {v} or N (S \ {v}), then v is on an S-path or S-cycle, and so is in
B0 , a contradiction. Thus, every vertex in U1 has non neighbours in V (G) \ V (H). Furthermore,
no two neighbours in U1 share any neighbours, as the vertices would then have been in B0 . Thus,
|NH (U1 ) \ S| ≥ 2|U1 |.
Furthermore, by Claim B, e((U2 ∪ NH (U2 )) \ S) ≥ D|U2 |/8, and therefore, by (3) (and as
|U2 | < m), |(U2 ∪ NH (U2 )) \ S| > 10|U2 |. Now, each vertex v ∈ U2 has at most one neighbour in
NH (U1 ), for otherwise v is in an S-path or S-cycle. Thus, |NH (U1 ) ∩ NH (U2 )| ≤ |U2 |. Thus,

|NH (U )| ≥ |NH (U1 )\S|−|U2 |+|NH (U2 )\S|−|NH (U1 )∩NH (U2 )| ≥ 2|U1 |+7|U2 | ≥ 2|U |.

Proposition 19. Let C be a sufficiently large constant, G = G(n, C/n) and S = {v : d(v) <
C/100}. Then, almost surely, we have the following properties.
a) There are at most (1 + oC (1))Ce−C n vertices with degree at most 1.
b) There are at most e−3C/2 n S-paths and S-cycles.

12
Proof. For illustration, we will show that the expected number of such vertices or paths and
cycles satisfy a) and b), and leave their proofs by the second moment method to an exercise.
a) Let p = C/n and let X be the number of vertices with degree at most 1. Then,
EX = n · ((1 − p)n−1 + (n − 1)p(1 − p)n−2 ) = (1 + o(1))pn(1 − p)n = (1 + o(1))Ce−C n.

b) Let Y be the number of cycles with length at most 4. Then,

4
X
EY ≤ nk pk ≤ 2C 4 .
k=3

By Markov’s inequality, we have P(Y ≥ log n) = o(1).

Let δ = 1/100 and let Z be the number of S-paths. Then,
4 2δC   !
X
`+1 `
X 2n i 2n−5−i
EZ ≤ n p · p (1 − p)
i=0
i
`=1
2δC  i
X 2enp
≤ 4nC ·4
e−2pn(1−o(1)) .
i=0
i
 enp 2δC
≤ 8δnC · 5
e−2C(1−o(1)) .
δC
 e 2δC
≤ nC 5 · e−2C(1−o(1)) .
δ
≤ n · C 5 exp(−2(1 − o(1) − 2δ log(e/δ))C) = oC (e−3C/2 ) · n.
Exercise. Using the second moment method, prove a) and b).
Proposition 20. Let p = C/n with C sufficiently large and C ≤ 2 log n. Let G = G(n, p) and
m = n/1016 . Then, almost surely, we have the following properties.
a) Every subset A ⊂ V (G) with |A| = m satisfies |N (A)| ≥ 4n/5.
b) There is no set A ⊂ V (G) with |A| ≤ 10m and eG (A) ≥ C|A|/107 .
Proof. a) follows directly from Proposition 13.
b) Let δ = 10−7 . For each 1 ≤ t ≤ 10m, let pt be the probability there is no set A ⊂ V (G)
with |A| = t and eG (A) ≥ δC|A|. Then,
  2 
n t /2 δCt
pt ≤ p
t δCt
 en t  etp δCt
≤
t 2δC
 2 t  δCt−t
e tpn etp
≤
2δCt 2δC
 2 t  δCt−t  3 δCt/2
e et e t
= ≤ .
2δ 2δn 4δ 2 n
√ √
If t ≤ n, then, as δCt ≥ 10, pt = o(n−1 ). If t ≥ n, then, as e3 t/4δ 2 n ≤ 3/4, pt = O((3/4)t ) =
P10m
o(n−1 ). Thus, b) holds with probability at least 1 − t=1 pt = 1 − o(1).

13
 −(1−o(1))C
Theorem 21. G(n, C
n ) almost surely has a cycle with at least (1 − e )n vertices.
Proof. Let λ = C − 1000 and G = G(n, λ/n). We will show that, almost surely, G contains a
(2, n/5)-expander on at least e−(1−oC (1))C n vertices. The result will then follow by Lemma 17
and Proposition 3.
Let S = {v : d(v) < λ/100}, X = |{v : d(v) ≤ 1}| and let Y be the number of S-paths and
S-cycles. Almost surely, by Proposition 19, X + 7Y ≤ (1 + o(1))λe−λ n ≤ e−(1−oC (1))C n. Almost
surely, a) and b) in 20 hold in G with C replaced by λ. Then, by Lemma 18 with m = n/1016
and D = λ/100, G is a (2, n/5)-expander.
We can also use Lemma 17 to find a (very) sharp threshold for the appearance of Hamilton
cycles in the binomial random graph, for which we need the following.
Proposition 22. Let log n ≤ λ ≤ 2 log n, G = G(n, λ/n) and S = {v : d(v) ≤ λ/100}. Then,
almost surely, G has no S-paths or S-cycles.
Proof. Let δ = 1/100 and p = λ/n. Let X be the number of S-paths. Then,
 X 2 2δλ  
n X 2n i
EX ≤ pk+1 nk · p (1 − p)2n−5−i
2 i=0
i
k=0
 2δλ
2enp
2 2
≤ 3n pλ · (2δλ + 1) e−(2+o(1))np
2δλ
≤ (n log4 n) · exp(2δλ log(1/δ) − (2 + o(1))λ) = o(1).

Then, almost surely, there are no paths with length at most 4 and endpoints in S.
Let Y be the number of S-cycles. Then, similarly,
3 δλ  
X X n−3
EY ≤ n · pk+1 nk · pi (1 − p)n−3−i
i=0
i
k=2
 enp δλ  e δλ
≤ 2λ4 · (δλ + 1) · e−(1+o(1))np ≤ λ5 · e−(1+o(1))λ = o(1).
δλ δ
Thus, almost surely, there are no such cycles in G.
This gives the following theorem, due originally to Bollobás [5] and Komlós and Szemerédi [18].
log n+log log n+ω(1)
Theorem 23. If p = n , then G(n, p) is almost surely Hamiltonian.

Proof. Let q = p − 1000/n and G = G(n, q). By Lemma 17 and Proposition 3, it is sufficient to
show that G is almost surely a (2, n/5)-expander.
Note that q = log n+log nlog n+ω(1) . Thus, by Lemma 5, δ(G) ≥ 2 almost surely. Let δ = 10−5 ,
m = n/1016 and S = {v : d(v) < δqn}. Almost surely, by Proposition 20, any subset A ⊂ V (G)
with |A| = m satisfies |N (A)| ≥ 4n/5 and there is no set A ⊂ V (G) with |A| ≤ 10m and
eG (A) ≥ δqn/100. Almost surely, by Proposition 22, there are no cycles with a vertex in S or
paths with length at most 4 between vertices in S. Thus, by Lemma 18 with m and D = δqn,
G is a (2, n/5)-expander, as required.

14
4.5 Long paths in G(n, Cn ) V: Conditional existence of boosters
We will improve on Theorem 21 to give our final result on long cycles in G(n, C n ) using a
conditioning argument by Lee and Sudakov [21].
Given any fixed (2, n5 )-expander H ⊂ Kn , H has many boosters, so that it is very likely one
of them will appear in G = G(n, p). Typically there are many such graphs H which will appear
in G, but by restricting such graphs further we can find a large set of (2, n5 )-expanders in G
which all have boosters in G. The graphs H being sparse in comparison to G is sufficient for
this to hold, as follows.
Lemma 24. Let p ≥ 1/n and G = G(n, p). Almost surely, any (2, n5 )-expander H ⊂ G with
e(H) ≤ pn2 /104 has a booster in G.
Proof. Let δ = 10−4 . Let H be the set of (2, n/5)-expander graphs H with V (H) ⊂ [n] and
e(H) ≤ δpn2 . For each H ∈ H, let BH be the set of boosters in H which are not in E(H), so
that, by Lemma 10,
|BH | ≥ n2 /50 − δpn2 ≥ n2 /100.
For each H ∈ H, BH contains no edges in H, so that
2
P(BH ∩ E(G) = ∅|H ⊂ G) = P(BH ∩ E(G) = ∅) = (1 − p)|BH | ≤ e−pn /100
.

Let q be the probability that, for some H ∈ H, H ⊂ G and e(G) ∩ BH = ∅. Then,

X X
q≤ P((H ⊂ G) ∧ (BH ∩ E(G) = ∅)) = P(BH ∩ E(G) = ∅|H ⊂ G) · P(H ⊂ G).
H∈H H∈H

Therefore,
X
P(∃H ∈ H with H ⊂ G and BH ∩ E(G) = ∅) ≤ P((H ⊂ G) ∧ (BH ∩ E(G) = ∅))
H∈H
X
= P(BH ∩ E(G) = ∅|H ⊂ G) · P(H ⊂ G)
H∈H
2 X
≤ e−pn /100
P(H ⊂ G)
H⊂Kn ,e(H)≤δpn2
2
δpn
n2 t
 
−pn2 /100
X
≤e p
t=0
t
2
δpn t
en2 p

−pn2 /100
X
≤e
t=0
t
2
 e δpn2
≤ e−pn /100
· n2 ·
δ
≤ n2 · exp((−1/100 + δ log(e/δ))pn2 )
≤ n2 · exp(−pn2 /1000) = o(1).

Corollary 25. Let p ≥ 105 /n and G = G(n, p). Almost surely, for any (2, n5 )-expander H ⊂ G
with e(H) ≤ 5pn2 /105 , G[V (H)] is Hamiltonian.

15
Proof. Let δ = 5/105 . Almost surely, G has the property from Lemma 24. Let then H ⊂ G be
a (2, n5 )-expander with e(H) ≤ δpn2 . Let H0 = H and let ` ≤ n be the largest integer for which
there is a sequence H0 ⊂ H1 ⊂ . . . ⊂ H` ⊂ G such that, for each 1 ≤ i ≤ `, Hi is formed by
adding a booster to Hi−1 .
Now, e(H` ) ≤ δpn2 + ` ≤ pn2 /104 and H` is a (2, n/5)-expander, so that, by the property
from Lemma 24, H` has a booster in G. Thus, by the choice of `, we must have ` = n. Thus,
H` , and hence G[V (H)], is Hamiltonian.
We can now prove two essentially best possible results about long cycles in G(n, C/n) and
Hamiltonicity in G(n, p). The first of these was originally proved by Frieze [14], and the second
by Ajtai, Komlós and Szemerédi [2].

Theorem 26. For any sufficiently large constant C, G(n, C/n) almost surely has a cycle on at
least (1 − (1 + oC (1))Ce−C )n vertices.
log n+log log n+cn
Theorem 27. If p = n and c ∈ R, then

 0 if cn → −∞
−c
P(G(n, p) is Hamiltonian) → e−e if cn → c
1 if cn → ∞.


These results follow from the following theorem.

Theorem 28. There exists some sufficiently large C0 such that the following holds for any
C0 ≤ λ ≤ 2 log n. Let G = G(n, λ/n), X = |{v : d(v) ≤ 1}|, S = {v : d(v) < λ/100} and let
Y be the number of S-paths and S-cycles in G. Then, almost surely, G has a cycle on at least
n − X − 7Y vertices.
Proof. Almost surely, by Corollary 25, for any (2, n/5)-expander subgraph H ⊂ G with e(H) ≤
5λn/105 , G[V (H)] is Hamiltonian. Let m = n/1016 . Almost surely, by Proposition 20, the
following holds.

a) There are no sets A ⊂ V (G) with |A| ≤ 10m and eG (A) ≥ λ|A|/107 .

Form G0 ⊂ G, a graph with vertex set V (G) by, for each v ∈ V (G)\S, adding min{d(v), λ/105 }
edges adjacent to v. Note that e(G0 ) ≤ λn/105 .
Let G1 ⊂ G be a subset of G with each edge chosen independently at random with probability
1/105 . Noting that G1 ∼ G(n, λ/105 n), by Lemma 12 we almost surely have e(G1 ) ≤ λn/105
and, by Proposition 13, we almost surely have the following.

b) For all A ⊂ V (G) with |A| = m, |N (A)| ≥ 4n/5.

Now, G0 ∪ G1 ⊂ G has X vertices with degree at most 1 and, letting S 0 = {v : dG0 (v) <
λ/107 } ⊂ S, at most Y S 0 -paths and S 0 -cycles, and a) and b) hold. Thus, by Lemma 18
there is some (2, n/5)-expander H ⊂ G0 ∪ G1 on at least n − X − 7Y vertices. Hence, as
e(H) ≤ e(G0 ∪ G1 ) ≤ 5λ/105 , G[V (H)] is Hamiltonian. I.e., G contains a cycle with at least
n − X − 7Y vertices.
Theorem 26 follows from Proposition 19 and Theorem 28.
Exercise. Prove Theorem 27 by studying P(δ(G(n, p)) ≥ 2).

16
4.6 Hamiltonicity in the random graph process
Let G0 be the empty graph with vertex set [n] and create the

following sequence of random
graphs G0 ⊂ G1 ⊂ . . . ⊂ G(n) , where, for each 1 ≤ M ≤ n2 , GM is formed from GM −1 by,
2
uniformly at random, adding a non-edge. We call {GM }M ≥0 the random graph process. Our
aim here is to show that, almost surely, the very edge we add in this process which gives the
graph minimum degree 2 also creates the first Hamilton cycle. That is, the hitting time for
Hamiltonicity and the hitting time for minimum degree at least 2 almost surely coincide.
Definition. Let Gn,M be the random graph with vertex set [n] and M edges, with all such
graphs equally likely.
Definition. Say an event holds in almost every (a.e.) random graph process if the event holds
almost surely in the random graph process.
Switching between graph models. Note that in the n-vertex random graph process {GM }M ≥0
each GM is distributed as Gn,M . Often G(n, p) is easier to work with

due to the independence
between edges. Fortunately, the random graph Gn,M with
M = p n2 edges is closely related to
G(n, p). For example G(n, p) almost surely has around p n2 edges, as follows.

Proposition 29. Let p ≥ 10/n satisfy p = o(1). Then, almost surely, −n ≤ e(G) − p n2 ≤ n.

Proof. Let M = p n2 and ε = n/M = 2/(n − 1)p ≤ 1, so that ε2 M = n2 /M = Ω(1/p) = ω(1).

By Lemma 12 with ε, we have

P(|e(G) − M | ≤ n) ≤ 2 exp(−ε2 M/3) = o(1).

This allows us to move simply between models, as follows.

Proposition 30. If P is a graph property, then the following hold.
a) Let p ≥ 10/n satisfy

p = o(1). Suppose P almost surely contains G(n, p). Then, there is
some M with p n2 − n ≤ M ≤ p n2 + n for which, almost surely, Gn,M ∈ P.

b) Suppose that in almost every graph process, every graph has property P. Then, for any
p = p(n), G(n, p) is almost surely in P.
Proof. a) Let M = {M : p n2 − n ≤ M ≤ p n2 + n}. Let ε > 0 and let n be sufficiently large

that P(G(n, p) ∈ P) ≥ 1 − ε/2 and, using Proposition 29, that P(e(G(n, p)) ∈ M) ≥ 1 − ε/2.
Then,
P(e(G(n, p) ∈ M ∧ G(n, p) ∈ P)) ≥ 1 − ε,
so that
X
P(Gn,M ∈ P) · P(e(G(n, p)) = M ) = P(e(G(n, p)) ∈ M ∧ G(n, p) ∈ P) ≥ 1 − ε.
M ∈M

Thus, for some M ∈ M, P(Gn,M ∈ P) ≥ 1 − ε.

b) For each
ε > 0, we must have, for sufficiently large n, that P(Gn,M ∈ P) ≥ 1 − ε for each
0 ≤ M ≤ n2 . Thus, for any p,

(n2 )
X (n2 )
X
P(G(n, p) ∈ P) = P(G(n, p) = M )·P(Gn,M ∈ P) ≥ P(G(n, p) = M )·(1−ε) = 1−ε.
M =0 M =0

17
 We will show the following theorem, which was originally proved by Bollobás [6]
Theorem 31. In almost every random graph process, for each 0 ≤ M ≤ n2 , δ(GM ) ≥ 2 =⇒

GM is Hamiltonian.

Thus, we have by Proposition 30b) applied with P = {G : δ(G) ≥ 2 =⇒ G is Hamiltonian},

we have the following.
Corollary 32. For any p = p(n), we have

P(G(n, p) is Hamiltonian) = P(δ(G(n, p)) ≥ 2) + o(1).

Before we prove Theorem 31, we will record a couple of useful propositions, the first giving
a simple bound on the likely maximum degree.
Proposition 33. If p ≤ 2 log n/n, then, almost surely, ∆(G(n, p)) ≤ 10 log n.
Proof. Let X be the number of vertices with degree more than 10 log n. Then,
   10 log n
n 10 log n enp
P(X ≥ 1) ≤ EX ≤ n · p ≤n· ≤ n · (e/5)10 log n = o(1).
10 log n 10 log n
1 log n
Proposition 34. Almost surely, G = G(n, (1 − 100 ) n ) satisfies the following properties with
S = {v : d(v) < log n 16
100 } and m = n/10 .

A1 δ(G) ≤ 1.
A2 There is no set A ⊂ V (G) with |A| ≤ 10m and eG (A) ≥ |A| log n/107 .
A3 Every subset A ⊂ V (G) with |A| = m satisfies |N (A)| ≥ 4n/5.

A4 There are no S-paths or S-cycles.

A5 |S| ≤ n1/4 .
Proof. A1 follows from Lemma 5. A2 and A3 restate part of Proposition 20, while A4 restates
Proposition 22.
Proof of A5: Let X = |S| and δ = 1/100. Then,
δX
log n   δX
log n 
n i n−1−i enp i −(1−o(1))pn
EX ≤ n p (1 − p) ≤n e
i=0
i i=0
i
≤ n · log n · (e/δ)δ log n e−(1−δ−o(1)) log n = o(n1/4 ).

Thus, by Markov’s inequality, P(X ≤ n1/4 ) = o(n−1 ).

When combined with Proposition 30a), this gives the following corollary.
1 n log n
Corollary 35. There exists some M0 = M0 (n) ≤ (1 − 200 ) 2 such that in almost every
n-vertex random graph process {GM }M ≥0 , GM0 satisfies A1 – A5.
We can now prove Theorem 31, using a method from the work of Krivelevich, Lubetzky and
Sudakov [20].

18
Proof of Theorem 31. Let {GM }M ≥0 be the n-vertex random graph process, and, for each M ,
let eM be the edge added to GM −1 to create GM . Let δ = 1/200, and take M0 ≤ (1 − δ) n log 2
n
n log n n log n
from Corollary 35. Let M1 = 2 and M2 = (1 + δ) 2 .
Reveal edges e1 , e2 , . . . , eM0 . By the choice of M0 , almost surely, A1–A5 hold in GM0 . Let
S = {v : dGM0 (v) < log n
100 }.
Now, let HM0 = GM0 . For each M0 < i ≤ M2 , reveal whether ei contains a vertex in S or
not.

• If it does, let XM = 0, reveal eM (i.e., its exact location) and let HM = HM −1 + eM .

• If it does not, let XM = 1 and HM = HM −1 .

When finished, let I = {M : M0 < M ≤ M1 , XM = 1}.

Claim A. Almost surely, |I| ≥ δn log n/4.
Proof of Claim A. As, by A5, |S| ≤ n1/4 , we have, for each M , M0 < M ≤ M1 ,
!
|S| · n 3
E(XM |XM0 , . . . , XM −1 ) ≥ 1 − n
≥ .
2 − (M − 1) 4
Pi
M1
Thus, M =M0 +1 (XM − 3/4) i=M0 +1 is a submartingale. By Theorem 15, then, for t =
δ log n/8, !
M1
t2
X  
P XM < 3(M1 − M0 )/4 − t ≤ exp − ,
2(M1 − M0 )
M =M0 +1
PM1
so that, as |I| = M =M0 +1 XM and M1 − M0 ≥ δn log
2
n
,
 
δn log n
P |I| < = o(1).
4

We know the edges eM , M ∈ I, lie within [n] \ S, but nothing further about their location,
other than they are not in HM0 .
Claim B. Almost surely, for each M0 ≤ M ≤ M2 , the following event EM holds.

EM : There are no S-paths or S-cycles in HM .

Proof of Claim B. Note that we know EM0 holds by A5. If EM −1 holds then, for EM not to
hold eM must lie within SM := S ∪NHM −1 (S)∪NHM −1 (NHM −1 (S)). If, in addition, ∆(HM −1 ) ≤
10 log n, then, by A5,
|SM | ≤ 2(10 log n)2 |S| = o(n1/3 ).
Thus, letting ĒM be the event that EM fails,
!
n2/3
P(ĒM |EM −1 ∧ (∆(HM −1 ) ≤ 10 log n)) = o n

= o(n−4/3 ).
2 −M

Now,

P(ĒM for some M0 ≤ M ≤ M2 )

≤ P(∆(HM2 ) > 10 log n) + P((ĒM for some M0 ≤ M ≤ M2 ) ∧ (∆(HM2 ) ≤ 10 log n)).

19
By Propositions 33 and 30a), P(∆(HM2 ) > 10 log n) ≤ P(∆(Gn,M2 ) > 10 log n) = o(1). There-
fore,

P(ĒM for some M0 ≤ M ≤ M2 )

M2
X
≤ o(1) + P(ĒM ∧ EM −1 ∧ (∆(HM2 ) ≤ 10 log n))
M =M0 +1
M2
X
≤ o(1) + P(ĒM ∧ EM −1 ∧ (∆(HM −1 ) ≤ 10 log n))
M =M0 +1
M2
X
≤ o(1) + P(ĒM |EM −1 ∧ (∆(HM −1 ) ≤ 10 log n))
M =M0 +1

≤ o(1) + M2 · n−4/3 ≤ o(1) + (n log n) · n−4/3 = o(1).

Note that, for each M0 ≤ M ≤ M2 , δ(HM ) ≥ 2 exactly when δ(GM ) ≥ 2. Thus, by Lemma 5
and Proposition 30, we almost surely have δ(HM2 ) ≥ 2. Let N be the smallest N ≤ M2 such
that δ(HN ) ≥ 2. By A1, N > M0 , so that GN is the first graph in the sequence {GM }M ≥0 with
minimum degree 2. Let L0 = HN . By A2–A3, Claim B and Lemma 18, and as δ(L0 ) ≥ 2, L0
is a (2, n/5)-expander.
Let m = |I| ≥ δn log n/4 (by Claim A) and label I = {ji : i ∈ [m]} so that j1 < . . . < jm .
PM
For each 1 ≤ M ≤ m, let LM = L0 + ( i=1 eji ). For each 1 ≤ M ≤ m, let YM be the indicator
function for the event that ejM is a booster for LM −1 .
For each 1 ≤ M ≤ m, As L0 , and thus LM −1 , is a (2, n5 )-expander, by Lemma 10, the number
2
of boosters for LM −1 within [n] \ S but not in E(LM −1 ) is at least n50 − n · |S| − M ≥ n2 /100.
Thus, for each 1 ≤ M ≤ m,
 2  . 
n n 1
E(YM |Y1 , . . . , YM −1 ) ≥ ≥ .
100 2 50
Pi
By applying
Pm Theorem 15 to the submartingale ( M =1 (YM − 1/50))m i=1 , we have that, almost
surely, M =1 YM ≥ m/100 ≥ n. Thus, Lm is formed from L0 by adding iteratively n boosters
(and some other edges), and thus is Hamiltonian. Therefore, GN = Lm , the first graph in the
sequence with minimum degree 2 is Hamiltonian.

5 Random Directed Graphs

Orienting the edges of a graph can often create an interesting new problem. For example, giving
each edge uv = {u, v} an orientation and looking for a long directed cycle — where the edges
have the same direction around the cycle — nullifies our previous approach using Pósa rotation,
as rotation does not maintain a directed path.
Definition. An oriented graph D has vertex set V (D) and a set E(D) ⊂ {uv
~ : u, v ∈ V (D), u 6=
v} of directed edges, where at most one of uv
~ and vu
~ are in E(D).
Definition. A directed graph (digraph) D has vertex set V (D) and a set E(D) ⊂ {uv
~ : u, v ∈
V (D), u 6= v} of directed edges, where uv
~ and vu
~ may both be in E(D).
We will work in random digraphs, though in practice we will typically have few u, v ∈ V (D)
with uv,
~ vu~ ∈ E(D)

20
Definition. Let D(n, p) be a random digraph with vertex set [n], where each of the possible
n(n − 1) edges occurs independently at random with probability p.
An elegant coupling argument of McDiarmid [22] allows us to give a good bound on the
appearance of oriented subgraphs in D(n, p), if we have a good bound on the appearence of the
underlying subgraph in G(n, p).
Theorem 36. Let F be a digraph created from a graph H by orienting edges. Then, for any
n ∈ N and p ∈ (0, 1),
P(D(n, p) contains a copy of F ) ≥ P(G(n, p) contains a copy of H).
Proof. We will interpolate between D0 ∼ G(n, p) and DN ∼ D(n, p), with N = n2 , using a

sequence of random digraphs D0 , D1 , . . . , DN defined as follows using an arbitrary enumeration

e1 , . . . , eN of the edges of Kn , with ei = {vi , wi }.
For each 0 ≤ i ≤ N , let Di be the random digraph where, for each j ∈ [N ],
• if j ≤ i, then vj~wj and w~j vj are included independently of each other with probability p,
and
• if j > i, then vj~wj and w~j vj are included together with probability p, and otherwise
omitted.
Note that DN ∼ D(n, p), and D0 can be formed by replacing each uv ∈ E(G(n, p)) with uv
~ and
vu.
~ Considering the latter, in such a coupling any copy of H in G(n, p) corresponds to some
copy of F in D0 . Thus, using ⊂
∼ to denote ‘contains a copy of’,
P(F ⊂
∼ D0 ) = P(H ⊂
∼ G(n, p)).

Now, let i ∈ [N ] and note that Di and Di−1 differ only on {vi~wi , w~i vi }. Let D̃i = Di −
{vi~wi , w~i vi } = Di−1 − {vi~wi , w~i vi }.
Either a) F ⊂ ∼ D̃i , so that P(F ⊂
∼ Di |D̃i ) = P(F ⊂ ∼ Di−1 |D̃i ) = 1.
or b) F 6⊂
∼ D̃i , but there is some e ∈ {vi~wi , w~i vi } such that F ⊂ ∼ D̃i + e, so that
⊂ ⊂
P(F ∼ Di |D̃i ) ≥ p = P(F ∼ Di−1 |D̃i ).
or c) F 6⊂
∼ D̃i + {vi~wi , w~i vi }, so that P(F ⊂ ∼ Di |D̃i ) = P(F ⊂
∼ Di−1 |D̃i ) = 0.
Note that in this division we have used that no copy of F contains both vi~wi and w~i vi .
Thus, in each case
P(F ⊂∼ Di |D̃i ) ≥ P(F ⊂ ∼ Di−1 |D̃i ),
so that P(F ⊂∼ Di ) ≥ P(F ⊂ ∼ Di−1 ).
Therefore,
P(F ⊂ ⊂ ⊂ ⊂
∼ D(n, p)) = P(F ∼ DN ) ≥ P(F ∼ D0 ) = P(H ∼ G(n, p)).
Remark. Nothing of the specific structure of graphs is used in the proof of Theorem 36. We
equally could have proved the following
Theorem 36’. Let [n]p and ([n] × [2])p be subsets of [n] and [n] × [2] with each element chosen
independently at random with probability p. Let π : [n] × [2] → [n] be the projection onto the first
co-ordinate. Let A ⊂ P([n]) and let B ⊂ P([n] × [2]) be such that for all B ∈ B and x ∈ [n],
{x} × [2] 6⊂ B, and A ⊂ π(B). Then,
P(∃B ∈ B : B ⊂ ([n] × [2])p ) ≥ P(∃A ∈ A : A ⊂ [n]p ).

21
Definition. A digraph D with n vertices is Hamiltonian if it contains a spanning directed cycle
(i.e. one with vertex set V (D)).
Combining Theorem 36 and Theorem 23, gives the following.
log n+log log n+ω(1)
Corollary 37. If p = n , then D(n, p) is almost surely Hamiltonian.
Note. Corollary 37 was subsequently improved by Frieze [15].
Corollary 38. For each constant C > 0, D(n, C
n ) almost surely contains a directed cycle with
at least (1 − (1 + oC (1))Ce−C )n vertices.
Are these corollaries best possible? Note that the best degree condition that holds in every
Hamiltonian digraph D is that δ ± (D) ≥ 1.
Exercise. Show, for each v ∈ [n], if p = log n+ω(1)
n , D = D(n, p), G = G(n, p) and j ∈ {+, −},
then
P(djD (v) ≥ 1) = P(dG (v) ≥ 1) = 1 − o(n−1 ),
so that
P(δ ± (D) ≥ 1) ≥ 1 − 2n · o(n−1 ) = 1 − o(1).
log n+ω(1)
Exercise ++ . Show that if p = n , then D(n, p) is almost surely Hamiltonian.
Exercise ++ . Show that D(n, C
n ) almost surely contains a directed cycle with at least (1 − (1 +
oC (1))e−C )n vertices.

6 The chromatic number of dense random graphs

One simple restriction on χ(G), the chromatic number of G, follows from the size of the largest
independent set, α(G). That is, as α(G) is an upper bound on the size of any colour class,
χ(G) ≥ |G|/α(G).
1
Lemma 39. Let p ∈ (0, 1) be constant, and b = 1−p . Then, almost surely α(G(n, p)) ≤
d2 logb ne.
Proof. Let k = d2 logb ne and let X be the number of independent sets with size k. Then,
k  k
en(1 − p)(k−1)/2
  
n (k
) e
P(X > 0) ≤ EX = (1 − p) ≤2 ≤ = o(1).
k k k(1 − p)1/2
1
Corollary 40. Let p ∈ (0, 1) be constant, and b = 1−p . Then, almost surely, χ(G(n, p)) ≥
n
2 log n .
b

Idea. If we iteratively remove maximal independent sets, we can (almost surely) find a colouring
n
with (1 + o(1)) 2 log n colours.
b

To show such large independent sets exist, we will use the following definition and theorem.
Definition. For any graphs G and H, G ∩ H is the graph with vertex set V (G) ∩ V (H) and
edge set E(G) ∩ E(H).
We will use the following version of Janson’s inequality, which, for example, follows directly
from Theorems 8.1.1 and 8.1.2 in [3].

22
Theorem 41. (Janson’s inequality) Let p ∈ (0, 1) and let {Hi }i∈I be a family of subgraphs of
the complete graph on the vertex set [n]. For each i ∈ I, let Xi denote the indicator random
variable for the event that Hi ⊂ G(n, p) and, for each ordered pair (i, j) ∈ I × I, with i 6= j,
write Hi ∼ Hj if E(Hi ) ∩ E(Hj ) 6= ∅. Then, for
X X
X= Xi , µ = EX = pe(Hi ) ,
i∈I i∈I
X X
δ= E[Xi Xj ] = pe(Hi )+e(Hj )−e(Hi ∩Hj )
(i,j)∈I×I,Hi ∼Hj (i,j)∈I×I,Hi ∼Hj

and any 0 < γ < 1, we have

γ 2 µ2
P[X < (1 − γ)µ] ≤ e− 2(µ+δ) .

6.1 Diversion: Upper and lower tails

log n
Let n≤ p ≤ n−2/3 and let X be the number of copies of C4 in G(n, p). Let Hi , i ∈ I, be the
copies of C4 in Kn and H some specific such copy of C4 in Kn . Note that if Hi ∼ H then
Hi
must share either 1 or 2 edges with H. Using the notation in Theorem 41, we have µ = 3 n4 p4
and, by symmetry,
  X
n
δ=3 p8−e(Hi ∩H) = µ·p4 ·O(n2 p−1 +np−2 ) = µ·O(n2 p3 +np2 ) = µ·O(n2 p3 ) = O(µ).
4
i∈I:Hi ∼H

Thus, by Theorem 41, we have, for each γ ∈ (0, 1), that

2
P(X < (1 − γ)µ) ≤ e−Ω(γ µ)
. (2)
That is, the lower tail of the distribution of X is exponentially small in µ.
The upper tail is harder to analyse (but has been done by Vu [27]), and does not decrease as
√
fast. To see this, keeping the same notation, let λ = 2 4 µ = o(n) and pick disjoint sets A, B ⊂ [n]
with |A| = |B| = λ. The complete bipartite graph with vertex sets A and B, Kn [A, B], contains
λ4 /4 = 4µ copies of C4 . Thus,
2 √
P(X ≥ 4µ) ≥ P(Kn [A, B] ⊂ G(n, p)) = pλ = e− µ log(1/p)
= e−o(µ) ,
where we have used that p ≥ log n/n. Thus, (2) cannot hold for the upper tail.

6.2 The chromatic number of G(n, p).

1
Lemma 42. Let p ∈ (0, 1) be constant and b = 1−p . Let k ∈ N satisfy k = 2 logb n −
ω((log log n) ). Then, with probability at least 1 − exp(−ω(n log3 n), G(n, p) has an indepen-
2

dent set with size k.

Proof. First note that
n
= exp(ω((log log n)2 )) = ω(log10 n),
bk/2
and that we can assume that k = ω(log n/ log log n). Let Hi , i ∈ I, be the independent k-sets
in K̄n , say H is one, and let us use the notation in Theorem 41. Then,
  k−1
!k
n ( k
) n(1 − p) 2
µ= (1 − p) 2 ≥ = ω(log8k n) = ω(n log3 n). (3)
k k

23
Furthermore,
  X
n
δ= (1 − p)k(k−1)−e(Hi ∩H)
k
Hi ∼H
k−1  
µ2 X k

n i
≤ n

· (1 − p)−(2)
k i=2
i k−i
k−1  i
2k i
(1 − p)−(2)
X
2 i
≤µ · k ·
i=2
n
2 k−1 X 2k 2 (1 − p)−i/2 i−1

2 2k
≤µ · ·
n i=2 n
k−1
!
2
µ2
 
2 2k
X
−6

i−1
=µ · ·o log n =o . (4)
n i=2
n log3 n
3
Thus, by (3), (4), and Theorem 41, we have P(X = 0) ≤ e−ω(n log n)
.
1
Theorem 43. Let p ∈ (0, 1) be constant and b = 1−p . Almost surely,
n
χ(G(n, p)) = (1 + o(1)) · . (5)
2 logb n
Note. This theorem was first proved by Bollobás, using a different technique as the result
predated Theorem 41.
Proof of Theorem 43. Let ` = n/ log2 n and k = 2 logb n − (log log n)3 = 2 logb ` − ω((log log `)2 ).
By Lemma 42, the probability there is some set with size ` in G = G(n, p) with no independent
set with size k is at most
2n · exp(−ω(` log3 `)) = o(1).
Then, iteratively remove maximal independent sets until fewer than ` vertices remain. This gives
a colouring using at most
n n
+ ` = (1 + o(1)) ·
k 2 logb n
colours. When combined with Corollary 40, this gives (5).

7 Squared Hamilton cycles in random graphs

Definition. The kth power of a graph H, H k is the graph with vertex set V (H) and edge set
E(H) = {uv : dH (u, v) ≤ k}, where dH (u, v) is the length of a shortest u, v-path in H if one
exists.
It is known (as follows from a general theorem of Riordan [25]) that, for all k ≥ 3, if p =
ω(n−1/k ), then, almost surely, Cnk ⊂ n!
∼ G(n, p). We expect 2n · p
2n
copies of Cn2 in G(n, p), so the
2 −1/2
threshold for Cn ⊂∼ G(n, p) must be at least n . This is its conjectured value, which remains
open, while the best current bound is due to Nenadov and Škorić [23]. Using improved path
connection techniques, we will show the following.
 2 
Theorem 44. If p = ω logn n , then G(n, p) almost surely contains a squared Hamilton cycle.

24
7.1 Overview of the proof of Theorem 44
We will prove Theorem 44 using absorption and path connection techniques. Absorption was
introduced as a general concept by Rödl, Ruciński and Szemerédi [26]. We will build up the
squared path, initially with a special structure that allows us to absorb vertices to lengthen the
path without altering its endpoints, as specified by the following definition.
Definition. A squared (a, b, c, d)-path P ⊂ G is an absorber for a vertex set X ⊂ V (G) \ V (P )
in G if, for any X 0 ⊂ X there is a squared (a, b, c, d)-path in G with vertex set V (P ) ∪ X 0 .
The set X is known as the reservoir. Using vertices from the reservoir we can then complete
the squared cycle while covering all the unused vertices not in the reservoir. The unused vertices
in the reservoir are then absorbed into the cycle to complete the squared Hamilton cycle.
We will seperate the absorbing and covering stage into the following two lemmas.
 
Lemma 45. Let p = ω log √ n . Almost surely, G = G(n, p) contains a set X with size at least
n
n
100 log2 n
and a squared path P with V (P ) ⊂ V (G) \ X which is an absorber for X in G.

For the second lemma, we use the following definition.

Definition. Let u, v, x, y ∈ [n] be distinct vertices. We say a graph P is a (u, v, x, y)-connector
with length k + 3, if we can label its vertices as {u, v, w1 , . . . , wk , x, y} so that P is the square
of the path uvw1 . . . wk xy with the edges uv and xy removed. We call w1 , . . . , wk the interior
vertices of P .
2
Lemma 46.2
Let X ⊂ [n] satisfy |X| ≥ n/100 log n, and let a, b, c, d ∈ [n] \ X be distinct. If
p = ω logn n , then, almost surely, G(n, p) contains an (a, b, c, d)-connector that covers [n] \ X.

Proof of Theorem 44. Reveal edges in G within [n] with probability p/2. Almost surely, by
n
Lemma 45, G contains a set X with size at least 100 log 2 n and a squared path P with V (P ) ⊂

V (G) \ X which is an absorber for X in G. Say P is a squared (a, b, c, d)-path.

Reveal more edges in G with probability p/2. By Lemma 46, G − (V (P ) \ {a, b, c, d}) contains
a (b, a, d, c)-connector, Q say, which covers [n] \ (V (P ) ∪ X). Let X 0 = X \ V (Q), and let P 0 be a
squared (a, b, c, d)-path with vertex set V (P )∪X 0 (which exists by the definition of an absorber).
Combining Q and P 0 gives a squared Hamilton cycle.

For both the proof of Lemma 45 and Lemma 46, we will find connectors.

7.2 Finding connectors

We wish to find connectors with length around log n between Ω(n/ log n) many distinct pairs of
vertex pairs. We will show that some connector is very likely to exist using Janson’s inequality.
Lemma
  47. Let ` = n/10 and let ui , vi , xi , yi , i ∈ [`], be distinct vertices in [n]. Let p =
log n
ω √
n
, G = G(n, p) and k = log n. Then, with probability at least 1 − e−ω(n log n) , there
is for some i ∈ [`] a (ui , vi , xi , yi )-connector in G with length k + 3 and interior vertices in
V := [n] \ {ui , vi , xi , yi : i ∈ [`]}.
2
Proof. Note that we can assume that p = o( log √ n ). Let P be a (u1 , v1 , x1 , y1 )-connector with
n
length k + 3 and interior vertices in V . Let U = {u1 , v1 , x1 , y1 }. Let H be the set of copies of P

25
in Kn with interior vertices in V making a (ui , vi , xi , yi )-connector for some i ∈ [`]. Note that
 k
2
nk+1 ≥ |H| = ` · |V |(|V | − 1) . . . (|V | − k + 1) ≥ ` · |V |k · 1 − = Ω(nk+1 ). (6)
n

We will use Theorem 41 to bound the probability no graph in H appears in G(n, p). Let X
be the number of graphs in H appearing in G(n, p) and let µ = EX. Then, by (6),

µ = Ω(|H| · pe(P ) ) = Ω(nk+1 p2k+3 ) = Ω(p · (np2 )k+1 ) = ω(n log n). (7)

For each H, H 0 ∈ H, write H ∼ H 0 if e(H ∩ H) > 0, and let

X 0 0 X
δ= pe(H)+e(H )−e(H∩H ) = |H| p2e(P )−e(H∩P ) ,
H,H 0 ∈H:H∼H 0 H∈H:H∼P

where we have used symmetry.  2 

µ
Using some subsequent claims, we will show that δ = o n log n . Whence, with (7) and
Theorem 41, the claim in the lemma will follow.
In order to bound δ, we divide into cases depending on whether H ∈ H with H ∼ P is a
(u1 , v1 , x1 , y1 )-connector (and so, like P , contains U ) or not. For this, let

J1 = {P ∩ H : H ∈ H, e(P ∩ H) > 0, U ∩ V (J) = ∅}

and
J2 = {P ∩ H : H ∈ H, e(P ∩ H) > 0, U ⊂ V (J)}.
First we shall calculate, for each possible intersection J, a bound on how many graphs in H
intersect with P in J.
Claim A1. For each J ∈ J1 , |{H ∈ H : H ∩ P = J}| ≤ 4k |J|−1 nk+1−|J| .
Proof of Claim A1. To choose H ∈ H with H ∩ P = J, we can choose where the copy of J
appears in H, choose i ∈ [`] (` ≤ n choices), and then choose the remaining vertices in H (at
most nk−|J| choices).
Recall that e(J) ≥ 1, choosing where J appears in H can be done by choosing the location
of some edge (at most e(P − U ) ≤ 4k choices), and then the location of the other vertices in J
(at most k |J|−2 choices). Thus, |{H ∈ H : H ∩ P = J}| ≤ 4k · k |J|−2 · ` · nk−|J| .
Claim A2. For each J ∈ J2 , |{H ∈ H : H ∩ P = J}| ≤ 4k |J−U |−1 nk−|J−U | .
Proof of Claim A2. Note that if e(J − U ) > 0, then the reasoning for Claim A1, noting that the
choice of i ∈ [`] is fixed at i = 1, implies the result.
If e(J − U ) = 0, then J must contain some edge incident to U , and therefore there are at
most 2 choices for its other vertex in the copy of H. Thus, there are at most 2k |J−U |−1 ways J
could be embedded in H. Along with at most nk−|J−U | choices for the other vertices of H, this
gives the result.
Next, we shall prove bounds on the number of edges in the intersection of graphs with P .
Claim B1. For each J ∈ J1 , e(J) ≤ 2(|J| − 1) − 1, so that
 |J|−1
−e(J) 1
p ≤p· .
p2

26
Proof of Claim B1. Consider the natural ordering of P beginning a, b and ending c, d. Count
the edges in J going to the left from each vertex in J using this order. The leftmost two vertices
between them have at most 1 edge going to the left, while every other vertex has at most 2 edges
going to the left. Thus, e(J) ≤ 2(|J| − 2) + 1 = 2(|J| − 1) − 1, as required.

Claim B2. For each J ∈ J2 ,

 |J−U |
−e(J−U ) 10
p ≤ . (8)
p2
Proof of Claim B2. If two consecutive vertices in P are not in J, then counting leftwards edges
and rightwards edges from that point, we have e(J − U ) ≤ 2|J − U |, so that (8) holds.
If this does not happen and |J − U | ≤ k − 3, then counting along leftward edges we have
e(J − U ) ≤ 2|J − U |, so that (8) holds.
Finally, if |J −U | ≥ k −2, then counting along leftward edges we have e(J −U ) ≤ 2|J −U |+3.
As 10k−2 = ω(n2 ) = ω(p−3 ), so that (8) holds.
Finally, we will prove a simple bound on the number of possible intersections J ∈ J1 and
J ∈ J2 of each order.
Claim C1. For each 2 ≤ i ≤ k, |{J ∈ J1 : |J| = i}| ≤ 4k i−1 4i .
Proof of Claim C1. Pick V ⊂ V (P ) with |V | = i with e(P [V ]) > 0 by choosing an edge (in at
most 2k ways) and then choosing the remaining i − 2 vertices (in at most k i−2 ways). As P [V ]
has at most 2i edges, there can be at most 22i subgraphs of P [V ] with vertex set V . Thus,
|{J ∈ J1 : |J| = i}| ≤ 4k i−1 4i , as required.
Claim C2. For each 1 ≤ i ≤ k, |{J ∈ J2 : |J − U | = i}| ≤ 8k i−1 4i .
Proof of Claim C2. For each 1 ≤ i ≤ k, |{J ∈ J2 : |J − U | = i, e(J − U ) > 0}| ≤ 4k i−1 4i ,
similarly to the proof of Claim C1. For each J ∈ J2 with e(J − U ) = 0, as e(J) > 0, V (J) must
contains some vertex in NP (U ). Thus, |{J ∈ J2 : |J − U | = i, e(J − U ) = 0}| ≤ 4k i−1 22i , and
the claim holds.
We can now bound δ, where we use, for example, A to indicate where the Claims A1 and A2
are applied. This gives us
X
δ = |H| p2e(P )−e(H∩P )
H∈H:E(H∩P )6=∅
!
X X X X
e(P ) −e(J) −e(J)
= µp · p + p
J∈J1 H∈H:H∩P =J J∈J2 H∈H:H∩P =J
!
A X X
e(P ) k+1−|J| |J|−1 −e(J) k−|J−U | |J−U |−1 −e(J)
≤ µp · n · 4k ·p + n · 4k ·p
J∈J1 J∈J2
!!
µ2 X
1−|J| 1 X −|J−U | |J−U | −e(J)
|J|−1 −e(J)
=O · n ·k ·p + n ·k ·p
n k
J∈J1 J∈J2

X  k |J|−1 |J−U | !!
µ2

B 1 X 10k
=O · ·p+
n np2 k np2
J∈J1 J∈J2

27
 k  i−1 k  i !!
C µ2 X k i−1 i 1 X 10k
=O · p· 2
· 4k ·4 + · 2
· 8k i−1 · 4i
n i=2
np k i=1
np
k  2 i−1 k   !!
2 i
µ2

X 4k 1 X 40k
=O · p· 2
+ .
n i=2
np k i=1
np2

Thus, as np2 = ω(k 2 ),

µ2 µ2
    
1
δ=o · p+ =o .
n k n log n
Therefore, by this, (7) and Theorem 41, the probability some graph in H exists in G(n, p) is at
least 1 − exp(−ω(n log n)).
 
In G(n, p) with p = ω log √ n , we almost surely have the following property. Given any
n
linearly sized subset of ‘unused’ vertices, revealing more edges is likely to connect any given pair
of vertex pairs. That is, we have the following.
 
Corollary 48. If p = ω log √n
n
and k = log n, then the following is almost surely true in
G = G(n, p) for any A ⊂ [n] with |A| ≥ n/2.

P Given any distinct vertices u, v, x, y ∈ [n] \ A, if more edges are revealed between {u, v, x, y}
and A with probability p, forming the edge set E, then, with probability at least 1 − n−2 ,
there is a (u, v, x, y)-connector in G + E with length k + 3 and interior vertices in A.

Proof. We will show that, for large n and any set A ⊂ [n] with |A| ≥ n/2, P does not hold with
probability exp(−ω(n log n)). As there are certainly at most 2n such sets A, the lemma follows
by a union bound.
Fix A ⊂ [n] with |A| ≥ n/2 and let q be the probability that P does not hold for A. Reveal
edges in G with probability p, where V (G) = [n], and suppose that P does not hold.
Let ` = n/10 and take disjoint vertices ui , vi , xi , yi , i ∈ [`] in [n] \ A. Reveal edges between
A and [n] \ A with probability p. As P does not hold, for each i ∈ [`] the probability there is no
(ui , vi , xi , yi )-connector with length k + 3 and interior vertices in A is at least n−2 . Furthermore
the that this happens is independent for each i ∈ [`], and each possible edge has been revealed
with probability at least p.
Therefore, in G(n, p), the probability that there is, for each i ∈ [`], no (ui , vi , xi , yi )-connector
with length k + 3 and interior vertices in A is at least q · n−2` = q · exp(−O(n log n)). But,
by Lemma 47, this probability is at most exp(−ω(n log n)), so that we must therefore have
q = exp(−ω(n log n)), as required.
We can use Corollary 48 to reveal edges and apply P iteratively to find many connectors.
Corollary 49. Let ` = n/5 log n, and let ui , vi , xi , yi , i ∈ [`], be distinct vertices in [n]. Let
log
√ n . Then, almost surely, in G(n, 2p) there are a set of disjoint
k = log n and p = ω n
(ui , vi , xi , yi )-connectors, i ∈ [`], with length k + 3.
Proof. Let G0 = G(n, p) and V := [n] \ {ui , vi , xi , yi : i ∈ [`]}. Almost surely, by Corol-
lary 48, each A ⊂ [n] with |A| ≥ n/2 satisfies P. We will, almost surely, iteratively find disjoint
(ui , vi , xi , yi )-connectors Pi , i ∈ [`], with length k + 3 and interior vertices in V \ (∪j<i V (Pj )).
For each i, 1 ≤ i ≤ `, reveal edges with probability p between {ui , vi , xi , yi } and V and add
them to Gi−1 to get Gi . By P, and as |V \ (∪j<i V (Pj ))| ≥ n − (k + 4)` ≥ n/2, with probability

28
at least 1 − n−2 , we can find a (ui , vi , xi , yi )-connector Pi with length k + 3 and interior vertices
in V \ (∪j<i V (Pj )). Thus, with probability at least 1 − `n−2 = 1 − o(1) we will find a disjoint
collection of connectors as desired.
Note that we have revealed each possible edge with probability at most 2p, so that we may
compare favourably the resulting random graph with G(n, 2p).

7.3 Finding a covering connector

To prove Lemma 46, we first find a square path which covers most of the vertices in [n] \ (X ∪
{a, b, c, d}), using the following result.

Lemma 50. For all p = p(n), G(n, p) almost surely contains a squared path covering all but at
most 10 plog
2
n
vertices.
√
Proof. Let ` = 10 plog 2
n
and note we can assume p ≥ 1/ n. Pick a vertex v0 ∈ [n] and reveal
the edges between v0 and [n] \ {v0 } with probability p/2 to get the graph G. Almost surely,
v0 has some neighbour, v1 say. Then, for each 2 ≤ i ≤ n − `, reveal more edges in G between
{vi−1 , vi−2 } and [n] \ {v0 , v1 , . . . , vi−1 } with probability p/2. With probability at least 1 − (1 −
2
(p/2)2 )` ≥ 1 − e−p `/4 = 1 − o(n−1 ), there is some neighbour, vi say, of vi−1 and vi−2 in
[n] \ {v0 , v1 , . . . , vi−1 }. Thus, this almost surely finds a squared path with at least n − ` vertices
in G. As each possible edge within [n] has been revealed with probability at most p, this gives
the result of the lemma.
Proof of Lemma 46. By removing vertices from X if necessary, assume that |X| ≤ n/2. Let
G0 = G(n, p/4). Almost surely, by Lemma 50, there is a squared path in G0 − (X ∪ {a, b, c, d})
covering all but at most 250 log n/p2 = o(n/ log3 n) vertices, say this path is P0 and is a squared
(x, y, u0 , v0 )-path. Let V = [n] \ (X ∪ V (P0 )) and ` = |V | ≤= o(n/ log3 n).
Reveal more edges with probability p/4 and find a matching from V into X (using, for
example, an iterative argument), and label its edges ui vi , i ∈ [`].
log n
Let X 0 = X \ {ui , vi : i ∈ [`]}, so that |X 0 | = ω(` log n) and p = ω √


0
. Revealing more
|X |
edges in two rounds with probability p/4, and applying Corollary 49 to vertex-disjoint pairs of
vertex pairs, we can find the following connectors which are internally vertex disjoint:
• an (a, b, x, y)-connector,
• for each i ∈ [`], a (ui−1 , vi−1 , ui , vi )-connector, and

• a (u` , v` , c, d)-connector.
Combining these connectors with the squared (x, y, u0 , v0 )-path P0 and the edges ui vi , i ∈ [`],
we get an (a, b, c, d)-connector covering [n] \ X.

7.4 Diversion: Realistic ambitions for constructing absorbers

We will create an absorber for a single vertex, before joining together many such absorbers with
connectors to create an absorber for a large set. The following (a, b, c, d)-path is an absorber for
x:

29
 x x x

H := = +
a b c d a b c d
a b c d

The graph H has 9 edges, so E|{copies of H in G(n, p)}| ≤ n5 p9 . Thus, if, for example
2
p = log
√ n , then we can expect only o(n2/3 ) copies of H in G(n, p), which would give rise to a
n
reservoir set X with size o(n2/3 ).
We want instead an absorber for which we can find disjoint copies covering Ω(n) vertices.
This is the same as finding a copy of the graph created by the disjoint union of different copies
of the absorber. When in general is there likely to be a copy of a graph H in G(n, p)? As before,
we can look to the expected number of copies of H, and deduce that certainly we will need

n|H| pe(H) = Ω(1).

Even for graphs H with fixed size, this is not
the whole story. For example, the expected number
1
of copies of the following graph in G n, n2/3 is ω(1):

K:

1


However, G n, n2/3 almost surely contains no copy of K5 , so in fact almost surely contains
no such copy of K. The problem is that K is not balanced, as it has a subgraph denser than the
overall graph, so we need to consider the following quantity:
n e(H 0 ) o
m0 (H) = max : H0 ⊂ H .
|H 0 |

Exercise. For any fixed graph H with at least 1 edge, prove that p = n−1/m0 (H) is a threshold
for G(n, p) to contain a copy of H. (This was originally shown for some graphs by Erdős and
Rényi [11] and all graphs by Bollobás [4])
When finding absorbers we wish to find vertex disjoint copies of some graph H in G(n, p). An
H-factor is bn/|H|c vertex disjoint copies of H in G(n, p). We wish to find an almost H-factor,
one covering Ω(n) vertices, so we will certainly need E|{copies of H 0 in G(n, p)}| = Ω(n) for each
H 0 ⊂ H. Define then the following quantity:
n e(H 0 ) o
0 0
m1 (H) = max : H ⊂ H, |H | ≥ 2 .
|H 0 | − 1

Exercise. For any fixed graph H with at least 1 edge, prove that p = n−1/m1 (H) is a threshold
for G(n, p) to contain vertex disjoint copies of H covering at least n/2 vertices.
As we have seen (for the Hamilton cycle), looking at the expectation of different subgraphs
does not tell the whole story. Many general H-factors also demonstrate this behaviour. For
example, in a remarkable paper, Johansson, Kahn and Vu [17] proved that for graphs H where

30
e(H 0 )/(|H 0 | − 1) obtains a unique maximum over all H 0 ⊂ H and |H 0 | ≥ 2 when H 0 = H, the
threshold for an H-factor in G(n, p) is the same as the threshold in G(n, p) for each vertex to be
contained in some copy of H.
Exercise. For each graph H, find the threshold in G(n, p) for each vertex to be contained in
some copy of H.
Exercise. Give an example of a graph H in which the threshold in G(n, p) for each vertex to be
contained in some copy of H is not a threshold in G(n, p) for an H-factor.
Ideal vertex absorbers. Say we have an absorber P for x and Px is a squared path with
same endpoint pairs as P and vertex set V (P ) ∪ {x}. We wish to find such a P with as few
vertices as possible so that we may still cover Ω(n) vertices with disjoint copies of P . We will
find such a P with O(log2 n) vertices, so to see whether we can find a smaller such P suppose
that |P | = O(log2 n) and we want to find Θ(n/ log2 n) disjoint copies of P ∪ Px .
2
Suppose p = log √ n , i.e., a fairly low edge probability that we still need to consider for
n
Lemma 45. For each H ⊂ P ∪ Px with e(H) ≥ 1 we need

n|H| pe(H) = Ω(n/ log2 n)

so that, as p = n−1/2+O(log log n/ log n) ,

n|H|−1 n−(1/2−O(log log n/ log n))e(H) = Ω(1).

Thus we need m1 (P ∪ Px ) ≤ 1/(1/2 − O(log log n/ log n) = 2 + O(log log n/ log n).
Now, to simplify the argument, let us work with paths rather than squares of paths. Here,
we consider an (a, b)-path P for a vertex x and an a, b-path Px with vertex set V (P ) ∪ {x}, and
wish to have m1 (P ∪ Px ) ≤ 1 + O(log log n/ log n) (so that it is plausible we can find n/2|P |
2
disjoint copies of P ∪ Px in G n, logn n ).

Let H be the graph with edge set E(P )4E(Px ) and no isolated vertices. As x ∈ / V (P ),
x ∈ V (H) and dH (x) = 2. Note that every vertex in H must in fact have at least 2 neighbours
in H. Thus, e(H) ≥ |H|, so that m1 (P ∪ Px ) ≥ e(H)/(|H| − 1) ≥ |H|/(|H| − 1) = 1 + O(1/|H|).
Hence, e(H) ≥ |H| = Ω(log n/ log log n).
Thus, e(P ∪ Px ) ≥ |P | + |E(P )4E(Px )|/2 = |P | + Ω(log n/ log log n), so that

e(P ∪ Px )  log n 
m1 (P ∪ Px ) ≥ ≥1+Ω .
|P | − 1 |P | · log log n

Thus, we must have |P | = Ω(log2 n/(log log n)2 ).

Similar, but more involved arguments, show that for squared paths the best we can do is
find an absorber for a vertex with Ω(log2 n/(log log n)2 ) vertices. We will get close to this,
constructing an absorber with O(log2 n) vertices – the difference being for ease of calculation
rather than a fundamental restriction of the construction.

7.5 Constructing absorbers

To find an absorber for a large set X we will construct an absorber for each x ∈ X before
connecting them using Corollary 49. Each absorber will be constructed from a subgraph we call
a gadget, many copies of which will exist in our random graph, as shown by Janson’s inequality.
Definition. As depicted in Figure 1, a k-gadget for x in G is a 2k-vertex subgraph H of G−{x},
with vertex set {ui , vi : i ∈ [k]} say, so that

31
 • for each 1 ≤ i ≤ k − 1, ui vi+1 , vi ui+1 ∈ E(H),
• for each 1 ≤ i ≤ k, ui vi ∈ E(H),
• u1 uk ∈ E(H) and xu1 , xv1 , xuk , xvk ∈ E(G).

u3
u4 u2
v3
v4 v2
u5 u1
v5 v1
x
v6 v10
u6 u10
v7 v9
v8
u7 u9
u8

Figure 1: A k-gadget for x, with k = 10.

Exercise. Show that if k = log n and H is a k-gadget, then m1 (H) = 2 + O(log log n/ log n).
Lemma 51. Let k = log n and ` = n/10k. If p = ω( logn n ), then, almost surely, G = G(n, p)
contains, disjointly, vertices xi , i ∈ [`], and subgraphs Hi , i ∈ [`], such that, for each i ∈ [`], Hi
is a (4k + 2)-gadget for xi in G.
Exercise. Prove Lemma 51 using an application of Janson’s inequality (see also the proof of
Lemma 47).

Lemma 52. Suppose H with vertex set {ui , vi : i ∈ [4k + 2]} is a (4k + 2)-gadget for x in
G as depicted in Figure 1. Suppose Pi , 1 ≤ i ≤ 2k, is a disjoint collection of connectors in
G − x so that, if i is odd, Pi is a (ui+2k+1 , vi+2k+1 , vi , ui )-connector and, if i is even, Pi is
a (ui , vi , vi+2k+1 , ui+2k+1 )-connector. Then, the following squared (v2k+1 , u2k+1 , u4k+2 , v4k+2 )-
path P , as depicted in Figure 2, is an absorber for x.

v2k+1 u2k+1 P1 P2 P3 P4 . . . P2k−1 P2k u4k+2 v4k+2 .

Proof. As depicted in Figure 3, the following is a squared (v2k+1 , u2k+1 , u4k+2 , v4k+2 )-path with
vertex set V (P ) ∪ {x}.

v2k+1 u2k+1 P2k P2k−1 P2k−2 P2k−3 . . . P2 P1 xu2k v2k .

We can now find disjoint gadgets and connect them together as in Lemma 52 to create a
squared path capable of absorbing any subset of a large set X, which proves Lemma 45.

32
 u3
u4 u2
v3
v4
v2
u5 v5 u1
v1
x
v6
u6 v10 u10
v7
v9
v8
u7 u9
u8

Figure 2: A (v5 , u5 , u10 , v10 )-squared path P which is an absorber for {x}. The dotted edges
exist in the parent graph to allow the path to absorb x.

Proof of Lemma 45. Let ` = n/100 log2 n and k = log n. By Lemma 51, G0 = G(n, p/2) almost
surely contains disjointly vertices xi , i ∈ [`], and subgraphs Hi , i ∈ [`], such that Hi is a (4k + 2)-
gadget for xi in G0 . For each i ∈ [`], label V (Hi ) as {ui,1 , vi,1 , . . . , ui,4k+2 , vi,4k+2 }, in the manner
of Figure 1. Let X = {x1 , . . . , x` }.
Note that `(2k +1) ≤ n/10k. Let G1 = G(n, p/2) and G = G0 ∪G1 . We will find the required
absorbing squared path almost surely in G. Almost surely, by Corollary 49, G1 − X contains
disjointly the following connectors.
• For each i ∈ [`] and j ∈ [2k] a (ui,j+2k+1 , vi,j+2k+1 , vi,j , ui,j )-connector, Qi,j say, if j is
odd, and a (ui,j , vi,j , vi,j+2k+1 , ui,j+2k+1 )-connector, Qi,j say, if j is even, and
• for each i ∈ [` − 1] a (ui,4k+2 , vi,4k+2 , vi+1,2k+1 , ui+1,2k+1 )-connector Pi .
By Lemma 52, for each i ∈ [`], the following squared (vi,2k+1 , ui,2k+1 , ui,4k+2 , vi,4k+2 )-path Qi ,
is an absorber for xi .

vi,2k+1 ui,2k+1 Qi,1 Qi,2 . . . Qi,2k−1 Qi,2k ui,4k+2 vi,4k+2

For each i ∈ [`], let Q̂i be a squared (vi,2k+1 , ui,2k+1 , ui,4k+2 , vi,4k+2 )-path with vertex set V (Qi )∪
{xi }, which exists as Qi is an absorber for xi .
Then the following squared (v1,2k+1 , u1,2k+1 , u`,4k+2 , v`,4k+2 )-path is an absorber for X.

Q1 P1 Q2 P2 . . . Q`−1 P`−1 Q`

Indeed, for any X 0 ⊂ X, for each i ∈ [`], let Q0i = Qi if xi ∈

/ X 0 and Q0i = Q̂i if xi ∈ X 0 , so that

Q01 P1 Q02 P2 . . . Q0`−1 P`−1 Q0`

is a squared (v1,2k+1 , u1,2k+1 , u`,4k+2 , v`,4k+2 )-path with vertex set V (P ) ∪ X 0 .

33
 u3
u4 u2
v3
v4
v2
u5 v5 u1
v1
x
v6
u6 v10 u10
v7
v9
v8
u7 u9
u8

Figure 3: The (v5 , u5 , u10 , v10 )-squared path P with x absorbed.

References
[1] M. Ajtai, J. Komlós, and E. Szemerédi. The longest path in a random graph. Combinatorica,
1(1):1–12, 1981.
[2] M. Ajtai, J. Komlós, and E. Szemerédi. First occurrence of Hamilton cycles in random
graphs. North-Holland Mathematics Studies, 115(C):173–178, 1985.

[3] N. Alon and J.H. Spencer. The probabilistic method. John Wiley & Sons, 2004.
[4] B. Bollobás. Random graphs. Proceedings of the Eigth Combinatorial Conference, University
College, pages 41–52, 1981.
[5] B. Bollobás. The evolution of random graphs. Transactions of the American Mathematical
Society, 286(1):257–274, 1984.
[6] B. Bollobás. The evolution of sparse graphs, Graph Theory and Combinatorics (Cambridge
1983), 35-57, 1984.
[7] B. Bollobás. Random graphs. Cambridge University Press, 2001.

[8] B. Bollobás and A.G. Thomason. Threshold functions. Combinatorica, 7(1):35–38, 1987.
[9] W.F. de la Vega. Long paths in random graphs. Studia Sci. Math. Hungar, 14(4):335–340,
1979.
[10] P. Erdős and A. Rényi. On random graphs I. Publicationes Mathematicae (Debrecen),
6:290–297, 1959.
[11] P. Erdős and A. Rényi. On the evolution of random graphs. Publ. Math. Inst. Hung. Acad.
Sci, 5:17–61, 1960.

34
[12] E. Friedgut. Sharp thresholds of graph properties, and the k-sat problem. Journal of the
American mathematical Society, 12(4):1017–1054, 1999.
[13] A. Frieze and M. Karoński. Introduction to random graphs. Cambridge University
Press, 2015.

[14] A.M. Frieze. On large matchings and cycles in sparse random graphs. Discrete Mathematics,
59(3):243–256, 1986.
[15] A.M. Frieze. An algorithm for finding Hamilton cycles in random directed graphs. Journal
of Algorithms, 9(2):181–204, 1988.

[16] S. Janson, T. Luczak, and A. Rucinski. Random graphs. John Wiley & Sons, 2011.
[17] A. Johansson, J. Kahn, and V. Vu. Factors in random graphs. Random Structures &
Algorithms, 33(1):1–28, 2008.
[18] J. Komlós and E. Szemerédi. Limit distribution for the existence of Hamiltonian cycles in
a random graph. Discrete Mathematics, 43(1):55–63, 1983.
[19] M. Krivelevich, C. Lee, and B. Sudakov. Long paths and cycles in random subgraphs of
graphs with large minimum degree. Random Structures & Algorithms, 46(2):320–345, 2015.
[20] M. Krivelevich, E. Lubetzky, and B. Sudakov. Cores of random graphs are born Hamilto-
nian. Proceedings of the London Mathematical Society, 2014.

[21] C. Lee and B. Sudakov. Dirac’s theorem for random graphs. Random Structures & Algo-
rithms, 41(3):293–305, 2012.
[22] C. McDiarmid. Clutter percolation and random graphs. In Combinatorial Optimization II,
pages 17–25. Springer, 1980.

[23] R. Nenadov and N. Škoric. Powers of cycles in random graphs and hypergraphs. arXiv
preprint arxiv:1601.04034, 2016.
[24] L. Pósa. Hamiltonian circuits in random graphs. Discrete Mathematics, 14(4):359–364,
1976.

[25] O. Riordan. Spanning subgraphs of random graphs. Combinatorics, Probability and Com-
puting, 9(2):125–148, 2000.
[26] V. Rödl, A. Ruciński, and E. Szemerédi. A Dirac-type theorem for 3-uniform hypergraphs.
Combinatorics, Probability and Computing, 15(1-2):229–251, 2006.
[27] V.H. Vu. A large deviation result on the number of small subgraphs of a random graph.
Combinatorics, Probability and Computing, 10(1):79–94, 2001.

35