Functionality of Random Graphs

John Sylvester∗ Viktor Zamaraev† Maksim Zhukovskii‡

arXiv:2412.19771v1 [math.CO] 27 Dec 2024

Abstract

The functionality of a graph G is the minimum number k such that in every induced subgraph
of G there exists a vertex whose neighbourhood is uniquely determined by the neighborhoods
of at most k other vertices in the subgraph. The functionality parameter was introduced in
the context of adjacency labeling schemes, and it generalises a number of classical and recent
graph parameters including degeneracy, twin-width, and symmetric difference. We establish the
functionality of a random graph G(n, p) up to a constant factor for every value of p.

1 Introduction
A common approach in algorithmic and structural graph theory to make a problem tractable is
by considering the problem on restricted classes of graphs, for example planar graphs or interval
graphs. One way to obtain a restricted graph class is to fix a graph parameter and consider the
graphs where the parameter is bounded from above by a fixed constant, for example graphs of
bounded degree or bounded treewidth.
Functionality is a graph parameter that was introduced in the context of adjacency labeling
schemes [ACLZ15, AAL21]. Speaking informally, the functionality of a graph G, denoted fun(G), is
the minimum number k such that every induced subgraph of G contains a vertex whose neighbour-
hood is uniquely determined by the neighbourhoods of at most k other vertices in the subgraph.
This parameter generalises maximum degree, degeneracy, twin-width, and symmetric difference in
the sense that if any of these parameters is bounded in a class of graphs, then functionality is also
bounded in the class. Despite its generality, functionality is still a relatively restrictive graph param-
eter as any graph class of bounded functionality is at most factorial, i.e., contains 2O(n log n) graphs
n
on n vertices [ACLZ15], which is significantly smaller than the number 2( 2 ) of all n-vertex graphs.
However, the structure of graph classes with bounded functionality is far from being understood.
√
It is known that the functionality of any n-vertex graph is at most O( n log n) and the incidence
√
graphs of projective planes have functionality Ω( n) [DFO+ 23]. A number of studies investigated
∗
Department of Computer Science, University of Liverpool, UK, [email protected],
†
Department of Computer Science, University of Liverpool, UK, [email protected],
‡
School of Computer Science, University of Sheffield, UK, [email protected],

1
functionality of graphs from some special classes. It is known that unit interval graphs have func-
tionality at most 2, line graphs have functionality at most 6, permutation graphs have functionality
at most 8 [AAL21], and interval graphs have functionality at most 8 [DLM+ 24]. On the other hand
there are examples of graph classes that are at most factorial, but have unbounded functionality,
e.g., hypercube graphs [AAL21] and box intersection graphs in Rd for any d ⩾ 3 [DLM+ 24] have un-
bounded functionality. It was shown in [DFO+ 23] that the
 functionality of  a random graph G(n, p)
3 log n 3 log n
is Ω(log n) with high probability1 (w.h.p.) for any p ∈ n ,1 − n , however nothing better
√
than the general upper bound of O( n log n) was known for the functionality of random graphs.

1.1 Our Contribution & Discussion

In this paper, we determine functionality of G(n, p) up to a constant factor for the entire range of
p ∈ [0, 1]. To avoid complicating the statement of Theorem 1.1 we state this only for p = Ω(1/n),
since for p = o(1/n) w.h.p. G(n, p) is acyclic and thus its functionality is at most 1 (in particular,
since functionality does not exceed the degeneracy). We note also that since functionality is invariant
under complementation, i.e. fun(G) = fun(G) for any graph G, it suffices to consider p ∈ [0, 1/2].

Theorem 1.1 (Main). Let Gn ∼ G(n, p), where Ω(1/n) = p := p(n) ⩽ 1/2. Then for any
w := w(n) ⩾ 1, w.h.p.   q

Θ np ln(ew) if p = ln n
;
fun(Gn ) =  ln n  q nw
Θ ln(ew) if p = w ln n
p n .
q
From Theorem 1.1, one can see that around the point p∗ = lnnn the functionality of G(n, p)
p
is maximised and becomes Θ( lnnn ). In the interval p ∈ [0, p∗ ] the functionality grows from 0 to
p p
Θ( lnnn ), and in the interval p ∈ [p∗ , 1/2] it decreases from Θ( lnnn ) to Θ(ln n). The behaviour
on the interval p ∈ [1/2, 1] mirrors that on the interval p ∈ [0, 1/2] since, as mentioned above,
fun(G) = fun(G) holds for every graph G. Figure 1 shows qualitative behaviour of the functionality
of G(n, p). q
As we have observed, the maximum functionality of G(n, p) is achieved around p∗ = ln n
n = o(1)
and not at p = Θ(1) where all other hereditary parameters mentioned above (namely, maximum
degree, degeneracy, twin-width, symmetric difference) are maximised. In fact, the functionality of
p
G(n, p) drops from Θ( lnnn ) to Θ(ln n) as p increases from p∗ to Θ(1). This behaviour might at
first seem mysterious, but it should hopefully be clear from the proofs of our bounds, outlined in
Section 1.2. In very rough terms, what is going is that for p below p∗ , if one wishes to encode the
neighbourhood N (v) of a vertex v in a random graph using adjacencies with some set of vertices
S ⊆ V , then one cannot do significantly better than just taking S = {v} ∪ N (v). However, for
p above p∗ , one can find a set S with |S| ≪ |N (v)| where for each x ∈ N (v) and y ∈
/ N (v), the
adjacencies of x and y with the set S are different.
1
With probability approaching 1 as n → ∞.

2
 fun(G(n, p))

q 
n
Θ ln n

Θ(ln n)

Θ(1)
p
1 1 1
q
ln n
n n 2

Figure 1: Qualitative behaviour of the functionality of G(n, p)

Second, random graphs G(n, p) almost achieve maximum functionality among n-vertex graphs,
p
up to a log-factor. Indeed, the maximum functionality of a random graph is Θ( lnnn ) w.h.p., and
√
the best known general upper bound on functionality [DFO+ 23] is O( n ln n). It is interesting to
√
note that incidence graphs of projective planes have functionality Θ( n) [DFO+ 23], and thus, in
contrast with many other width measures such as treewidth, rank-width [LLO12], and twin-width
[ACH+ 24] the functionality is not maximised (up to a constant) by some choice of parameters for
the random graph.
We discuss how this result is proved in the next subsection. However, here we state the following
result on dominating sets in random bipartite graphs as we believe it may be of independent interest.
We used a stronger version of this theorem (Theorem 4.2) in our proofs.

Theorem 1.2. Let α, β > 0 be fixed constants, n be sufficiently large, and f (n) be any function
satisfying f (n) → ∞ as n → ∞. Let ln2 n ⩽ a ⩽ b ⩽ n and p ∈ (0, 1] be functions of n which satisfy
ap ⩽ α · ln n and bp > nβ . Then, for G ∼ G(A, B, p), where |A| = a and |B| = b, we have
  h i
f (n)
P ∃ a dominating set D ⊂ B of A with |D| ⩽ ∀a ∈ A, N (a) ̸= ∅ ⩾ 1 − exp −b1−o(1) .
p

We note that if one is willing to accept a less strong probability guarantee then one can exchange
f (n) for a suitably large constant C > 0 to obtain a simpler bound from Theorem 4.2, such as
 
C  
∃C s.t. P ∃ a dominating set D ⊂ B of A with |D| ⩽ ∀a ∈ A, N (a) ̸= ∅ ⩾ 1 − exp −b0.9 .
p

3
 We further observe that Theorem 1.2 is optimal in the following sense. First, w.h.p. a minimum
dominating set of A in B has size Ω(1/p) whenever ap = Θ(ln n) (see Proposition 6.3). Second, the
probability that there exists a dominating set of size O(1/p) is at most 1 − exp [−O(b)] as soon as
ap = Ω(1) (see Proposition 6.4).
We are not aware of any such result for dominating sets in random bipartite graphs, but for
related results on dominating sets in random graphs see [NS93, BWZ24, GLS15, Web81, WG01].

1.2 Proof Strategies

Obtaining order optimal bounds on functionality requires different arguments for different regimes
of p. The summary of the bounds is depicted in Figure 2. We now proceed with an outline of the
proof strategies for different bounds. We consider p in the interval [0, 1/2] as the situation in the
interval [1/2, 1] is symmetric.

General upper bound O(np) [Theorem 3.1]

via degeneracy
O( lnpn ) [Theorem 3.4]
General upper bound
via distinguishing sets    
np ln(2w) ln(2w)
O ln n
O p
Upper bounds
via dominating sets [Theorem 4.1]

0 1
q 1 1 p
ln n
n1/2+ε p∗ = n n1/2−ε 2
Lower bounds
via bad subgraph 
np ln(2w)
 
ln(2w)

Ω ln n
[Lemma 5.1] Ω p
[Lemma 5.3]

Figure 2: Upper and lower bounds on the functionality of G(n, p). The brown segments show the
intervals on which the corresponding bounds hold. The thick brown segments show the intervals
where the bounds are order optimal.

General upper bounds (small and large values of p). We first obtain two general upper
bounds that hold for the entire interval [0, 1/2]. The first bound is order optimal on the interval
[0, n−1/2−ε ], and the second bound is order optimal on [n−1/2+ε , 1/2], for any fixed ε > 0.
The first bound is based on the observation that any vertex is a function of its neighbourhood,
and therefore fun(G) does not exceed the degeneracy of G, which is the maximum of the minimum
degree of a subgraph of G. Thus, we obtain a bound of O(np) on the functionality of G(n, p) from
the same bound on its degeneracy (Lemma 2.7).
The second bound is based on the observation (Observation 2.2) that if S ⊆ V (G) is a distin-
guishing set in G, i.e., a set such that all vertices in V (G)\S have pairwise different neighbourhoods
in S, then any vertex y ∈ V (G) \ S is a function of S. More specifically, in Theorem 3.4, we show

4
that in Gn ∼ G(n, p) w.h.p. for every subset S ′ ⊆ V (Gn ) of size ⌊ 3 lnp n ⌋, the set R ⊆ V (Gn ) \ S ′
of vertices with a non-unique neighbourhood in S ′ has at most 7 ln n
p elements. This is enough to
conclude (see Lemma 3.3) that in any induced subgraph H of Gn there exists a set S of at most
10 ln n 10 ln n
p vertices such that any vertex in V (H) \ S is a function of S and thus w.h.p. fun(Gn ) ⩽ p .

Upper bounds via dominating sets (p around p∗ ). In the middle interval [n−1/2−ε , n−1/2+ε ],
the above two general approaches (degeneracy and the maximum size of a minimum distinguishing
set over all induced subgraphs) do not give order optimal bounds on functionality, and we need to
use a more refined argument to find in every induced subgraph a vertex that is a function of some
set of vertices of the appropriate size, for the sake of this discussion let us denote this size d := d(p).
To bound the functionality of a vertex y one can find a set S such that any pair of vertices
that have different adjacencies with y also have different neighbourhoods in S (i.e., the two vertices
are distinguished by S). We use this idea to obtain the upper bound on the functionality for p
around p∗ . Namely, we build such a set S of order d by uniting several dominating sets of the
neighbourhood of a vertex y. We show that such a set exists with large enough probability that it
is also present in all induced subgraphs. The main steps are:

(i) We prove a general bound on dominating sets in random bipartite graphs (Theorem 4.2).

(ii) We show that given at least five distinct sets of order d which dominate the neighbourhood of a
vertex y, w.h.p. there are only order d many non-neighbours of y that cannot be distinguished
from a neighbour of y by the union of these sets (Claim 4.3). Thus, if we take these five
dominating sets together with the vertices which are not distinguished by them, then we have
a set S of order d such that y is a function S.

(iii) We then apply the general result for domination in random bipartite graphs from (i) to find
w.h.p. a dominating set of order d of the neighbourhood of a given vertex y (Claim 4.4).

(iv) We then use this to show that there exists a vertex whose neighbourhood has at least five
distinct dominating sets of order d (Claim 4.5) with even higher probability. We achieve this
higher probability of the event, over what would be guaranteed by (iii) for a fixed vertex, by
using the fact that most neighbourhoods in our random graph do not overlap much. We do
this probability boosting to overcome a large union bound in (v).

(v) We then wrap up the proof in Section 4.2, dealing with ‘large’ and ‘small’ induced subgraphs
separately. Since the probability guarantee by (iv) is so strong, we can take a union bound
over all ‘large’ induced subgraphs to show that w.h.p. each of them contains a vertex y whose
neighbourhood is dominated by at least five distinct sets of order d, which, by (ii), can be
converted into a witness of y having functionality of order d. For ‘small’ induced subgraphs
we use an elementary bound on the degeneracy to bound the functionality.

5
Lower bounds. We obtain our lower bounds by showing the existence w.h.p. of a subgraph H
in Gn ∼ G(n, p) such that no vertex y in H is a function of a “small” set of vertices. To do so,
we show that w.h.p. there exists H such that for every “small” set S ⊆ V (H) and every vertex
y ∈ V (H) outside S there exists a neighbour z1 and a non-neighbour z2 of y in H that have no
neighbours in S. This means that S cannot distinguish z1 and z2 and therefore cannot be used to
represent different adjacencies between y and z1 , and y and z2 , i.e., y cannot be a function of S.
This strategy requires slightly different implementations in the intervals [0, p∗ ] (Section 5.1) and
[p∗ , 1] (Section 5.2). Aside from technical differences arising from differing rates of decay for p, the
main distinction is that for small p the graph H is a proper subgraph of Gn , whereas for larger p
we can take H = Gn .

1.3 Related Graph Parameters

As was already mentioned, functionality generalises a number of graph parameters, in the sense
that functionality is upper bounded by a function of these parameters, and, thus, if a parameter
is bounded in a class of graphs, then functionality is bounded too. Among these parameters are
maximum degree, degeneracy, symmetric difference, and twin-width. All of these parameters are
hereditary, i.e., their value on induced subgraphs cannot exceed the value on the graph. The
maximum degree (and even the degree distribution) of a random graph G(n, p) is well-understood
in the entire range p ∈ [0, 1] (see e.g. [Bol01, Chapter 3]). The behaviour of the recently introduced
parameter twin-width of random graphs has been studied in [AHKO22, ACH+ 24, HNST24].
Two more parameters, similar in nature to functionality, that were investigated for random
graphs are metric dimension [BMP13] and the minimum size of identifying sets [FMM+ 07]. The
metric dimension of a graph G is the minimum size of a vertex set S of G such that the vertices
outside S can be distinguished by their distances to the vertices in S. Similarly, the minimum size
of an identifying set is defined as the size of a smallest vertex set S such that the vertices outside
S are distinguished by their adjacencies to the vertices in S. The main conceptual difference
between these notions and functionality is that in the definition of the latter a “distinguishing”
set must exist in every induced subgraph, rather than only in the graph itself. This hereditary
property of functionality poses new challenges, both quantitative and qualitative, that do not arise
in the analysis of metric dimension and identifying sets. In particular, achieving a tight upper
bound necessitates overcoming the union bound over 2n induced subgraphs, which demands at least
exponentially small probability bounds for rare events. This makes the proofs of the upper bounds
especially intricate and requires establishing new tight domination results in bipartite random graphs
(Theorem 4.2).

2 Preliminaries
For a positive integer k, we denote by [k] the set {1, . . . , k}.

6
Graphs. We consider only finite, undirected graphs, without loops and multiple edges. The vertex
set and the edge set of a graph G are denoted V (G) and E(G), respectively. The neighbourhood
of a vertex x ∈ V (G), denoted N (x), is the set of vertices of G adjacent to x, and the degree of x,
denoted deg(x), is the size of its neighbourhood. The minimum vertex degree and the maximum
vertex degree in G are denoted by δ(G) and ∆(G), respectively. We denote by N [x] the closed
neighbourhood of x, that is, the set N (x) ∪ {x}. For a set S ⊆ V (G) we write NS (x) to denote
the neighbourhood of x in S, i.e., NS (x) = N (x) ∩ S. For a set S ⊆ V (G) we denote by G[S] the
subgraph of G induced by S. We write H ⊑ G to denote the fact that H is an induced subgraph G.
Given a set S ⊆ V (G) and a number t ∈ N, a set D ⊆ V (G) \ S is said to t-dominate S if every
vertex in S has at least t neighbours in D; in which case we also say that D is a t-dominating set
for S. A set D ⊆ V is a t-dominating set in G if it t-dominates V \ D. If t = 1, then we simply say
“dominates” and “dominating set” instead of “t-dominates” and “t-dominating set”, respectively.
Let G(n, p) denote the distribution on all graphs on [n] where each edge is included independently
with probability p := p(n) ∈ [0, 1], i.e., for every graph G on [n] and Gn ∼ G(n, p),

n
P(Gn = G) = p|E(G)| (1 − p)( 2 )−|E(G)| .

Similarly for two disjoint sets A, B and p := p(n) ∈ [0, 1], we let G(A, B, p) denote the random
bipartite graph where each edge with one end point in A and the other in B is included independently
with probability p.

Functionality. Let G be a graph and let A = AG be the adjacency matrix of G. We say that a
vertex y ∈ V (G) is a function of (a set of ) vertices S = {x1 , x2 , . . . , xk } ⊆ V (G) \ {y} if there is a
Boolean function f of k variables such that for every vertex z ∈ V (G) different from y, x1 , x2 , . . . , xk
we have A(y, z) = f (A(z, x1 ), . . . , A(z, xk )). Alternatively, y is a function of S if for every z ∈
V (G) \ (S ∪ {y}) the adjacency between y and z in G depends only on N (z) ∩ S.
The following observations are simple implications of definitions.

Observation 2.1. Let G = (V, E) be a graph, y ∈ V , and S ⊆ V \ {y}. Then y is a function of

S if and only if there does not exist a pair of vertices z1 , z2 ∈
/ S ∪ {y} such that NS (z1 ) = NS (z2 ),
and z1 and z2 have different adjacencies with y.

For a graph G = (V, E), a set of vertices S ⊆ V is called distinguishing set in G if no two vertices
outside S have the same neighbourhood in S, i.e., N (v) ∩ S ̸= N (w) ∩ S for any two distinct vertices
v, w ∈ V \ S.

Observation 2.2. Let G be a graph and S ⊆ V (G) be a distinguishing set in G. Then any vertex
y ∈ V (G) \ S is a function of S.

We observe that every vertex is a function of some other vertices. In particular, every vertex is
a function of its neighbours, for the function f ≡ 0, and a function of its non-neighbours, for the

7
function f ≡ 1. The minimum k such that y is a function of k other vertices is the functionality of
y and is denoted funG (y), or simply fun(y) if the graph is clear from the context. The functionality
of G is denoted and defined as follows:

fun(G) = max min fun(y), (1)

H⊑G y∈V (H)

If in this definition we replace fun(y) with deg(y), then we obtain the definition of degeneracy:

dgn(G) = max min deg(y).

H⊑G y∈V (H)

From the two definitions and the observation that fun(y) ⩽ deg(y), we immediately obtain

Observation 2.3. For any graph G it holds fun(G) ⩽ dgn(G).

Probability. For random variables Y, Z we say that Y is stochastically dominated by Z, denoted

by Y ⪯ Z, if P(Y ⩾ x) ⩽ P(Z ⩾ x) for all real x. We use Bin(n, p) for the binomial distribution
with n ⩾ 1 trials, each with success probability p ∈ [0, 1]. We now state some standard Chernoff
bounds for the binomial distribution.

Lemma 2.4 ([JLR00, Theorem 2.1]). Let n ∈ N, p ∈ [0, 1], X ∼ Bin(n, p), t > 0. Then
   
t2 t2
P (X < np − t) ⩽ exp − and P (X > np + t) ⩽ exp − .
2np 2(np + t/3)

Lemma 2.5 ([AS15, Theorem A.1.12]). Let n ∈ N, p ∈ [0, 1], X ∼ Bin(n, p), λ > 0. Then

P (X > λnp) ⩽ (e/λ)λnp .

Let Geo(p) denote the geometric distribution with success probability p ∈ (0, 1). That is, if a
random variable X ∼ Geo(p) then P (X = k) = (1 − p)k−1 p for any integer k ⩾ 1. We will also need
the following Chernoff bound for sums of geometric random variables.
Pn
Lemma 2.6 ([Jan18, Theorem 2.1]). For any n ⩾ 1 and p1 , . . . , pn ∈ (0, 1), let X = i=1 Xi ,
where Xi ∼ Geo(pi ), i ∈ [n], are independent geometric random variables. Let p∗ = mini∈[n] pi and
P
µ = E (X) = ni=1 p1i . Then for any λ ⩾ 1,

P (X ⩾ λµ) ⩽ exp [−p∗ · µ · (λ − 1 − ln λ)] .

Properties of random graphs. We require the following two standard technical results, which
we prove here for completeness.

Lemma 2.7. Let p := p(n) ∈ [0, 1] and Gn ∼ G(n, p). Then w.h.p. dgn(Gn ) ⩽ max{10np, 2}.

8
Proof. If p < 0.5/n, then w.h.p. every connected component of Gn contains at most one cycle
[JLR00, Theorem 5.5], which implies dgn(Gn ) ⩽ 2. Thus, we now assume that p ⩾ 0.5/n and show
that dgn(Gn ) ⩽ 10np.
|E(H)|
It is sufficient to prove that w.h.p., for every H ⊑ Gn , ⩽ 5np. Let us fix k ∈ [n] and
|V (H)|


consider any k-element subset V ⊂ [n]. Then the number of edges induced by V is ξ ∼ Bin( k2 , p).
By Lemma 2.5,

!5npk  
e k2 p 9n
P (ξ > 5npk) ⩽ ⩽ exp −5npk ln .
5npk ke

By the union bound, probability that there exists a subset V ⊂ [n] of size k that induces more than
5npk edges is at most
         
n ne 9n ne 9n 9n
P (ξ > 5npk) ⩽ exp k ln − 5np ln ⩽ exp k ln − 2 ln ⩽ exp −k ln .
k k ke k ke ke

Finally, by the union bound,

√
  Xn   ⌊X n⌋   n
X
|E(H)| 9n 1
P ∃V ⊂ [n] : > 5np ⩽ exp −k ln ⩽ exp − k ln n + exp [−k] ,
|V (H)| ke 2 √
k=1 k=1 k=⌈ n⌉

which is o(1), as claimed.

ln n
Lemma 2.8. Let p ≫ n and Gn ∼ G(n, p). Then w.h.p. every vertex has degree np(1 + o(1)) and
every two vertices have at most max{2 ln n, 10np2 } common neighbours in Gn .
ln n
Proof. Let w = w(n) → ∞ as n → ∞ be such that p/w ≫ n . By the Chernoff bound (Lemma 2.4),
np
a fixed vertex has degree which is at least w -far from np with probability at most exp[−Ω(np/w)].
Thus, by the union bound, w.h.p. every vertex has degree in the interval [np − np/w, np + np/w].
Observe that, for any fixed pair of vertices, the number of vertices in their common neighbourhood is


stochastically dominated by a random variable with distribution Bin n, p2 . Thus, by the Chernoff
1
bound, if np2 ⩾ 5ln n, then a fixed pair of vertices have at least 10np2 common neighbours with
 (9np2 )2  −2 ). On the other hand, if np2 < 1 ln n, then, again by
probability at most exp − 2(np 2 +3np2 ) = o(n 5
the Chernoff bound, a fixed pair of vertices has at least 2 ln n common neighbours with probability
 ln n)2 
at most exp − 2( 1 (2 2
ln n+ ln n)
= o(n−2 ). The result thus follows from the union bound.
5 3

3 General Upper Bounds

In this section we establish two general upper bounds. The first one is based on degeneracy, and
the second one is based on distinguishing sets.

Upper bound via degeneracy. The bound in Theorem 3.1 below follows immediately from
the fact that degeneracy upper bounds functionality (Observation 2.3) and the upper bound on

9
degeneracy of G(n, p) from Lemma 2.7.

Theorem 3.1. Let p := p(n) ∈ [0, 1] and Gn ∼ G(n, p). Then w.h.p. fun(Gn ) ⩽ max{10np, 2}.

Upper bound via distinguishing set. For this bound we use the notion of a distinguishing set
and the fact that any vertex outside a distinguishing set is a function of this set (Observation 2.2).
In order to use this fact to upper bound the functionality, we show that any sufficiently large induced
subgraph of G(n, p) has a distinguishing set of the desired size.
We proceed with some notation and auxiliary statements. For a graph G = (V, E) and a set
U ⊆ V we define the repetition number of U in G, denoted repG (U ), as the number of vertices
outside U that have the same neighbourhood in U as some other vertex outside U , i.e.,

repG (U ) := |{v ∈ V \ U : ∃w ∈ V \ U s.t. NU (v) = NU (w)}|.

The repetition number of the set U shows how far U is from being a distinguishing set in the
sense that U can be extended to a distinguishing set in G by adding at most repG (U ) vertices to
it. A simple, but important observation is that the repetition number of a set cannot increase in
induced subgraphs.

Observation 3.2. For any graph G, H ⊑ G, and U ⊆ V (H), we have repH (U ) ⩽ repG (U ).

We use this observation to show that if every small set of vertices in a graph has small repetition
number, then the graph has small functionality.

Lemma 3.3. For any graph G and k, r ∈ N, if repG (U ) ⩽ r for all U ⊆ V (G) with |U | = k, then
fun(G) ⩽ k + r.

Proof. To prove the lemma, we will show that any induced subgraph of G has a vertex that is a
function of at most k + r other vertices.
Consider any induced subgraph H ⊑ G and note that we can assume that V (H) ⩾ k + r + 1
or else any vertex of this subgraph is a function of at most k + r other vertices. Now, take any
set U ⊆ V (H) with |U | = k. By hypothesis we have repG (U ) ⩽ r, and thus repH (U ) ⩽ r by
Observation 3.2. It follows that there is a set R of at most r vertices, such that all vertices in
V (H) \ (U ∪ R) have distinct neighbourhoods in U , and therefore U ∪ R is a distinguishing set in H.
Hence, by Observation 2.2, any vertex in V (H) \ (U ∪ R) is a function of at most |U | + |R| ⩽ k + r
vertices.

Using this result we establish our second general upper bound.

10 log n
Theorem 3.4. Let p ∈ (0, 1/2] and Gn ∼ G(n, p). Then w.h.p. fun (Gn ) ⩽ p .

Proof. Let Gn = (V, E) ∼ G(n, p). For r, k ⩾ 0, let the event E(k, r) be given by

E(k, r) = {for all U ⊆ V satisfying |U | = k we have repGn (U ) ⩽ r}.

10
We prove the theorem by showing that for
 
3 ln n 7 ln n
k= and r= ,
p p

the event E(k, r) holds w.h.p., which together with Lemma 3.3 implies the result.
Let U ⊆ V be a set of vertices of size k. Fix an arbitrary ordering of vertices in V \ U , say
v1 , v2 , . . . , vn−k , and for every vi ∈ V \ U , define

1 if there exists some j < i satisfying NU (vi ) = NU (vj );
R(Gn , U, vi ) :=
0 otherwise.
P
Let R(Gn , U ) := vi ∈V \U R(Gn , U, vi ) and observe that this quantity is independent of the ordering
of V \ U . We claim that repGn (U ) ⩽ 2R(Gn , U ). Indeed, let C1 , C2 , . . . , Ct ⊆ V \ U be the non-
singleton equivalence classes of the vertices in V \ U , where two vertices are equivalent if and only if
they have the same neighbourhood in U . Then, every such equivalence class Ci contributes |Ci | − 1
to R(Gn , U ) and |Ci | to repGn (U ). Thus, we have

t
X t
X t
X
R(Gn , U ) = (|Ci | − 1) ⩽ repGn (U ) = |Ci | ⩽ 2(|Ci | − 1) = 2R(Gn , U ),
i=1 i=1 i=1

where the second inequality holds, because |Ci | ⩾ 2 for every i ∈ [t].
Now, the probability that vi ∈ V \ U has a specific set of j neighbours in U is equal to pj (1 −
p)k−j ⩽ (1 − p)k , where the inequality holds because p ⩽ 1/2. As the worst case is when all previous
vertices v1 , . . . , vi−1 have had distinct neighbourhoods, for each i ⩽ k,

2
P (R(Gn , U, vi ) = 1) ⩽ i · (1 − p)k ⩽ n · e−pk < .
n2


Observe also that R(Gn , U, vi ) ⪯ γi where γi ∼ Ber 2/n2 are i.i.d. Bernoulli random variables,


thus R(G, U ) ⪯ Bin n, 2/n2 . Using Lemma 2.5 with t = rn/2 > 0, for any fixed U we have

 r/2   3.5 ln n
2


4e p⩽1/2 e p
−
3.5(ln n)2
P (R(Gn , U ) ⩾ r/2) ⩽ P Bin n, 2/n ⩾ r/2 ⩽ ⩽ ⩽e p .
rn (3.5 ln n)n

We now observe that by the union bound

   
n
n 3 ln n
−
3.5(ln n)2
P (¬E(k, r)) ⩽ · P repGn (U ) ⩾ r ⩽ · P (R(Gn , U ) ⩾ r/2) ⩽ n p · e p = o(1),
k k

which concludes the proof.

11
4 Tight Upper Bound for p around p∗
Our aim in this section is to prove the following bound on functionality, which gives the right
behaviour for p close to the value p∗ , i.e. near where the functionality of G(n, p) is maximised.

Theorem 4.1. Let Gn ∼ G(n, p), where p ∈ [n−1/2−ε , n−1/2+ε ] and ε > 0 is a sufficiently small
constant. Then there exists a constant c such that w.h.p.
 q
 c · np ln(ew) , for p = ln n
, w = w(n) ⩾ 1;
ln n q wn
fun (Gn ) ⩽
 c · ln(ew) , for p = w ln n
p n , w = w(n) ⩾ 1.

We prove Theorem 4.1 by showing that, in all large enough induced subgraphs of G(n, p) there
exist a vertex y and a set D of an appropriate size that t-dominates most of the neighbourhood of
y such that there are only few non-neighbours of y that have same neighbourhoods in D as some
neighbours of y. This proof heavily relies on an auxiliary result about dominating sets in random
bipartite graphs with parts of different sizes stated below.
Recall that for two independent sets A, B and p ∈ [0, 1], the random bipartite graph G(A, B, p)
denotes the distribution over bipartite graphs on A, B where each edge ab with a ∈ A and b ∈ B,
is present independently with probability p.

Theorem 4.2. Let α, β > 0, and 0 < δ < β be fixed constants, and n be sufficiently large. Let
a, b ∈ N where ln2 n ⩽ a ⩽ b ⩽ n, and h, p ∈ R where 1 ⩽ h ⩽ nδ and p ∈ (0, 1], all be real functions
of n which satisfy aph ⩽ α · ln n and bp > nβ . Then there exists C := C(α, δ) ⩾ 1 such that in
G(A, B, p), where |A| = a and |B| = b, we have
 
′ ′ 1 ′ ln(eh)
P ∃A ⊆ A with |A | ⩽ and a dominating set D ⊂ B of A \ A with |D| ⩽ C ·
ph ph
 −δ 
n ·b
⩾ 1 − exp − .
C

We believe that this result might be of independent interest, so we state and prove it separately
in Section 6. The probability bound in this result allows us to establish necessary properties of
neighbourhoods and their dominating sets, this is done in Section 4.1. The proof of Theorem 4.1 is
completed in Section 4.2.

4.1 Neighbourhood Domination

In this section we apply our bound on the size of dominating sets (Theorem 4.2) to find witnesses
for functionality which are based on dominating the neighbours of a vertex. However, using just a
single dominating set here is not enough. For a natural number t, a set D ⊆ V \ S is t-dominating
set for S if every vertex in S has at least t neighbours in D. Thus, a 1-dominating set is just a

12
dominating set. We begin with Claim 4.3, which shows that for t ⩾ 5, w.h.p. t-dominating sets can
be used to identify neighbourhoods in a random graph when p is close to p∗ .
All results in this section hold for Gn ∼ G(n, p) with
r r
ln n ln n
⩽p⩽ (2)
n1+ε n1−ε

for some small ε > 0.

Claim 4.3. Let d ⩾ t ⩾ 5. Then, w.h.p. for every y ∈ [n], every S ⊆ N (y), and every t-dominating
set D for S of size at most d, the number of vertices w ∈
/ N (y) such that, for some z ∈ S,
ND (w) = ND (z), is at most 4d.

Proof. We will prove that w.h.p. the following stronger event B holds. This event B states that for
every y ∈ [n] and every set D of size at most d, there are at most 4d vertices w ∈
/ N (y) such that,

for some z ∈ N (y) that has at least t neighbours in D, ND (w) = ND (z).

The event B contains an intersection over many events, this makes it hard to bound P(¬B) using
a union bound. To overcome this we use a classic trick to bound difficult events. This is the more
general observation that if there exists an ‘affable’ event A and a ‘cover’ event C, both of whose
probability we can bound easily, then provided (¬B) ∩ A ⊆ C, we have

P (¬B) = P ((¬B) ∩ A) + P (¬A) ⩽ P (C) + P (¬A) . (3)

Let A be the event that every vertex has degree at most 2np, then P (¬A) = o(1) by Lemma 2.8.
We now assume that the vertex set [n] has an arbitrary but fixed ordering. Let C be the event there
exists y ∈ [n], a set D of size at most d, and at least 4d + 1 vertices w ∈
/ N (y), such that

one of the first 2np neighbours z ∈ N (y) has at least t neighbours in D and ND (w) = ND (z).

Observe that (¬B) ∩ A ⊆ C holds, as if A holds then the first 2np neighbours of y is all of N (y).
We will now bound P(C) by the union bound. To begin, the number of choices for y is n, and
n


for D is at most d ⌊d⌋ . Now, as soon as y and D are fixed and the first 2np vertices F ⊆ N (y)
are exposed, then letting ξ be the number of vertices w ∈
/ N (y) such that, for some z ∈ F ∩ N (D),
ND (w) = ND (z), we get that ξ is stochastically dominated by Bin(n, 2np · pt ). This domination
follows since there are less than n choices for w, at most 2np choices for z ∈ F , and, as each y has
at least t neighbours in D, the probability w has the same neighbourhood in D as y is at most pt .
4d
Thus, by applying Lemma 2.5 with λ = 2n2 pt+1
> 0,
   
n t+1

d+2 2en2 pt+1
P(C) ⩽ nd · P Bin(n, 2np ) > 4d ⩽ n · exp 4d ln .
⌊d⌋ 4d

13
 2en2 pt+1
Now, as p ⩽ n−1/2+ε and d ⩾ t ⩾ 5, we have 4d ⩽ n2 p6 ⩽ n−1+6ε ⩽ n−1/2 . Thus,

P(C) ⩽ exp [(d + 2) ln n − 2d ln n] = o(1).

It now follows from (3) that P(¬B) = o(1), as claimed.

q q
For nln1+ε
n
⩽ p ⩽ nln1−ε
n
, with some small ε > 0, and a constant C > 0, we define the function
 q
 C· np ln(ew)
= C · ln(ew) p = ln n
, 1 ⩽ w ⩽ nε ;
ln n pw , q wn
d := d(p, C) = (4)
 ln(ew)
C· p , p = w ln n ε
n , 1⩽w ⩽n .

We can now adapt Theorem 4.2 to the setting of our upper bound.
Claim 4.4. There exists C > 0 such that for d := d(p, C) given by (4) and any set S ⊂ [n] \ {1}
where |S| ⩽ 3np/2, we have

P ∃D ⊂ [n] \ N [1] ∃N ′ ⊂ N (1) : D is dominating for N (1) \ N ′ ; |D|, |N ′ | ⩽ d N (1) = S
⩾ 1 − exp[−n1−ε /C].

Proof. Let us expose the neighbourhood of 1 and assume that |N (1)| ⩽ 3np/2. Clearly, we may
also assume that |N (1)| ⩾ ln2 n — otherwise, we can extend the set N (1) arbitrarily
q and then
ln n
a dominating set for the extension dominates N (1) as well. The case p = wn follows from
Theorem 4.2 by taking A = N (1), B = [n] \ N [1], and h = w. To check the assumptions are
satisfied, we have ln2 n ⩽ a := |A| ⩽ 3np/2 ⩽ b, and also

3 3 ln n 3 ln n  3np 
ap ⩽ np2 = = · and bp ⩾ n − p ⩾ n1/2−ε .
2 2w 2 h 2
q
So, from now on, assume p = w ln n ε
n , where 1 ⩽ w ⩽ n . Let

ln(ew)
k= .
p

We divide [n/2] \ N [1] into roughly equal parts W1 , . . . , Wm , each of size either ⌊k⌋ or ⌈k⌉, where

n
m = (1 − o(1)) .
2k

For each j ∈ [m], let ξj be the number of vertices in N (1) that do not have neighbours in Wj . Then
ξj is a binomial random variable with

3 3  np 4 np
· np(1 − p)⌊k⌋ =
Eξj ⩽ + o(1) ⩽ · .
2 2 ew 5 w
np

We now apply Lemma 2.4 with t := 15 · np
w to bound P ξj > w , whereby stochastic domination we

14
can assume that the estimate on Eξj above is an equality. Thus,
  " #
 np  t2 1/52 np h np i
P ξj > ⩽ exp − = exp − 4 1 · ⩽ exp − .
w 2(Eξj + t/3) 2( 5 + 5·3 ) w 50w

Since all ξj are independent, and w ⩽ nε the probability that each Wj does not dominate more
np
than w vertices of N (1) is at most

h np i  
n2 p
exp − · m = exp −(1 − o(1))
50w 100wk
 
(1 − o(1))n2 p2
= exp −
100w ln(ew)
  h n i
(1 − o(1))n ln n
= exp − ⩽ exp − .
100 ln(ew) 100
np
Now, fix j ∈ [m] such that Wj dominates all but at most w vertices of N (1) and let A be the set of
2
vertices of N (1) that are not dominated by Wj . If |A| < ln n then add some arbitrary vertices from
N (1) until |A| ⩾ ln2 n. We can now apply Theorem 4.2 to this A with B = [n] \ ([n/2] ∪ N [1]) and
np
h = 1, since aph ⩽ w ·p·1 = ln n and bp ⩾ n1/2 , and we can take δ = ε/2 > 0. This shows that there
exists C > 0 such that with probability at least 1−exp[−n−δ b/C] ⩾ 1−exp[−(1/2−o(1))n1−ε/2 /C],
there exist D ⊂ [n] \ ([n/2] ∪ N [1]) of size at most C/p and N ′ ⊂ A of size at most 1/p such that D
dominates A \ N ′ . The set D ⊔ Wj is the desired set (called D in the statement), which dominates
N (1)\N ′ where |N ′ | ⩽ 1/p ⩽ d, and this holds with the desired probability by the union bound.

Let Q be the event that there are at least n0.2 vertices among the first ⌈n0.6 ⌉ vertices of Gn that
have degrees at most 3np/2. For a positive number x, let Q′ (x) be the event that any two of the
first ⌈n0.6 ⌉ vertices in Gn have at most x neighbours in common.
Claim 4.5. Let x = o(n0.7 · p). Then, there exists a C > 0 such that for d := d(p, C) given by (4),
any t ∈ Z>0 , and

A = ∃y ∃D ⊂ [n] \ N [y] ∃N ′ ⊂ N (y) : D t-dominates N (y) \ N ′ ; |D|, |N ′ | ⩽ 2td , (5)

we have P ((¬A) ∩ Q ∩ Q′ (x)) ⩽ exp[−n1.1 ].

Proof. Let Q′ := Q′ (x). For the event A given by (5) we have

P((¬A) ∩ Q ∩ Q′ ) ⩽ P(¬A | Q ∩ Q′ ).

Thus, it suffices to show that P(¬A | Q ∩ Q′ ) ⩽ exp[−n1.1 ], which we will do.

Let r = ⌈n0.6 ⌉ and s = ⌈n0.2 ⌉. Expose the neighbours of vertices 1, . . . , r. If Q holds, then
assume without loss of generality that the sets N (1), . . . , N (s) each have size at most 3np/2. Let

15
 W
N (y)

[r] W1

y
e (y)
N

W2
[s]

e (z)
N

W3

N (z)

Figure 3: Illustration of the proof of Claim 4.5. The vertex y ∈ [s] has a 3-dominating set for two
of the vertices in its neighbourhood.

W = [n] \ (N (1) ∪ . . . ∪ N (s) ∪ [r]). Partition W into t parts W1 , . . . , Wt with sizes as equal as
possible. Then each Wℓ , for ℓ ∈ [t], has nt (1 − o(1)) vertices, see Figure 3.
e (i) to be the part of the i’s neighbourhood not contained N (j) for any
For i ∈ [s], we define N
other j ̸= i ∈ [s], that is
[
e (i) := N (i) \
N N (j).
j∈[s],j̸=i

Observe that conditional on Q ∩ Q′ we have |N e (i)| = (1 + o(1))|N (i)| for all i ∈ [s]. Additionally,
for i ∈ [s], let Ei be the event that all the sets W1 , . . . , Wt contain a dominating set D of Ne (i) \ N ′
e (i) where |D|, |N ′ | ⩽ d. Take S1 , . . . , Ss ⊂ [n] to be any sets such that the event
for some N ′ ⊂ N

NS1 ,...,Ss := {N (1) = S1 , . . . , N (s) = Ss }

implies Q ∩ Q′ . To bound P (¬Ei | NS1 ,...,Ss ) we will first bound the probability that, for a given
set Wℓ , where ℓ ∈ [t], there exist the desired sets D, N ′ such that D ⊂ Wℓ dominates N e (i) \ N ′ .
If we take C > 0 suitably large, then it follows from applying Claim 4.4h to the random graph i on
e ′ ((1−o(1))n/t)1−ε
{i} ∪ N (i) ∪ Wℓ that the probability no such D, N exist is at most exp − C . Thus

16
by the union bound, we have that for any i ∈ [k]
h ((1 − o(1))n/t)1−ε i
P (¬Ei | NS1 ,...,Ss ) ⩽ t · exp − ⩽ exp[−n0.9 ].
C

e (i)| ⩽ s · x = o(n0.2 · n0.7 p) = o(d) for all i ∈ [s], as ε > 0

Observe that if Q′ holds then |N (i) \ N
is small. Thus, if additionally Ei holds, for some i ∈ [s], then, by adding up all the dominating sets
from each Wℓ , we have a t-dominating set of N (i) of size t|D| ⩽ td, and we can add |N (i) \ N e (i)|
along with all the sets N ′ to give a subset N ′′ ⊆ N (i) (this is N ′ in the statement) of vertices that
e (i)| ⩽ 2td. Note that for any i, j ∈ [s], conditional on
are not covered with |N ′′ | ⩽ t|N ′ | + |N (i) \ N
e (i) ∩ N
the event NS1 ,...,Ss , the events ¬Ei and ¬Ej are independent as N e (j) = ∅, i.e. the overlapping
parts of the neighbourhoods of i and j are not considered within these events. Thus,

P(¬A | Q ∩ Q′ ) ⩽ P(¬E1 ∧ . . . ∧ ¬Es | Q ∩ Q′ )

X P(¬E1 ∧ . . . ∧ ¬Es | NS ,...,S )P(NS ,...,S )
1 s 1 s
=
P(Q ∩ Q′ )
S1 ,...,Ss
X P(¬E1 | NS ,...,S ) . . . P(¬Es | NS ,...,S )P(NS ,...,S )
= 1 s 1 s 1 s
⩽ exp[−s · n0.9 ].
P(Q ∩ Q′ )
S1 ,...,Ss

This completes the proof.

4.2 Proof of Theorem 4.1

Let ε > 0 be sufficiently small so that for all p satisfying (2), all statements from Section 4.1
hold true. We can assume w.l.o.g. that ε < 0.001, and then fix the epsilon in the statement of
Theorem 4.1 to be ε/8. Thus in terms of our ε > 0, we have

p ∈ [n−1/2−ε/8 , n−1/2+ε/8 ]. (6)

Now, let t = 5 and C > 0 be a sufficiently large constant, and for this C let d := d(p, C) be as given
by (4). To bound the functionality of Gn ∼ G(n, p) from above we must bound the functionality in
all induced subgraphs. We break into two cases depending on k, the size of the induced subgraph.

Large (k ⩾ n1−ε/3 ). We denote by A the event from Claim 4.5, given by (5). For U ⊂ [n], let QU
be the event that there are at least |U |0.2 vertices among the first ⌈|U |0.6 ⌉ vertices of Gn [U ] that

have degrees at most 3|U |p/2, i.e. QU = Gn [U ] ∈ Q , recalling that Q and Q′ (x) were defined in

Section 4.1. In the same way, we set Q′U = Gn [U ] ∈ Q′ (k 0.65 · p) and AU = {Gn [U ] ∈ A}. We
also let Q′′ be the event that any two vertices in Gn have at most n0.6 · p neighbours in common.
Notice that Q′′ implies Q′U for any U with |U | ⩾ n1−ε/3 . Finally, since n0.6 p ⩾ max{2 ln n, 10np2 },
Lemma 2.8 implies that
P(¬Q′′ ) = o(1). (7)

17
 Our aim in this case is to show that w.h.p. for all U with |U | ⩾ n1−ε/3 , in the induced subgraph
Gn [U ] there is y ∈ U , D ⊂ U \ N [y] and some N ′ ⊂ N (y) such that D t-dominates N (y) \ N ′ and
|D|, |N ′ | ⩽ 2td, where d = d(p(n), C). That is, w.h.p. for all such U the event Gn [U ] ∈ A holds.
We will achieve this by applying a union bound over such sets U , then use Claim 4.5 to bound
the probability Gn [U ] ∈ A. There is a small issue to overcome here as each Gn [U ] ∼ G(|U |, p(n)),
however Claim 4.5 must be applied to a random graph from G(n, p(n)). Let Gk ∼ G(k, p(n)), where
n1−ε/3 ⩽ k ⩽ n. Note that w.l.o.g. k is a bijective function of n, so there is a function p′ (k) ∈ [0, 1]
such that, for each n ∈ N and k = k(n), p′ (k) = p(n) and thus for this p′ we have Gk ∼ G(k, p′ (k)).
1
As n1−ε/3 ⩽ k ⩽ n corresponds to k ⩽ n ⩽ k 1−ε/3 and we can assume ε < 0.001, we have
r  1 −1/2−ε/4 r
ln k − 12 − 12−4ε
5ε ln k
⩽ k = k 1−ε/3 ⩽ p′ (k) = p(n) ⩽ k −1/2+ε/4 ⩽ .
k 1+ε k 1−ε

Since ε > 0 is small enough, we can apply Claim 4.5 to Gk ∼ G(k, p′ (k)) and obtain

P(Gk ∈ ¬A ∩ Q ∩ Q′ ) ⩽ exp[−k 1.1 ]. (8)

The bound has a good rate of decay, however the event Gk ∈ A gives |D|, |N ′ | ⩽ 2td(p′ (k), C),
whereas we want |D|, |N ′ | ⩽ 2td(p(n), C). The following claim shows this is not an issue.
Claim 4.6. Let C > 0 and d := d(p, C) be given by (4). Then for any p = p(n) ∈ [0, 1], all n large
enough, and n1−ε/3 ⩽ k ⩽ n,
d(p′ (k), C) ⩽ d(p(n), C).
q
Proof. First suppose that p(n) = = p′ (k) for some 1 ⩽ w ⩽ nε . Then, since we can assume
ln n
wn
q
ln n ln k ′ ′ ln k ln(ex)
k, n are large, n ⩽ k and thus there is a w ⩾ w ⩾ 1 such that p (k) = w ′ k . Now since x
is decreasing in x ⩾ 1, in this case,

C ln(ew′ ) C ln(ew)
d(p′ (k), C) = ′ ′
⩽ = d(p(n), C).
p (k)w p(n)w
q
Otherwise, if p(n) = w ln n
n
= p′ (k) for some 1 ⩽ w ⩽ nε , then there exists a n−ε ⩽ w′ ⩽ w such
q
′
that p′ (k) = w kln k . First suppose that 1 ⩽ w′ ⩽ w, then

C ln(ew′ ) C ln(ew)
d(p′ (k), C) = ′
⩽ = d(p(n), C).
p (k) p(n)
q
Otherwise, n−ε ⩽ w′ < 1, and then p′ (k) = ln n
(1/w′ )n , where 1/w′ > 1. Therefore,

C ln(e/w′ ) (⋆) C (†) C ln(ew)

d(p′ (k), C) = ⩽ ′ ⩽ = d(p(n), C),
p′ (k)/w′ p (k) p(n)

18
 ln(ex)
where (⋆) holds since x is decreasing in x ⩾ 1, and (†) holds since w ⩾ 1.

Now, by (7) and the above discussion around Gk ∈ A vs Gn [U ] ∈ AU , we have

P(∃U ⊂ [n] : |U | ⩾ n1−ε/3 and ¬AU )

⩽ P(¬Q′′ ) + P(∃U ⊂ [n] : |U | ⩾ n1−ε/3 and ¬AU ∩ Q′′ )
⩽ o(1) + P(∃U ⊂ [n] : |U | ⩾ n1−ε/3 and ¬AU ∩ Q′U )
X n 
′
⩽ o(1) + P(Gk ∈ ¬A ∩ Q ∩ Q ) + P(Gk ∈ / Q) . (9)
1−ε/3
k
k⩾n

We are now in a position to apply (8), but we will also need to following to bound the other
probability on the RHS of (9).
/ Q) = o(2−n ).
Claim 4.7. For every k ⩾ n1−ε/3 , P(Gk ∈

Proof. Recall that we assume ε < 0.001. Let S be the set of the first ⌈k 0.6 ⌉ vertices in [k]. By
Lemma 2.4, with probability 1−o(2−n ) there are o(k 1.6 p) edges in Gk [S]. Similarly, with probability
1 − o(2−n ) there are at most 1.1k 1.6 p edges between S and [k] \ S. However, the event Gk ∈
/ Q
implies that almost all vertices in U have degrees at least 32 kp in Gk , which is only possible when
either Gk [U ] has more than 0.1k 1.6 p edges or there are more than 1.1k 1.6 p edges between S and
[k] \ S. This completes the proof.

Now, inserting (8) and Claim 4.7 into (9) gives

X n
1−ε/3
P(∃U ⊂ [n] : |U | ⩾ n and ¬AU ) ⩽ o(1) + exp[−k 1.1 ]
k
k⩾n0.99
X  
⩽ o(1) + 2n exp −n0.99·1.1 = o(1).
k⩾n0.99

So, w.h.p. in any induced subgraph H of size at least n1−ε/3 , there exists a vertex y, a set
N ′ ⊂ N (y) ∩ V (H) of size at most 2td, and a t-dominating set of N (y) ∩ V (H) \ N ′ of size at most
2td. Let us fix such a set D := D(H; y) ⊂ V (H), which t-dominates the set S := N (y) ∩ V (H) \ N ′ .
By Claim 4.3, we can assume that the number of vertices w in H that do not belong to N (y) such
that, for some z ∈ S, ND (w) = ND (z), is at most 4 · 2td. Denote the set of these vertices by B and
note that, by Observation 2.1, y is a function of D ∪ N ′ ∪ B. Thus, w.h.p.

max min fun(y) ⩽ 2td + 2td + 8td = 60d.

H⊏Gn : V (H)⩾n1−ε/3 y∈V (H)

Small (k ⩽ n1−ε/3 ). For this case we will prove the stronger bound: w.h.p.,

max min fun(y) ⩽ n1/2−ε/7 < d.

H⊏Gn : V (H)<n1−ε/3 y∈V (H)

19
This follows immediately from the next claim, completing the proof of Theorem 4.1.
Claim 4.8. Let ε > 0 be sufficiently small, n−1/2−ε/8 ⩽ p ⩽ n−1/2+ε/8 , and Gn ∼ G(n, p). Then,
w.h.p., every H ⊏ Gn with V (H) < n1−ε/3 has degeneracy at most n1/2−ε/7 .
n1/2−ε/7 k
Proof. Let k < n1−ε/3 . Then, by Lemma 2.5 with λ = k2 p
, we have

     
ek 2 p
P |E(Gk )| > n1/2−ε/7 k ⩽ P Bin(k 2 , p) > n1/2−ε/7 k ⩽ exp n1/2−ε/3 k · ln .
n1/2−ε/7 k

Then, as
    
ekp ε 1 ε 1 ε 11
ln ⩽1+ 1− − − − − ln n + 1 = 1 − ε · ln n,
n1/2−ε/7 3 2 8 2 7 168

for large n we have

  h ε i
P |E(Gk )| > n1/2−ε/7 k ⩽ exp − · n1/2−ε/7 k ln n .
20


By the union bound and as nk ⩽ exp(k ln n), the probability that there exist H ⊏ Gn with
V (H) < n1−ε/3 and degeneracy at least n1/2−ε/7 is at most

1−ε/3   1−ε/3
nX
n h ε i nX h ε i
1/2−ε/7
exp − · n k ln n ⩽ exp k ln n − · n1/2−ε/7 k ln n
k 20 20
k=1 k=1
1/2−ε/7
= n−(ε/20+o(1))·n = o(1),

which completes the proof since functionality is at most degeneracy by Observation 2.3.

5 Lower Bounds
By Observation 2.1, a vertex y ∈ V (G) is not a function of a set S ⊆ V (G) \ {y} if there exists a
pair of vertices z1 , z2 ∈
/ S ∪ {y} such that z1 and z2 have the same neighbourhood in S but different
adjacencies with y.
We apply this idea to prove our lower bounds by showing that such a pair z1 , z2 exists for all
such y, and sets S of a certain size. We split the proof into two cases for small and large p, Sections
5.1 and 5.2 respectively, as the way in which this idea is applied is dependent on p.

5.1 Lower Bound for p ⩽ p∗

q
ln n
Lemma 5.1. Let Gn ∼ G(n, p), where p = wn , w = w(n) ⩾ 1. Then w.h.p.
 
np ln(ew)
fun (Gn ) ⩾ .
20 ln n

20
 np ln(ew)
Proof. Without loss of generality we can assume that np ⩾ 19, since otherwise 20 ln n < 1, and so
the statement holds trivially. This can be seen for w ⩽ n1.02 as ln(ew) becomes sufficiently small,
and for w > n1.02 as the numerator np ln(ew) becomes sufficiently small. Note that the assumption
np ⩾ 19 implies that w ⩽ n ln n.
Consider the following process. Iteratively, remove from Gn vertices of degree at most np/5. Let
k be the number of removed vertices, and Hn be the remaining graph on n0 := n − k vertices. Let
n np ln n o
D := ∆ ⩽ max{ln n, 4np} ⩽ ,
19

and define the following two events

 
A := n0 ⩾ n/ ln n ∩ D, and Alarge := n0 = n ∩ D, (10)

and note that Alarge ⊆ A. We now bound the probability of these events.
4 ln n
Claim 5.2. P(A) = 1 − o(1). Furthermore, if n ⩽ p ⩽ 12 , then P(Alarge ) = 1 − o(1).

Proof of Claim 5.2. Due to the union bound and Chernoff bound, w.h.p. D holds.
The second part of the claim is straightforward since all degrees in G(n, p) are at least np/5
4 ln n
w.h.p. when p ⩾ n by the union bound and Chernoff bound (Lemma 2.4).
For the first part, observe that the total number of edges in Gn is at most knp/5 + n0 ∆, where


∆ is the maximum degree in Gn . Thus, since w.h.p. |E(Gn )| = n2 p(1 + o(1)), we have that w.h.p.
 
n knp 1
p(1 + o(1)) ⩽ + n0 ∆ ⩽ np (k + n0 ln n) .
2 5 5

Thus, for n large enough, we have 2n ⩽ k + n0 ln n, which together with the fact that k ⩽ n implies
the desired n0 ⩾ n/ ln n. ♢

In the rest of the proof we will show that w.h.p. the functionality of Hn is at least
 
np ln(ew)
m := .
20 ln n

By definition, this will imply the desired bound for the functionality of Gn . To do so we will use
Observation 2.1 to prove that for any set S ⊆ V (Hn ) with |S| = m and any vertex y outside S,
the vertex y is not a function of S. Specifically, we will show that for any such S and y there exist
z1 , z2 ̸∈ S ∪ {y} such that z1 and z2 have different adjacencies with y, and neither z1 nor z2 has
neighbours in S.
Fix S ⊂ [n] of size m, y ∈ [n], and denote by N the neighbourhood of y in Hn (though we are
only interested in S ⊂ V (Hn ) and y ∈ V (Hn ), we want to treat S and y as deterministic objects,
thus we choose them from the entire set of vertices). Assuming that y belongs to Hn and A holds,

21
 np ln n  
we have |V (Hn )| ⩾ n/ ln n, and ∆ ⩽ 19 , and since m ⩽ (1 + o(1)) np
20 we get that

np np
|N \ S| ⩾ −m⩾ , (11)
5 7

and
n np ln n n
|V (Hn ) \ (N ∪ S ∪ {y})| ⩾ − − m − 1 = (1 − o(1)) . (12)
ln n 19 ln n
Letting Z ⊆ [n] \ S be any set, we have

P(every z ∈ Z has a neighbour in S) = (1 − (1 − p)m )|Z| ⩽ (mp)|Z| . (13)

q
19 4 ln n 4 ln n ln n
Let us now separately consider two cases: n ⩽p< n and n ⩽p⩽ n .

19 4 ln n
Small p. Let n ⩽p< n . For this range of p, we have w ∈ [n/(16 ln n), (n ln n)/361], and thus

j np k (ln n)2
m = (1 + o(1)) and mp ⩽ . (14)
20 n

We start by showing P(B1 ) = 1 − o(1), where B1 is the event that for every y ∈ [n], S ⊂ [n] \ {y}
of size m, and N ⊂ N (y) of size at least np/5, there exists z1 ∈ N \ S with no neighbours in S.
We will use the trick (3) from Claim 4.3, and define an event C1 satisfying (¬B1 ) ∩ A ⊆ C1 .
It then suffices to show P(C1 ) = o(1), as P(¬A) = o(1) by Claim 5.2. To define C1 we first fix
an ordering of [n]. Now assign, for each vertex y, a set Fy containing the first (at most) ⌈16 ln n⌉
vertices of N (y) and, if |N (y)| < ⌈np/5⌉, then Fy also contains the first ⌈np/5⌉ − |N (y)| vertices in
order that are not present in N (y). Let C1 be the event that there exist y ∈ [n], S ⊂ [n] \ {y} of
size m, and N ⊂ Fy of size exactly ⌈np/5⌉, where every vertex in N \ S has a neighbour in S. Now,
by (11), (13), and the union bound,
   
n ⌈16 ln n⌉
np/7
P(C1 ) ⩽ n · · · (mp)np/7 ⩽ nm+1 mp(ln n)3 . (15)
m ⌈np/5⌉

Since np ⩾ 19, due to (14), for large n we have that

 19/14
m+1


3 np/7 m+1 (ln n)2
np/14
np/14
n mp(ln n) ⩽n · (ln n)3 · mp(ln n)3 ⩽ nm mp(ln n)3 .
n

1
Thus, again, as nm ⩽ exp[ 19 np ln n] due to (14), the desired probability satisfies
   
1 1
1 1 − o(1)
P(C1 ) ⩽ exp np ln n + np ln mp(ln n)3 = exp np ln n − np ln n = o(1). (16)
19 14 19 14

Next we show P(B2 ) = 1 − o(1), where B2 is the event that for every set U ⊂ [n] of size at
least n/ ln n, y ∈ U , S ⊂ U \ {y} of size m, and N ⊂ N (y) ∩ U of size at least np/5 there exists

22
z2 ∈ U \ (N ∪ S ∪ {y}) with no neighbours in S. We show this by setting up a similar event C2 using
sets Fy . The event C2 states that there exist U ⊂ [n] of size at least n/ ln n, y ∈ U , S ⊂ U \ {y} of
size m, and N ⊂ Fy ∩ U of size exactly ⌈np/5⌉, such that every vertex in U \ (N ∪ S ∪ {y}) has a
neighbour in S. Indeed, from (12) and (13), by the union bound,
   
n ⌈16 ln n⌉
P(C2 ) ⩽ 2n · n · · · (mp)(1−o(1))n/ ln n .
m ⌈np/5⌉
Observe that bound on P(C2 ) contains the first bound (15) we obtained on P(C1 ) as a factor. Thus
applying the bound (16) that we have just obtained for (15) to this factor yields

(14)
(1−o(1))n/ ln n
P(C2 ) ⩽ 2n · o(1) · (mp)(1−o(1))n/ ln n−np/7 ⩽ 2n (ln n)2 /n = 2n e−(1−o(1))n = o(1).

Finally, let B be the event that for every choice of y ∈ V (Hn ) and S ⊂ V (Hn ) \ y of size m, there
exist z1 ∈ N (y) ∩ V (Hn ) \ S and z2 ∈ V (Hn ) \ (N (y) ∪ S ∪ {y}) that both do not have neighbours in
S. By Observation 2.1 and the definition of functionality, if B holds then fun(Gn ) ⩾ fun(Hn ) ⩾ m
as desired. However, B ⊇ B1 ∩ B2 and thus P(¬B) ⩽ P(C1 ) + P(C2 ) + P(¬A) = o(1).
q q
4 ln n ln n ln n
Large p. Let n ⩽p⩽ n . In this case w ∈ [1, n/(16 ln n)], and thus as p = wn , we have

np ln(ew) ln(ew)
m = (1 + o(1)) and mp ⩽ (1 + o(1)) . (17)
20 ln n 20w

Similarly to before, we give an arbitrary order to the vertices of Gn and assign for each vertex y, a
set Fy such that either Fy = N (y) if |N (y)| ⩾ ⌈np/5⌉ or, otherwise, Fy contains all vertices from
N (y) and also the first ⌈np/5⌉ − |N (y)| vertices in order that are not present in N (y). Let C be the
event that there exist y ∈ [n] and S ⊂ [n] \ {y} of size m such that either every z1 ∈ Fy \ S has a
neighbour in S, or every z2 ∈ [n] \ (Fy ∪ S ∪ {y}) has a neighbour in S. We now bound P(C) by
the union bound, using the fact that both Fy and [n] \ (Fy ∪ S ∪ {y}) have size at least np/5 while
m < np/19, which gives
 
n
P(C) ⩽ 2 · n · · (mp)np/7 ⩽ nm+1 (mp)np/7 .
m

Now, by (17) and using the inequality (∗) that 20w > (ln(ew))2 for all w ⩾ 1, we have

(17)

np ln(ew) np ln(20w/ ln(ew))
P(C) ⩽ exp (1 + o(1)) − (1 + o(1))
20 7
(∗)
  
1
⩽ exp + o(1) (7np ln(ew) − 10np ln(20w)) = o(1). (18)
140

Let B be the event that for every choice of y ∈ [n] and S ⊂ [n] \ y of size m, there exist
z1 ∈ N (y) \ S and z2 ∈ [n] \ (N (y) ∪ S ∪ {y}) that both do not have neighbours in S. Again, B

23
implies that fun(Gn ) ⩾ m. However (¬B) ∩ Alarge ⊆ C, and so P(¬B) = o(1) by (18) and Claim 5.2,
as desired.

5.2 Lower Bound for p ⩾ p∗

q
w ln n
Lemma 5.3. Let Gn ∼ G(n, p), where p = n ⩽ 12 , w = w(n) ⩾ 1. Then w.h.p.

ln(ew)
fun (Gn ) ⩾ .
4p

Proof. It is sufficient to prove that the following event B holds w.h.p.: for every set S of size
 
ln(ew) 1
k := > (ln n − ln2 ln n)
4p 2

and every vertex y ∈

/ S, there exist two vertices z1 , z2 ∈ [n] \ (S ∪ {y}) with no neighbours in S,
where z1 is adjacent to y, while z2 is non-adjacent to y. Note that ¬B ⊆ E1 ∪ E2 , where

E1 is the event that there exist S ⊂ [n] of size k and y ∈

/ S such that every neighbour of y in
[n] \ (S ∪ {y}) has a neighbour in S;

E2 is the event that there exist S ⊂ [n] of size k and y ∈

/ S such that every non-neighbour of y in
[n] \ (S ∪ {y}) has a neighbour in S.

Bound on P (E1 ). Let us fix S ⊂ [n] of size k and a vertex y ∈

/ S. For every z1 ∈ [n] \ (S ∪ {y}),
the probability that z1 is adjacent to y and N (z1 ) ∩ S = ∅ is exactly p(1 − p)k . Then

P (∄z1 ∈ [n] \ (S ∪ {y}) : N (z1 ) ∩ (S ∪ {y}) = {y}) = (1 − p(1 − p)k )n−k−1 ,

and by the union bound

nk h i
P (E1 ) ⩽ · n · (1 − p(1 − p)k )n−k−1 ⩽ exp k ln n − (1/2 − o(1))pn(1 − p)k . (19)
k!

Bound on P (E2 ). Let R be the event that every vertex has at least n−np(1+o(1)) ⩾ n(1/2−o(1))


non-neighbours, and note that P(¬R) = o(1) by Lemma 2.8. Fix S ′ ∈ [n] ′ / S ′ . Then, by
k and y ∈
the union bound and since k! > n, for sufficiently large n, P (E2 ) is bounded from above by
   
[n]
P ∃S ∈ ∃y ∈
/ S ∀z2 ∈ / (S ∪ {y}) : N (z2 ) ∩ (S ∪ {y}) ̸= ∅
k
 
n

⩽ · n · P (n − |N [y ′ ]| ⩾ n(1/2 − o(1))) ∧ (∀z2 ∈ [n] \ (N [y ′ ] ∪ S) : N (z2 ) ∩ S ′ ̸= ∅) + P(¬R)
k
 
⩽ nk · P Bin(n(1/2 − o(1)), (1 − p)k ) = 0 + o(1).

24
Thus we have
 n(1/2−o(1)) h i
P (E2 ) ⩽ nk 1 − (1 − p)k + o(1) ⩽ exp k ln n − (1/2 − o(1))n(1 − p)k + o(1). (20)

Bound on P (¬B). From (19) and (20), we get that

h i
P (¬B) ⩽ P (E1 ) + P (E2 ) ⩽ 2 exp k ln n − (1/2 − o(1))pn(1 − p)k + o(1). (21)

We now show that the exponent on the right-hand side of (21) goes to −∞ as n goes to +∞.
Claim 5.4. Let ρ(n) := k ln n − (1/2 − o(1))pn(1 − p)k . Then, ρ(n) = −ω(1).

Proof of Claim 5.4. Set δ := 1/10. First, let n−δ ⩽ p ⩽ 1/2. Since φ(p) := ln(1 − p) + 2p ln 2
increases on (0, 1 − 1/(2 ln 2)), decreases on (1 − 1/(2 ln 2), 1/2), φ(0) = 0, and φ(1/2) = 0, φ(p) is
ln(1−p)
non-negative on [0, 1/2] and thus 2p ⩾ − ln 2 for p ∈ [n−δ , 1/2]. Therefore,
  
δ 2 1−δ ln(ew)
ρ(n) ⩽n ln n − (1/2 − o(1))n · exp + 1 · ln(1 − p)
4p
 
δ 2 1−δ ln 2
⩽n ln n − (1/2 − o(1))n · exp − · ln (ew) + ln(1 − p)
2
   
δ 2 1−δ ln 2
⩽n ln n − n · exp − + o(1) ln n
2
⩽n2δ − n1−δ−ln 2/2−o(1) = −ω(1),

large n.
for suitablyq
Now, if lnnn ⩽ p < n−δ , then,

ln(ew) 1
ρ(n) ⩽ (1 + o(1)) ln n − np exp [−(1/4 + o(1)) ln(ew)] .
4p 2

Then recalling that w = np2 /(ln n) and noting that ln x − (2/e) · x5/7 decreases on [e, ∞), we have

(1 + o(1)) ln(ew) − 2w · (ew)−1/4−o(1)

ρ(n) ⩽
4p/ ln n


(1 + o(1)) ln(ew) − 2w · (ew)−2/7
⩽
4p/ ln n


(1 + o(1)) ln e − (2/e) · e5/7
⩽
4p/ ln n


(1 + o(1)) nδ · ln n
⩽− = −ω(1),
8

as claimed. ♢

Finally, by (21) and Claim 5.4, we have P (¬B) = o(1), which completes the proof.

25
6 Domination in Random Bipartite Graphs
Let G = (V, E) be a graph and S ⊆ V . We say that a set D ⊆ V \ S is a dominating set for S if
every vertex in S has a neighbour in D. Recall that for two independent sets A, B and p ∈ [0, 1],
the random bipartite graph G(A, B, p) denotes the distribution over bipartite graphs on A, B where
each edge ab with a ∈ A and b ∈ B, is present independently with probability p.
The following result (Theorem 4.2) bounds dominating sets in the binomial random graph, and,
as mentioned in the introduction, this bound is tight. Theorem 4.2 is a key ingredient in the proof of
the upper bound on functionality, and as it is written primarily with this task in mind, the statement
is a little opaque. After proving Theorem 4.2 we will show in Section 6.1 how Theorem 1.2, the
simplified version of Theorem 4.2, follows from Theorem 4.2.

Theorem 4.2. Let α, β > 0, and 0 < δ < β be fixed constants, and n be sufficiently large. Let
a, b ∈ N where ln2 n ⩽ a ⩽ b ⩽ n, and h, p ∈ R where 1 ⩽ h ⩽ nδ and p ∈ (0, 1], all be real functions
of n which satisfy aph ⩽ α · ln n and bp > nβ . Then there exists C := C(α, δ) ⩾ 1 such that in
G(A, B, p), where |A| = a and |B| = b, we have
 
′ ′ 1 ′ ln(eh)
P ∃A ⊆ A with |A | ⩽ and a dominating set D ⊂ B of A \ A with |D| ⩽ C ·
ph ph
 −δ 
n ·b
⩾ 1 − exp − .
C

Proof. We will now sketch the proof, which uses a greedy algorithm to find a subset of vertices in
B that dominates most of A. To begin, we assign an arbitrary ordering σ to the vertices of B. The
algorithm then proceeds in rounds, where in round i ⩾ 1, for “thresholds” ki and ℓi to be defined,
we expose the edges from vertices in B vertex-by-vertex, following σ, and we add v ∈ B to the
dominating set D if it has more than ki neighbours in A that are not yet dominated. The round
ends when there are less than ℓi vertices of A that are not yet dominated. The thresholds ki and
ℓi are both non-increasing in i, but ki decays at a slower rate than ℓi . This algorithm terminates
when there are at most 1/(ph) vertices of A which are not dominated, or if we have exposed all
edges incident with vertices in B. Assuming we do not exhaust B, the set D will contain the desired
number of vertices by definition of ki and ℓi . Thus, if the algorithm terminates before we have “run
out” of vertices in B, then we have achieved our goal. Our task is to bound the probability of this
event.
Notice that by the conditions ln2 n ⩽ a and aph ⩽ α ln n we have

α ln n α
ph ⩽ = . (22)
ln2 n ln n

Further, observe that if a ⩽ 1/ph, then we can set A′ = A and the theorem holds trivially. Thus,
we can assume that
aph > 1. (23)

26
We now set, with foresight, the constant
 
20α
c := max , 1 , (24)
δ
and define the “thresholds” ki , ℓi in the following way: ℓ0 := a, and, for i ⩾ 1,
 
aph  
ki := and ℓi := a · e−i·c . (25)
ic2 ln(2h)

Consider the setting where, given our ordering σ of B, we extend σ to an ordering of a countably
infinite set U — thus B is set of the first b elements of U in this ordering. We are then considering
running the algorithm on a random bipartite graph G(A, U, p). Before starting the algorithm, colour
each vertex of U green. In the i-th round of the algorithm, where i ⩾ 1, we initialise Xi = ∅. We
then process vertices one-by-one by choosing the first green vertex from U in the ordering σ, say
v, and first colour it red. We then expose N (v) and add v to Xi if |N (v) \ N (Di )| ⩾ ki , where
S
Di = 1⩽j⩽i Xj . The i-th round ends when |A \ N (Di )| < ℓi , and we let τi ⩾ 0 denote the number
of vertices in U which changed colour from green to red during the i-th round. The algorithm ends
after some (random) number T of rounds such that |A \ N (DT )| ⩽ 1/(ph), and we let D := DT .
As we shall see shortly (Claim 6.1), ki ⩽ ℓi holds for large n and all i ∈ [T ] almost surely, and
thus, since p > 0, this algorithm terminates almost surely. We say that the algorithm is successful
P
if the number of vertices of U which are turned red is at most b, that is the event { Ti=1 τi ⩽ b}
holds. In this case we find the desired set D ⊆ B.
We now prove three bounds, one for T , and two which relate ki to ℓi . To begin, observe that

1 l m l 1 m 1
−((1/c)·ln(2aph)+1)·c −⌈(1/c)·ln(2aph)⌉·c −⌈(1/c)·ln(2aph)⌉·c
= a·e ⩽ a·e ⩽ a·e ⩽ ⩽ , (26)
2ec ph 2ph ph

since 1/(ph) = Ω(ln n) by (22), aph > 1 by (23), and n is suitably large. Thus, from the definition
of T and the definition of ℓi given in (25), it follows that

T ⩽ T ′ := ⌈(1/c) · ln(2aph)⌉ almost surely. (27)

Therefore, almost surely, for all 1 ⩽ i ⩽ T , we have ℓi = ω(1) and so from (25) we obtain

ic2 ln(2h) −i·c

p · ℓi ⩽ p · ℓT ′ ⩽ (1 + o(1)) · ki · ·e . (28)
h

Claim 6.1. For any n large enough and all i ∈ [T ], we have ki ⩽ ℓi almost surely.

Proof of Claim 6.1. Note that since c ⩾ 1 by (24), it follows that i · e−i·c is decreasing in i ⩾ 1.
Thus by (27), almost surely, for all 1 ⩽ i ⩽ T

aph 1 aph 1 aph 1

· −i·c
⩽ · −T ·c
⩽ · .
ic2 ln(2h) a·e 2
T c ln(2h) a · e 2
⌈(1/c) · ln(2aph)⌉ · c a · e −⌈(1/c)·ln(2aph)⌉·c

27
Plugging in the bound from (26), aph ⩽ α ln n, and a ⩾ ln2 n, for large n we have

aph 1 aph c 2ec (aph)2 2ec (α ln n)2

· ⩽ · 2e ph = ⩽ ⩽ 1, (29)
ic2 ln(2h) a · e−i·c ln(2aph) ln(2aph) · a ln(2α ln n) · ln2 n

where the penultimate inequality uses the fact that x2 / ln(2x) is increasing on the interval x > 1.
Now, since x ⩽ y implies ⌈x⌉ ⩽ ⌈y⌉, it follows from (29) that ki ⩽ ℓi for large n and all i ∈ [T ]
almost surely. ♢

Observe that by the definition of the rounds we only have an upper bound on how many vertices
of A are not dominated at the start of each round, also it may happen that some round i is actually
skipped (i.e. τi = 1) if we find a vertex in a previous round that dominates many vertices of A.
Let Ei,j be the event that the j-th vertex uncovered during the i-th round has at least ki ⩽ ℓi
neighbours among the un-dominated vertices in A, and let Bi,j be the event that, at the time when
the algorithm processes the j-th vertex in the i-th round we still have that the un-dominated part
of A has cardinality at least ℓi . Then
   
ℓi ki ℓi −ki p · ℓi ki −2·p·ℓi
P (Ei,j | Bi,j ) ⩾ P (Bin(ℓi , p) = ki ) = p (1 − p) ⩾ e , (30)
ki ki

where the last inequality holds since e−2p ⩽ 1 − p holds for all p ∈ [0, .79], and p ⩽ ph ⩽ α/ ln n
by (22), and n is large by assumption. We now bound the right-hand side of (30) from below. Let
pi := P (Bin(ℓi , p) = ki ).
Claim 6.2. For any 1 ⩽ i ⩽ T ′ , we have pi ⩾ e−2c−1 · n−δ .

Proof of Claim 6.2. We have two cases for different values of h.

aph
The first case is ln(2h) ⩾ ic2
. Observe that in this case ki = 1 holds by (25). Thus by (28),
since x/ex is maximized by x = 1 and ln(2x)/x is maximized by x = e/2, for large n we have

ic2 ln(2h) c ln(e) c

p · ℓi ⩽ (1 + o(1)) · ki · i·c
· ⩽ (1 + o(1)) · 1 · · ⩽ .
e h e e/2 2

Observe also that p · ℓi ⩾ p · ℓT ′ ⩾ 1/(2ec h) by (26). Thus, by (30) and h ⩽ nδ , and recalling that
ki = 1, we have
1
pi ⩾ (p · ℓi ) · e−2·p·ℓi ⩾ · e−c ⩾ e−2c−1 · n−δ .
2ec h
aph aph
The second case is ln 2 ⩽ ln(2h) < ic2
. This condition implies that ic2 ln(2h)
> 1, and thus,
ki ⩾ 2. Therefore,

aph ic2 ln(2h) −i·c ki ic2 ln(2h) −i·c

p · ℓi ⩾ p · a · e−i·c = · ·e > · ·e .
ic2 ln(2h) h 2 h

28
 ic2 ln(2h)
However, for large n, we have p · ℓi ⩽ 2ki · hei·c
by (28). Using these bounds in (30) gives

 k  
1 ic2 ln(2h) −i·c i ic2 ln(2h)
pi ⩾ · ·e · exp −2 · 2ki ·
2 h hei·c
     2  
4c ln(2h) ic ln(2h)
= exp − ic 1 + − ln · ki . (31)
hei·c 2h

From the conditions i, h, c ⩾ 1 we have ln(2h) < h and ce−i·c < 1/2, which give the bounds
     
4c ln(2h) ic2 ln(2h) 2h 2h
<2 and − ln = ln ⩽ ln < ln (3h) .
hei·c 2h 2
ic ln(2h) ln 2

Inserting these bounds into (31) yields

   
aph 6aph 2aph ln(3h)
pi > exp −2 · (3ic + ln(3h)) · = exp − − .
ic2 ln(2h) c ln(2h) ic2 ln(2h)
Using i, h, c ⩾ 1, 1/ ln(2h) ⩽ 3/2, ln(3h)/ ln(2h) ⩽ ln(3)/ ln(2) < 2, and aph ⩽ α ln n gives
 
9aph 4aph
pi > exp − − > n−13α/c > n−δ ,
c c

where the last inequality holds by our choice of c from (24). ♢

PT
We now show that the “success” event { i=0 τi ⩽ b} holds with the desired probability. First,
observe that in round i ⩾ 1, there are at most ℓi−1 − 1 vertices of A which are not dominated at the
start of round i, when each vertex is added to Xi it covers at least ki undominated vertices in A,
and before the last vertex of round i is added to Xi there are at least ℓi undominated vertices in A.
Thus, we have
ℓi−1 − 1 − (|Xi | − 1) · ki ⩾ ℓi .

This, together with (28), implies that for 1 ⩽ i ⩽ T , almost surely, we have

ℓi−1 − 1 − ℓi ℓi−1 ℓi c2 ec ln(2h)

|Xi | ⩽ +1⩽ + 1 ⩽ (1 + o(1))ec · + 1 ⩽ (1 + o(1)) · · ie−i·c + 1.
ki ki ki ph
P∞ −i e
Thus, as c ⩾ 1 by (24), i=1 ie = (e−1)2
< 0.93, 1/ph = Ω(ln n) by (22), T = O(ln ln n) by (27),
and n is large, we have

X c2 ec ln(2h) X −i·c c2 ec ln(2h)

|Xi | ⩽ T + (1 + o(1)) · · ie ⩽ almost surely. (32)
ph ph
i∈[T ] i∈[T ]

Due to (30), there exist independent geometric random variables ηji ∼ Geo(pi ), j ∈ [|Xi |], such

29
 P|Xi | i
that, almost surely, for each i ∈ [T ], τi ⩽ j=1 ηj . Let
 
c2 ec ln(2h)
p∗ := e−2c−1 · n−δ and N := . (33)
ph
P
We know that, almost surely, mini∈[T ] pi ⩾ p∗ by Claim 6.2, and i∈[T ] |Xi | ⩽ N by (32). Hence
there exist independent ζk ∼ Geo(p∗ ) for k ∈ [N ] such that

|Xi | N
X XX X
τi ⩽ ηji ⩽ ζk almost surely.
i∈[T ] i∈[T ] j=1 k=1

PN
Thus, if Z := k=1 ζk , then it suffices for our goals to bound Z from above. To begin, observe that
by (33),  2 c 
1 2c+1 δ c e ln(2h)
µ := E (Z) = ·N =e n · . (34)
p∗ ph
By assumption we have bp > nβ , h ⩽ nδ , and δ < β are constants, so by (34) we have b/µ = ω(1).
Thus, for large n, by applying Lemma 2.6 with λ = b/µ, and p∗ = e−2c−1 · n−δ , we have
 
X
P τi > b ⩽ P (Z > b)
i∈[T ]

⩽ exp [−p∗ · µ · (λ − 1 − ln λ)]

⩽ exp [−(1 − o(1)) · p∗ · b]
h i
⩽ exp −e−2c−2 · n−δ · b .

P
The result follows by taking C := max{c2 ec , e2c+2 } in the statement, as conditional on i∈[T ] τi ⩽b
P
the desired set D := DT has size at most i∈[T ] |Xi | < C ln(eh)
ph almost surely by (32).

6.1 Simplified Version of Theorem 4.2

In this section, we show how to derive from Theorem 4.2 its simplified version stated in the intro-
duction, which we restate below for convenience.

Theorem 1.2. Let α, β > 0 be fixed constants, n be sufficiently large, and f (n) be any function
satisfying f (n) → ∞ as n → ∞. Let ln2 n ⩽ a ⩽ b ⩽ n and p ∈ (0, 1] be functions of n which satisfy
ap ⩽ α · ln n and bp > nβ . Then, for G ∼ G(A, B, p), where |A| = a and |B| = b, we have
  h i
f (n)
P ∃ a dominating set D ⊂ B of A with |D| ⩽ ∀a ∈ A, N (a) ̸= ∅ ⩾ 1 − exp −b1−o(1) .
p

Proof. Fix h = 1 in the Theorem 4.2 and thus h ⩽ nδ for any δ ⩾ 0. For α, β > 0 in the statement
of Theorem 1.2 and any δ > 0, let Dδ denote the event in the statement of Theorem 4.2, and

30
let C := {∀a ∈ A, N (a) ̸= ∅}. Observe that C and Dδ are both monotone increasing properties.
Thus, as a consequence of the FKG inequality [AS15, Theorem 6.3.3], we have P (Dδ | C) ⩾ P (Dδ ).
Conditional on the event C, the set A′ described in the event Dδ can be dominated using at most
1/p additional vertices from B, giving a dominating set D′ ⊆ B of A with size at most C ′ /p where
C ′ := 1 + C(α, δ) ln 2. Now, since f (n) → ∞, it follows that for any ε > 0, there exists constants
N, δN > 0 such that for n ⩾ N we have n−δN b/C(α, δN ) ⩾ b1−ε and C ′ = (1+C(α, δN ) ln 2) ⩽ f (n).
The result then follows by applying Theorem 4.2 over such a sequence of δN ’s.

We conclude this section with two propositions showing the optimality of Theorem 1.2. Namely,
w.h.p. a minimum dominating set of A in B has size Ω(1/p) when ap = Θ(ln n) (Proposition 6.3),
 
and the probability of existence of a dominating set of size O(1/p) is at most 1 − exp −b1+o(1)
when ap = Ω(1) (Proposition 6.4).

Proposition 6.3. Let a, b and p satisfy the conditions of Theorem 1.2, and additionally ap ⩾ c ln n
for some constant c > 0. Then, for large n,
 
c
P ¬∃ a dominating set D ⊂ B of A with |D| ⩽ ⩾ 1 − exp [−a/3] .
2p

Proof. We can assume w.l.o.g. that c ⩽ 2/e. Let X be the number of dominating sets D ⊆ B of A
c 1
which satisfy |D| ⩽ d := 2p ⩽ ep . Then,

  (i) (ii)
b
E (X) ⩽ d · · (1 − (1 − p)d )a ⩽ b · bd · (dp)a ⩽ n · ed ln n · e−a ⩽ n · e−a/2 ⩽ e−a/3 ,
d

where (i) holds as d ln n ⩽ c ln n

2p ⩽ a/2 and (ii) holds for large n since a ⩾ ln2 n by the hypothesis
from Theorem 1.2. The result then follows by Markov’s inequality.

Proposition 6.4. For any constant C ⩾ 1 and a, b and p satisfying the conditions of Theorem 1.2,
and additionally ap > C
 
C
P ∃ a dominating set D ⊂ B of A with |D| ⩽ ∀a ∈ A, N (a) ̸= ∅ ⩽ 1 − exp [−3abp] .
p

Proof. Choose any ln2 n ⩽ a ⩽ b ⩽ n and p ∈ [0, 1] which satisfy bp ⩾ nβ . Let

E := {∃ a dominating set D ⊂ B of A with |D| ⩽ C/p}, and C := {∀a ∈ A, N (a) ̸= ∅},

be the two events in the statement, and further let M := {E(G) is a matching of A in B}. Observe
that since a > C/p we have M ⊆ ¬E. Consequently, as P (C) > 0, we have

P (M)
P (E | C) = 1 − P (¬E | C) ⩽ 1 − P (M | C) = 1 − ⩽ 1 − P (M) .
P (C)

31
 b


Now since there are a · a! ways to choose a matching of A in B we have
 
b
P (M) = a! · pa (1 − p)(b−1)a
a
⩾ (b − a)a pa (1 − p)ab
2 /b
⩾ ba e−2a · pa e−2abp = exp[a ln(bp) − 2a2 /b − 2abp].

Recalling that nβ ⩽ bp ⩽ n and observing 2a2 /b ⩽ abp gives P (M) ⩾ exp[−3abp].

7 Conclusions
In this paper we determined the functionality of G(n, p) up to a constant factor for all p ∈ [0, 1]. In
p
particular, w.h.p. the functionality of G(n, p) is at most Θ( lnnn ) for any p ∈ [0, 1] (cf. Figure 1).
√
This is lower than the largest known functionality of an n-vertex graph, which is Ω( n) and is
achieved on the incidence graph of a finite projective plane [DFO+ 23]. On the other hand, the best
√
known general upper bound on the functionality of an n-vertex graph is O( n ln n) [DFO+ 23]. We
believe that this general upper bound can be improved, and suspect that it should be possible to
√
do this all the way to O( n).

Problem 7.1. Does there exist a constant C such that the functionality of every n-vertex graph is
√
at most C n?

In order to give some evidence supporting our intuition, we first recall that both upper and lower
bounds from [DFO+ 23] utilize the observation that in a graph G = (V, E), a vertex v is a function
of a set S if and only if S intersects every set (N (u1 )△N (u2 )) ∪ {u1 , u2 }, where u1 is a neighbour
and u2 is a non-neighbour of v, and u2 ̸= v. In other words, v is a function of S if and only if S is
a transversal in the hypergraph H(v) := (V, {(N (u1 )△N (u2 )) ∪ {u1 , u2 } : u1 ∈ N (v), u2 ̸∈ N [v]}).
√
Using this observation, the upper bound of O( n ln n) is obtained by showing that if, for every
√
two vertices u, v, the symmetric difference N (u)△N (v) has size at least Ω( n ln n), then a uniformly
√
random set of size Θ( n ln n) intersects each N (u)△N (v) with a positive probability. The lower
√
bound of Ω( n) is proved by showing that in the incidence graph P of a finite projective plane with
p
n points and n lines, for every vertex v, no set of size less than n/2 is a transversal in H(v).
A possible way to improve the lower-bound construction is to add or delete some random edges
in P . This is supported by the following two observations. First, while the functionality of Kn is
0, after removing every edge from Kn with probability 1/2 the functionality of the resulting graph
becomes fun(G(n, p)) = Θ(log n) w.h.p. Second, it is known that the minimum transversal of a
hypergraph with suboptimal transversal number typically increases after deleting half of vertices
from every hyperedge uniformly at random (see, [HY18, Theorem 4]). However, this strategy does
√
not seem to work for P . Indeed, the maximum degree (and thus functionality) of P is O( n) and
the removal of edges can only lower the maximum degree. On the other hand, we might add at

32
 √
least Θ(n n ln n) random edges to P instead, but it makes the influence of deterministic edges
insignificant. In particular, it can be shown that any distinguishing set S in the original graph of
√
size Θ( n), w.h.p. is also a distinguishing set in the randomly perturbed graph.
Finally, we note that there exist k-uniform hypergraphs on n vertices with the transversal number
Ω(n log k/k) (see [Alo90, TY07, HY18]), which would be promising constructions for attempting
√ √
to lower bound functionality with Ω( n ln n) (by taking k = Θ( n ln n)). Unfortunately, it seems
unlikely that for such a hypergraph there exists a corresponding graph with the symmetric differences
coinciding with the hypergraph’s hyperedges.

References
[AAL21] Bogdan Alecu, Aistis Atminas, and Vadim V. Lozin. Graph functionality. J. Comb.
Theory, Ser. B, 147:139–158, 2021.

[ACH+ 24] Jungho Ahn, Debsoumya Chakraborti, Kevin Hendrey, Donggyu Kim, and Sang-il Oum.
Twin-width of random graphs. Random Structures & Algorithms, 2024.

[ACLZ15] Aistis Atminas, Andrew Collins, Vadim V. Lozin, and Victor Zamaraev. Implicit rep-
resentations and factorial properties of graphs. Discret. Math., 338(2):164–179, 2015.

[AHKO22] Jungho Ahn, Kevin Hendrey, Donggyu Kim, and Sang-il Oum. Bounds for the twin-
width of graphs. SIAM Journal on Discrete Mathematics, 36(3):2352–2366, 2022.

[Alo90] Noga Alon. Transversal numbers of uniform hypergraphs. Graphs and Combinatorics,
6:1–4, 1990.

[AS15] Noga Alon and Joel H. Spencer. The probabilistic method. Hoboken, NJ: John Wiley &
Sons, fourth edition, 2015.

[BMP13] Béla Bollobás, Dieter Mitsche, and Paweł Prałat. Metric dimension for random graphs.
Electronic Journal of Combinatorics, 20(4), 2013.

[Bol01] Béla Bollobás. Random graphs, volume 73 of Cambridge Studies in Advanced Mathe-
matics. Cambridge University Press, Cambridge, second edition, 2001.

[BWZ24] Tom Bohman, Lutz Warnke, and Emily Zhu. Two-point concentration of the domination
number of random graphs. arXiv, 2401.10486, 2024.

[DFO+ 23] Pavel Dvorák, Lukáš Folwarczný, Michal Opler, Pavel Pudlák, Robert Sámal, and
Tung Anh Vu. Bounds on functionality and symmetric difference - two intriguing graph
parameters. In Graph-Theoretic Concepts in Computer Science - 49th International
Workshop, WG 2023, volume 14093 of Lecture Notes in Computer Science, pages 305–
318. Springer, 2023.

33
[DLM+ 24] Clément Dallard, Vadim V. Lozin, Martin Milanic, Kenny Storgel, and Viktor Zama-
raev. Functionality of box intersection graphs. Results Math, 79:48, 2024.

[FMM+ 07] Alan Frieze, Ryan Martin, Julien Moncel, Miklós Ruszinkó, and Cliff Smyth. Codes
identifying sets of vertices in random networks. Discrete Mathematics, 307(9-10):1094–
1107, 2007.

[GLS15] Roman Glebov, Anita Liebenau, and Tibor Szabó. On the concentration of the domi-
nation number of the random graph. SIAM J. Discret. Math., 29(3):1186–1206, 2015.

[HNST24] Kevin Hendrey, Sergey Norin, Raphael Steiner, and Jérémie Turcotte. Twin-width of
sparse random graphs. Combinatorics, Probability and Computing, page 1–20, 2024.

[HY18] Michael A. Henning and Anders Yeo. Transversals in uniform linear hypergraphs. arXiv,
1802.01825, 2018.

[Jan18] Svante Janson. Tail bounds for sums of geometric and exponential variables. Statist.
Probab. Lett., 135:1–6, 2018.

[JLR00] Svante Janson, Tomasz Luczak, and Andrzej Rucinski. Random graphs. Wiley-
Interscience series in discrete mathematics and optimization. Wiley, 2000.

[LLO12] Choongbum Lee, Joonkyung Lee, and Sang-il Oum. Rank-width of random graphs. J.
Graph Theory, 70(3):339–347, 2012.

[NS93] Sotiris E. Nikoletseas and Paul G. Spirakis. Near-optimal dominating sets in dense
random graphs in polynomial expected time. In Graph-Theoretic Concepts in Com-
puter Science, 19th International Workshop, WG 1993, volume 790 of Lecture Notes in
Computer Science, pages 1–10. Springer, 1993.

[TY07] Stéphan Thomassé and Anders Yeo. Total domination of graphs and small transversals
of hypergraphs. Combinatorica, 27(4):473–487, 2007.

[Web81] Karl Weber. Domination number for almost every graph. Rostock. Math. Kolloq.,
(16):31–43, 1981.

[WG01] Ben Wieland and Anant P. Godbole. On the domination number of a random graph.
Electron. J. Comb., 8(1), 2001.

34