A Bound for the Diameter of Random Hyperbolic Graphs
Marcos Kiwi∗ Dieter Mitsche†
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
Abstract
Random hyperbolic graphs were recently introduced by
Krioukov et. al. [KPK+ 10] as a model for large networks.
Gugelmann, Panagiotou, and Peter [GPP12] then initi-
ated the rigorous study of random hyperbolic graphs us-
ing the following model: for α > 12 , C ∈ R, n ∈ N, set
R = 2 ln n + C and build the graph G = (V, E) with
|V | = n as follows: For each v ∈ V , generate i.i.d. po-
lar coordinates (rv , θv ) using the joint density function
f (r, θ), with θv chosen uniformly from [0, 2π) and rv with
α sinh(αr)
density f (r) = cosh(αR)−1 for 0 ≤ r < R. Then, join
two vertices by an edge, if their hyperbolic distance is at
most R. We prove that in the range 12 < α < 1 a.a.s. for
any two vertices of the same component, their graph dis-
2
tance is O(logC0 +1+o(1) ), where C0 = 2/( 12 − 34 α + α4 ),
thus answering a question raised in [GPP12] concerning
the diameter of such random graphs. As a corollary from
our proof we obtain that the second largest component
has size O(log2C0 +1+o(1) n), thus answering a question of
Bode, Fountoulakis and Müller [BFM13]. We also show
that a.a.s. there exist isolated components forming a path
of length Ω(log n), thus yielding a lower bound on the size
of the second largest component.
1 Introduction
Building mathematical models to capture essential
properties of large networks has become an important
objective in order to better understand them. An in-
teresting new proposal in this direction is the model
of random hyperbolic graphs recently introduced by
Krioukov et. al. [KPK+ 10] (see also [PKBnV10]). A
good model should on the one hand replicate the
characteristic properties that are observed in real
∗ Depto. Ing. Matemática & Ctr. Modelamiento Matemático
(CNRS UMI 2807), U. Chile. Beauchef 851, Santiago, Chile,
Email: mkiwi@dim.uchile.cl. Gratefully acknowledges the
support of Millennium Nucleus Information and Coordination
in Networks ICM/FIC P10-024F and CONICYT via Basal in
Applied Mathematics.
† Université de Nice Sophia-Antipolis, Laboratoire J-A
Dieudonné, Parc Valrose, 06108 Nice cedex 02, Email:
dmitsche@unice.fr. The main part of this work was per-
formed during a stay at Depto. Ing. Matemátics & Ctr. Mod-
elamiento Matemático (CNRS UMI 2807), U. Chile, and the
author would like to thank them for their hospitality.
26
world networks (e.g., power law degree distributions,
high clustering and small diameter), but on the other
hand it should also be susceptible to mathemati-
cal analysis. There are models that partly succeed
in the first task but are hard to analyze rigorously.
Other models, like the classical Erdös-Renyi G(n, p)
model, can be studied mathematically, but fail to
capture certain aspects observed in real-world net-
works. In contrast, the authors of [PKBnV10] ar-
gued empirically and via some non-rigorous meth-
ods that random hyperbolic graphs have many of the
desired properties. Actually, Boguñá, Papadopoulos
and Krioukov [BnPK10] computed explicitly a maxi-
mum likelihood fit of the Internet graph, convincingly
illustrating that this model is adequate for reproduc-
ing the structure of real networks with high accuracy.
Gugelmann, Panagiotou, and Peter [GPP12] initi-
ated the rigorous study of random hyperbolic graphs.
They compute exact asymptotic expressions for the
maximum degree, the degree distribution (confirming
rigorously that the degree sequence follows a power-
law distribution with controllable exponent), and also
estimated the expectation of the clustering coeffi-
cient.
In words, the random hyperbolic graph model
is a simple variant of the uniform distribution of n
vertices within a disc of radius R of the hyperbolic
plane, where two vertices are connected if their
hyperbolic distance is at most R. Formally, the
random hyperbolic graph model Gα,C (n) is defined
in [GPP12] as described next: for α > 12 , C ∈ R,
n ∈ N, set R = 2 ln n + C, and build G = (V, E) with
vertex set V = [n] as follows:
• For each v ∈ V , polar coordinates (rv , θv )
are generated identically and independently dis-
tributed with joint density function f (r, θ), with
θv chosen uniformly at random in the interval
[0, 2π) and rv with density:
 α sinh(αr) , if 0 ≤ r < R,

f (r) = C(α, R)
0, otherwise,

where C(α, R) = cosh(αR)−1 is a normalization
constant.
Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 • For u, v ∈ V , u 6= v, there is an edge with end-
points u and v provided d(ru , rv , θu − θv ) ≤ R,
where d = d(r, r0 , θ − θ0 ) denotes the hyper-
bolic distance between two vertices whose native
representation polar coordinates are (r, θ) and
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
ball of sufficiently small radius around the origin and
partition it into 3 sectors. It can be shown that in
each of the sectors a.a.s. there will be at least one
vertex, and also that every other vertex is connected
to at least one of the three vertices. For α = 12 the

(r0 , θ0 ), obtained by solving probability of this configuration to exist depends on

C (see [BFM]). For α > 1, there exists no giant com-
(1.1) ponent (see [BFM13]); the case α = 1 is a matter
cosh(d) requiring further study.
= cosh(r) cosh(r0 ) − sinh(r) sinh(r0 ) cos(θ−θ0 ). We show (see Theorem 3.1) that for 12 < α < 1,
a.a.s., for any two vertices of the same component,
The restriction α > 1/2 and the role of R, informally their graph distance is O(logC0 +1+o(1) ), where C0 =
speaking, guarantee that the resulting graph has a 2/( 1 − 3 α + α2 ). Our proof argument also yields
2 4 4
bounded average degree (depending on α and C (see Corollary 3.2) that the size of the second largest
only). If α < 1/2, then the degree sequence is so component is O(log2C0 +1+o(1) ). As a complementary
heavy tailed that this is impossible. result (see Theorem 4.1), we establish that a.a.s. there
Research in random hyperbolic graphs is in a exists a component forming an induced path of length
sense in its infancy. Besides the results mentioned Θ(log n).
above, very little else is known. Notable exceptions
are the emergence and evolution of giant compo- Organization. This work is organized as fol-
nents [BFM13], connectedness [BFM], results on the lows. In Section 2 we introduce the geometric frame-
global clustering coefficient of the so called binomial work and give background results. Section 3 deals
model of random hyperbolic graphs [CF13], and on with the upper bound on the diameter as well as the
the evolution of graphs on more general spaces with upper bound on the size of the second largest com-
negative curvature [Fou12]. ponent. Section 4 is dedicated to the lower bound.

Notation. As typical in random graph the- 2 Conventions, background results and

ory, we shall consider only asymptotic properties as preliminaries
n → ∞. We say that an event in a probability space
Henceforth, for a point P in hyperbolic plane, we let
holds asymptotically almost surely (a.a.s.) if its prob-
(rP , θP ) denote its polar coordinates (0 ≤ rP < R
ability tends to one as n goes to infinity. We follow
and 0 ≤ θP < 2π). The point with polar coordinates
the standard conventions concerning asymptotic no-
(0, 0) is called the origin and is denoted by O.
tation. In particular, we write f (n) = o(g(n)), if
By (1.1), the hyperbolic triangle formed by the
limn→∞ |f (n)|/|g(n)| = 0. We will nevertheless also
geodesics between points A, B, and C, with opposing
use 1 − o(·) when dealing with probabilities. We also
side segments of length a, b, and c respectively, is such
say that an event holds with extremely high proba-
that the angle formed at C is:
bility, w.e.p., if it occurs with probability at least  cosh(a) cosh(b) − cosh(c) 
1 − e−ω(log n) . Throughout this paper, log n always θc (a, b) = arccos .
denotes the natural logarithm of n. sinh(a) sinh(b)
The constants α, C used in the model and the Clearly, θc (a, b) = θc (b, a).
constants C0 , δ defined below have only one special In fact, although some of the proofs hold for a
meaning, other constants such as C 0 , C 00 , c, c1 , c2 , c3
wider range of α, we will always assume 12 < α < 1,
change from line to line. Since we are interested since as already discussed, we are interested in this
in asymptotic results only, we ignore rounding issues regime. In order to avoid unnecessary repetitions, we
throughout the paper. henceforth omit this restriction from the statement
Results. The main problem we address in of this article’s results.
this work is the natural question, explicitly stated First, we recall some useful estimates. A very
in [GPP12, page 6], that asks to determine the ex- handy approximation for θc (·, ·) is given by the fol-
pected diameter of the giant component of a random lowing result.
hyperbolic graph G chosen according to Gα,C (n) for Lemma 2.1. ([GPP12, Lemma 3.1]) If 0 ≤
1 1
2 < α < 1. We look at this range, since for α < 2 min{a, b} ≤ c ≤ a + b, then
a.a.s. a very small central configuration yielding a di- 1
θc (a, b) = 2e 2 (c−a−b) 1 + Θ(ec−a−b ) .


ameter of at most 3 exists (see [BFM]): consider a

27 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 Remark 2.1. We will use the lemma also in this
form: let A and B be two points such that rA , rB >
R/2, 0 ≤ min{rA , rB } ≤ d := d(A, B) ≤ R. Then,
taking c = d, a = rA , b = rB in Lemma 2.1 and
r
noting that rA + rB ≥ R ≥ d, we get
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

ρO
|θA − θB | = θd (rA , rB )
1 θρA (r, rA )
= 2e 2 (d−rA −rB ) 1 + Θ(ed−rA −rB ) .

Note also that for fixed rA , rB > R/2, |θA − θB | A

R
is increasing as a function of d (for d satisfying
the constraints). Below, when aiming for an upper
bound, we always use d = R.
ρA A r
Throughout, we will need estimates for measures of
regions of the hyperbolic plane, and more specifically,
for regions obtained by performing some set algebra
involving a couple of balls. Hereafter, for a point P
of the hyperbolic plane, ρP will be used to denote the Figure 1: Region whose measure we want to estimate.
radius of a ball centered at P , and BP (ρP ) to denote
the closed ball of radius ρP centered at P . Also, we recalling that C(α, R) = cosh(αR) − 1 = 1 eαR +
2
denote by µ(S) the probability measure of a set S, O(e−αR ), for this part we obtain
i.e., Z 1 1 1
µ(S) = f (r, θ)drdθ. αe 2 (ρA −rA )  e(α− 2 )ρO − e(α− 2 )(ρA −rA ) 
+ O(1)
S πC(α, R) α − 12
We collect now a few results for such measures.  1
= Cα e−α(R−ρO )− 2 (ρO −ρA +rA ) − e−α(R−ρA +rA )
Lemma 2.2. ([GPP12, Lemma 3.2]) For any 0 ≤ 1

−α(R−ρO ) + Θ(e 2 (ρA −rA )−αR )
ρO ≤ R, µ(BO (ρO )) = e (1 + o(1)).
 1
Lemma 2.3. Let Cα = 2α/(π(α − 12 )). For rA ≤ ρA = Cα e−α(R−ρO )− 2 (ρO −ρA +rA )
and ρO + rA ≥ ρA , 
+ O(e−α(R−ρA +rA ) ) ,
µ(BA (ρA ) ∩ BO (ρO ))
1
−α(R−ρO )− 12 (ρO −ρA +rA ) −α(R−ρA +rA ) where the last identity is because e 2 (ρA −rA )−αR =
= Cα e + O(e ).
O(e−α(R−ρA +rA ) ) for α > 12 . Now, for the error term,
Proof. We want to bound (see Figure 1): since α < 32 , we obtain
3 3 3
µ(BA (ρA ) ∩ BO (ρO )) αe 2 (ρA −rA )  Θ(e(α− 2 )ρO − e(α− 2 )(ρA −rA ) ) 
+ O(1)
Z ρO Z θρA (r,rA )
f (r) πC(α, R) α − 32
= µ(BO (ρA − rA )) + 2 dθdr. 3
ρA −rA 0 2π αe 2 (ρA −rA )
= O(1) = O(e−α(R−ρA +rA ) ).
πC(α, R)
Relying on the approximation of θρA (r, rA ) given by
Lemma 2.1, Applying Lemma 2.2 to µ(BO (ρA − rA )), the desired
conclusion follows. 
µ(BA (ρA ) ∩ BO (ρO ))
1 From now on, denote R0 := R/2 and Ri :=
αe 2 (ρA −rA ) i
= µ(BO (ρA − rA )) + Re−α /2 for i ≥ 1. Observe that Ri−1 ≤ Ri for all i.
πC(α, R)
Z ρO
1 1
Lemma 2.4. Let ξ > 0 be some constant and i ≥ 1
× (e(α− 2 )r − e−(α+ 2 )r )(1 + Θ(eρA −r−rA ))dr.
be such that Ri < R − ξ. If Ri < rA ≤ Ri+1 , then
ρA −rA


µ(BA (R) ∩ BO (Ri ) \ BO (Ri−1 ) )
We will first solve the integral without the error
term Θ(eρA −r−rA ). Since Cα = 2α/(π(α − 12 )), and = Ω(µ(BA (R) ∩ BO (Ri ))),

28 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 where the constant C 0 = C 0 (ξ, α) hidden inside the
asymptotic notation can be assumed nondecreasing as
a function of ξ.
Proof. Since Ri−1 ≤ Ri , we have that BO (Ri−1 ) ⊆
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
C ∈ R. The intensity function at polar coordinates
(r, θ) for 0 ≤ r < R and 0 ≤ θ < 2π is equal to
α sinh(αr)
g(r, θ) := δeR/2 f (r, θ) = δeR/2
2πC(α, R)

BO (Ri ), so the left hand side of the stated identity R R R 2π

−C/2
can be re-written as µ(BA (R)∩BO (Ri ))−µ(BA (R)∩ with δ = e . Note that r=0 θ=0 g(r, θ)dθdr =
BO (Ri−1 )). Now, observe that by Lemma 2.3 applied δeR/2 = n, and thus E|P| = n.
with ρA = R, ρO = Ri (so Ri + rA > 2Ri ≥ R), The main advantage of defining P as a Poisson
point process is motivated by the following two
(2.2) µ(BA (R) ∩ BO (Ri )) properties: the number of points of P that lie in
= (1 + o(1))Cα e −α(R−Ri )− 12 (Ri −R+rA )
, any region S ∩ BO (R)R follows a Poisson distribution
with mean given by S g(r, θ)drdθ = nµ(S ∩ BO (R));
where Cα = 2α/(π(α − 12 ). By the same argument, and the number of points of P in disjoint regions of
this time applied with ρA = R, ρO = Ri−1 , we also the hyperbolic plane are independently distributed.
have Moreover, by conditioning P upon the event |P| = n,
we recover the original distribution. Note that for any
(2.3) µ(BA (R) ∩ BO (Ri−1 )) k, by standard estimates for the Poisson distribution,
−α(R−Ri−1 )− 12 (Ri−1 −R+rA )
we have
= (1 + o(1))Cα e .
√
(2.4) P(|P| = k) ≤ P(|P| = n) = O(1/ n).
It suffices to show that the ratio ρ(R, i) between the
right hand sides of the expressions in (2.2) and (2.3) In some of the arguments below, we will add a set
is at least a constant. We claim that the assertion of vertices Q of size |Q| = m, that are all chosen in-
regarding ρ(R, i) holds. Indeed, note that dependently according to the same probability distri-
1
bution as every vertex in Gα,C (n), we add them to P
ρ(R, i) = (1 + o(1))e(α− 2 )(Ri −Ri−1 ) and consider P ∪Q. Throughout this work all vertices
1 −αi /2
−e −αi−1 /2 named by Q, or in case there are more, Q1 , . . . , Qm
= (1 + o(1))e(α− 2 )R(e )
, are vertices added in this way. For m = o(n1/2 ), by
Stirling’s approximation, we have
and let i1 = O(1) be large enough so that 1 − α ≥
i1 i1 −1
αi1 −1 /2. Since 12 < α < 1, β := e−α /2 −e−α /2
> −n n
n−m

0 is a constant such that if i ≤ i1 , then ρ(R, i) ≥ (2.5) P(|P| = n − m) = e

1
(n − m)!
ρ(R, i1 ) = (1 + o(1))e(α− 2 )βR . Thus, the claim holds −m
 n n−m
for i ≤ i1 . =e Θ(n−1/2 )
n − m
Using now 1 − x ≤ e−x ≤ 1 − x + x2 /2, by our  m n−m
choice of i1 , and recalling that α < 1, = e−m 1 + Θ(n−1/2 ).
n−m
 i i−1
 αi−1  αi−1  Using that ex/(1+x) ≤ 1 + x if x > −1 and recalling
R e−α /2 − e−α /2 ≥ R 1−α−
2 4 that m = o(n1/2 ), we get
αi−1 1 − α
≥R · . m
P(|P| = n − m) ≥ e−m e(n−m) n Θ(n−1/2 )
2 2
√
= e−o(1) Θ(n−1/2 ) = Ω(1/ n).
Since by assumption Ri < R − ξ, we also have αi ≥
2ξ (α− 12 )(1−α)ξ/(2α)
R , and thus ρ(R, i) ≥ (1 + o(1))e , Thus, for such an m, by (2.5),
finishing the proof. 
√
(2.6) P(|P| = n − m) = Θ(1/ n).
In order to simplify our proofs, we will make
use of a technique known as de-Poissonization, By conditioning upon |P| = n − m, we recover the
which has many applications in geometric probabil- original distribution of n points. Moreover, for any
ity (see [Pen03] for a detailed account of the subject). m = o(n1/2 ), any event holding in P or P ∪ Q with
Throughout the paper we work with a Poisson point probability at least 1−o(fn ) must hold in the √ original
process on the hyperbolic disc of radius R and denote setup with probability at least 1 − o(fn n), and
its point set by P. Recall that R = 2 log n + C for in particular, any event holding with probability at

29 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 √
least 1 − o(1/ n) holds asymptotically almost surely for i = 1, . . . , k. Then, with probability at least
(a.a.s.), that is, with probability tending to 1 as 1 − o(n−3/2 ), for any 0 ≤ i ≤ j ≤ k,
n → ∞, in the original model. We identify below
points of P with vertices. We prove all results below |θvi − θvj | = O(log n/n).
for the Poisson model and then transform the results
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

to the uniform model. Proof. By Lemma 3.1, with probability at least 1 −

o(n−3/2 ), k = O(log n). By the proof of Lemma 3.1,
3 The upper bound on the diameter for any two vi , vj with rvi , rvj > R − ξ and d(vi , vj ) ≤
R, we must have |θvi − θvj | = O(1/n). The corollary
We now turn our attention towards the upper bound.
follows by the triangle inequality for angles. 
We first show that there cannot be long paths with all
vertices being close to the boundary (i.e., at distance 2
Recall that C0 = 2/( 12 − 34 α + α4 ) and note that
roughly R from the origin O).
C0 > 8. Define throughout the whole section i0 :=
i0
Lemma 3.1. Let ξ > 0. Let Q ∈ Q be a vertex logα ((2C0 log R)/R). Note that Ri0 = Re−α /2 =
αi0
added, and suppose rQ > R − ξ. There is a constant R(1 − (1 + o(1)) 2 ) = R − (1 + o(1))C0 log R. Also,
C 0 = C 0 (ξ, C) such that with probability at least observe that for any two adjacent vertices vi , vi+1
1 − o(n−3/2 ), there is no path Q =: v0 , v1 , . . . , vk , with rvi , rvi+1 > Ri0 , by Remark 2.1,
with k ≥ C 0 log n and rvi > R − ξ for i = 1, . . . , k. 1
(3.7) |θvi − θvi+1 | ≤ (2 + o(1))e 2 (R−rvi −rvi+1 )
Proof. First, note that by Remark 2.1, for two ver- 1

tices vi , vj with rvi , rvj > R − ξ and d(vi , vj ) ≤ R, we ≤ (2 + o(1))e 2 (R−2Ri0 )

1
must have |θvi − θvj | ≤ (1 + o(1))2e−R/2+ξ = C 00 /n = e− 2 R RC0 (1+o(1)) = O(logC0 +o(1) n/n).
for C 00 = (2 + o(1))e−C/2+ξ . Partition the disc of ra-
dius R into Θ(n) equal sized sectors of angle 2πC 00 /n For a vertex v ∈ P with rv > Ri0 , let ` be such
and order them counterclockwise. Note that any path that R` < rv ≤ R`+1 . Define a center path from
containing only vertices v with rv > R − ξ satisfies v to be a sequence of vertices v =: w0 , w1 , . . . , ws
the following property: if it contains a vertex vi in such that d(wi , wi+1 ) ≤ R, R`−i < rwi ≤ R`−i+1 for
a sector Si and a vertex vj in a sector Sj , either 0 ≤ i ≤ s − 1, and Ri0 −1 < rws ≤ Ri0 .
all sectors between Si and Sj in the counterclock- Lemma 3.2. Let ξ 0 > 0 and Q ∈ Q be a vertex
wise ordering or all sectors between Si and Sj in the added. Suppose that R < r ≤ R
` Q `+1 for ` ≥ i0 and
clockwise ordering have to contain at least one ver- R ≤ R − ξ 0 . Then, with probability bounded away
`
tex from the path. By symmetry, for each sector, from 1, there exists no center path Q =: ω , . . . , ω
0 s
the expected number of vertices inside the sector is from Q.
C 00 , and therefore this is also an upper bound for
the expected number of vertices v inside this sector Proof. Denote by E0 the event that there exists one
with the additional restriction of rv > R − ξ. Thus, vertex of P that belongs to Bw (R) ∩ (BO (R` ) \
0
the probability of having at least one vertex inside a BO (R`−1 )). By Lemma 2.4,
00
sector is at most 1 − e−C < 1. The probability to
have at least one vertex in C 000 log n given consecu- µ(Bw0 (R) ∩ (BO (R` ) \ BO (R`−1 )))
tive sectors starting from the sector that contains Q = Ω(µ(Bw0 (R) ∩ BO (R` ))).
00 000
is thus at most 2(1 − e−C )C log n = o(n−3/2 ) for
C 000 = C 000 (C 00 ) sufficiently large. Thus, with prob- By Lemma 2.3 (applied with ρA = R, ρO = R` , and
ability at least 1 − o(n−3/2 ), any such path starting R` < rA = rw ≤ R`+1 ), we have
0
at Q contains only vertices from at most C 000 log n
sectors. By Chernoff bounds together with a union µ(Bw0 (R) ∩ BO (R` ))
bound, for C 000 large enough, with probability at least 2α  −(α− 1 )(R−R` )− 1 rw 
−3/2 000 = e 0 + O(e−αrw0 )
1 − o(n ), any consecutive set of C log n sectors 2 2
π(α − 12 )
contains O(log n) vertices, and the statement follows. 1 1
 = Θ(e−(α− 2 )(R−R` )− 2 rw0 ),

Corollary 3.1. Let ξ > 0. Let Q ∈ Q be a where the last identity follows because rw0 > R −
vertex added, and suppose rQ > R − ξ, and let R` . Now, since 1 − e−x ≤ x, we have R − R` =
`
Q =: v0 , v1 , . . . , vk be a path with rvi > R − ξ R(1 − e−α /2 ) ≤ Rα` /2, and since for x = o(1),
5

30 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 e−x = (1 + o(1))(1 − x), we also have rw0 ≤ R`+1 = Clearly, since α−1 > 1, the sum converges. Note
that the constant C 0 coming from Lemma 2.4 is
`+1
Re−α /2 = R(1 − (1 + o(1))α`+1 /2). Thus
nondecreasing as a function of ξ 0 . Hence, by choosing
−(α− 12 )(R−R` )− 12 rw0
c
P(E0 ) = exp(−Ω(ne )) ξ 0 = ξ 0 (C 0 , α) big enough, the sum is less than 1, and
1 R ` R `+1 the statement follows. 
= exp(−Ω(ne−(α− 2 ) 2 α −(1−(1+o(1)) 4 α ) )).
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

The following lemmas will show that for vertices v

Recalling that e−R/2 = Θ(1/n), we have with rv ≤ Ri0 the probability of not connecting to a
(3.8) vertex in BO (R/2) can be bounded by a much smaller
1 R ` R `+1
P(E0 c ) = exp(−Ω(e−(α− 2 ) 2 α +(1+o(1)) 4 α )). expression than the analogous probability bound for
vertices close to the boundary R given by Lemma 3.2.
Note that if E0 holds, then a vertex inside the desired Define j0 = j0 (α) ≥ 1 to be a constant sufficiently
region is found. Assuming the existence of a vertex wi large so that e−αj /2 ≤ 1 − (1 − 1−α ) αj for j ≥ j
2 2 0
with R`−i < rwi ≤ R`−i+1 , we continue inductively (note that such j exists because e−x ≤ 1 − x + x2 if
0
for i = 1, . . . , s − 1 = ` − i0 in the same way: we |x| ≤ 1).
define the event Ei that there exists one element of P
that belongs to Bwi (R) ∩ (BO (R`−i ) \ BO (R`−i−1 )). Lemma 3.3. Let Q ∈ Q be a vertex added such that
By the same calculations (noting that Lemma 2.3 can Rj < rQ ≤ Rj+1 with j0 ≤ j ≤ i0 . W.e.p.,
still be applied and R`−i + rwi > R still holds), we

BQ (R) ∩ (BO (Rj ) \ BO (Rj−1 ) ∩ P = 6 ∅.
obtain
1 R `−i R `−i+1
Proof. As in the previous proof, by Lemma 2.4,
P(Ei c ) = exp(−Ω(e−(α− 2 ) 2 α +(1+o(1)) 4 α )).
(3.9) µ(BQ (R) ∩ (BO (Rj ) \ BO (Rj−1 )))
Denote by C the event of not having a center path, = Ω(µ(BQ (R) ∩ BO (Rj ))),
we have thus by independence of the events
and also as before, by Lemma 2.3 (applied with
s−1
X s−1
X ρA = R, ρO = Rj , and Rj < rA = rQ ≤ Rj+1 ),
P(C) ≤ P(E0 c ) + P(Ei c |E0 , . . . , Ei−1 ) = P(Ei c ). we have
i=1 i=0 1 1
µ(BQ (R) ∩ BO (Rj )) = Θ(e−(α− 2 )(R−Rj )− 2 rQ ).
Hence, j
Again, R − Rj = R(1 − e−α /2 ) ≤ Rαj /2, and since
`−i
X0 j
αj
1 R `−i
+(1+o(1)) R `−i+1 by assumption e−α /2 ≤ 1 − (1 − 1−α
2 ) 2 , we have
P(C) ≤ exp(−Ω(e−(α− 2 ) 2 α 4 α )) j+1
αj+1
i=0 rQ ≤ Rj+1 = Re−α /2 ≤ R(1 − (1 − 1−α 2 ) 2 ).
`−i
X0 Thus,
1 R `−i `−i+1
+(1+o(1)) R
≤ e−Ω(1−(α− 2 ) 2 α 4 α )
µ(Bz (R) ∩ BO (Rj ))
i=0
1 R j
−(1−(1− 1−α R j+1
`−i
X0 = Ω(e−(α− 2 ) 2 α 2 ) 4 α )
)
`−i
−Ω(1+(1+o(1)) R
4 α (1−α))
≤ e −R/2 R j 1 3 α2
= Ω(e e 2 α ( 2 − 4 α+ 4 )
).
i=0
`−i 2
X0 0
−(1+o(1))C 0 R `−i Note that 12 − 34 α + α4 > 0 for 12 < α < 1, so
≤ e−C 4 α (1−α)
,
the last displayed expression is clearly decreasing
i=0
in j. By plugging in our upper bound j = i0 =
where C 0 = C 0 (ξ 0 ) > 0 is the constant hidden in logα (2C0 log R/R), we get with our choice of C0 =
2
the asymptotic notation of Lemma 2.4. Clearly, the 2/( 12 − 34 α + α4 ),
closer Q is to the boundary, the more difficult it is  R2 
to find a center path, and we may thus assume that µ(BQ (R) ∩ BO (Rj )) = Ω R/2 = Ω((log n)2 /n).
R` < rQ ≤ R`+1 is such that R`+1 > R − ξ 0 . Then, e
`+1
noting that R`+1 = Re−α
/2 > R − ξ 0 implies that Hence, the expected number of vertices in BQ (R) ∩
` = logα (1 + o(1))2ξ 0 /R must hold and plugging BO (Rj ) is Ω(log2 n), and by Chernoff bounds
this into the previous, sum we get for Poisson random variables (see [AS08, Theo-
0
rem A.1.15]), w.e.p. there are at least Ω(log2 n) ver-
0 X 0 ξ (1−α)
P(C) ≤ e−C e−(1+o(1))C 2αi . tices in this region. By (3.9) the statement follows.
i≥0 

31 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 Finally, for vertices v ∈ P with R/2 < rv ≤ Rj0 Clearly, the bigger j, the bigger the last displayed
we will for the next lemma modify the definition of term is. Thus, we plug in the smallest possible value
Ri : since j0 = O(1), there exists some 12 < c < 1 j = 1, and together with (3.10) we obtain
j0
such that Rj0 = Re−α /2 =: cR. Let T ≥ 1 be the
largest integer such that c − T2 (1 − c)(1 − α) > 12 . µ(BQ (R) ∩ (BO (Rj0 ) \ BO (Rj+1
0
)))
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

0


Define now the new bands to be R0 := cR, and for
 R
− 2 −(1−c)(1−α) +(1−c)(2α−1+ 1−α
2
2 )+c
= Ω e
any i = 1, . . . , T , define Ri0 := R(c − 2i (1 − c)(1 − α)).
Note in particular, that Ri0 ≥ R/2 for all i in the
 R

− 2 (1−c)( 7α−3 −α2 )+c
=Ω e 2 .
range of interest. We have the following result.

Lemma 3.4. Let Q ∈ Q be a vertex added. Note that for 12 < α < 1, we have 7α−3 2 − α2 < 1,
and thus the constant factor multiplying − R2 in the
0 0
• If Rj < rQ ≤ Rj−1 for some 1 ≤ j ≤ T − 1, then exponent of the last term of the previously displayed
w.e.p., we have equation is clearly bounded by c < 1. Hence,
µ(BQ (R) ∩ (BO (Rj0 ) \ BO (Rj+1
0
))) = Ω(e−γR/2 ) for
BQ (R) ∩ (BO (Rj0 ) \ BO (Rj+1
0


) ∩P =6 ∅. some γ < 1. Hence, the expected number of vertices
inside BQ (R) ∩ (BO (Rj0 ) \ BO (Rj+1
0
)) is n1−γ . By
0 0
• If RT < rQ ≤ RT −1 , then w.e.p. we have Chernoff bounds the first part of the lemma follows.
 For the second part, by Lemma 2.3 (applied with
BQ (R) ∩ BO (R/2) ∩ P = 6 ∅. ρA = R, ρO = R/2, and RT0 < rA = rQ ≤ RT0 −1 ),
since rQ > R/2 still holds, we have
Proof. First, assume 1 ≤ j ≤ T − 1. By Lemma 2.3 1 R
µ(BQ (R) ∩ BO (R/2)) = Θ(e−(α− 2 ) 2 − 2 rQ )
1

(applied with ρA = R, ρO = Rj0 , and Rj0 < rA = 1 R 1 0

0
rQ ≤ Rj−1 ), since Rj0 + rQ > R still holds, we have = Ω(e−(α− 2 ) 2 − 2 RT −1 ).

1
µ(BQ (R) ∩ BO (Rj0 )) = Θ(e−(α− 2 )(R−Rj )− 2 rQ )
0 1
Note that RT0 −1 ≤ R2 + R(1 − c)(1 − α) must hold,
R j 1
as otherwise RT0 +1 = RT0 −1 − R(1 − c)(1 − α) > R/2
= Θ(e− 2 (2α−1)(1−c+ 2 (1−c)(1−α))− 2 rQ ), would hold, and then c − T +1 1
2 (1 − c)(1 − α) > 2 ,
contradicting the definition of T . Thus,
and also


−R α+(1−c)(1−α)
0
µ(BQ (R) ∩ BO (Rj+1 )) µ(BQ (R) ∩ BO (R/2)) = Ω(e 2 ).
R j+1 1
= Θ(e− 2 (2α−1)(1−c+ 2 (1−c)(1−α))− 2 rQ ). Since α + (1 − c)(1 − α) < 1, µ(BQ (R) ∩ BO (R/2)) =
0
n1−γ for some γ 0 > 0, and by the same argument as
Thus, before, the second part of the lemma follows. 
0
µ(BQ (R) ∩ BO (Rj+1 )) We also have the following lemma.
−R
= Θ(e 4 (2α−1)(1−c)(1−α) )µ(BQ (R) ∩ BO (Rj0 )), Lemma 3.5. Let u be a vertex with ru > Ri0 and let
u =: u0 , u1 , . . . , um be a center path starting at u, or
so we get
the longest subpath found in an attempt of finding a
(3.10) µ(BQ (R) ∩ (BO (Rj0 ) \ BO (Rj+1
0
))) center path starting at u, in case no center path is
found. Then
= (1 + o(1))µ(BQ (R) ∩ BO (Rj0 )).
|θu − θum | = o(logC0 n/n).
0
Now, since rQ ≤ Rj−1 ,
Proof. Let ` be such that R` < ru ≤ R`+1 , and
µ(BQ (R) ∩ BO (Rj0 )) note that ` ≥ i0 . Since by making paths longer,
R j 1 0 the regions exposed get only larger and differences in
= Ω e− 2 (2α−1)(1−c+ 2 (1−c)(1−α))− 2 Rj−1


R j j−1
angles get larger as well, it suffices to consider the case
= Ω e− 2 ((2α−1)(1−c+ 2 (1−c)(1−α))+c− 2 (1−c)(1−α))


where indeed a center path is found. Suppose there
R
−j(1−c)(1−α)2 +(1−c)(2α−1+ 1−α


exists such a center path starting from u, that is, a
= Ω e− 2 2 )+c . sequence of vertices denoted by u =: u0 , u1 , . . . , um ,
7

32 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 O
satisfying d(uj , uj+1 ) ≤ R for all 0 ≤ j ≤ m − 1,
R`−j < ruj ≤ R`−j+1 for any 0 ≤ j ≤ m − 1, R

and Ri0 −1 < rum ≤ Ri0 . By Remark 2.1, |θuj −

1
θuj+1 | ≤ 2e 2 (R−R`−j −R`−j+1 ) (1 + o(1)), and thus, by
the triangle inequality for angles,
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

m−1
X
|θu − θum | ≤ |θuj − θuj+1 |
j=0
m−1
1
X
≤ 2(1 + o(1)) e 2 (R−R`−j −R`−j+1 ) Ri0
j=0 Ri0 +1
m+i
X0 1
= 2(1 + o(1)) e 2 (R−Ri−1 −Ri ) Q R`
R`+1
i=i0 +1
1 A0 (Ri0 ) R`0
X
≤ 2(1 + o(1)) 2e 2 (R−Ri −Ri+1 ) .
1 0
i≥i0 A (Ri0 ) \ A (Ri0 )

i A2 (Ri0 ) \ A1 (Ri0 )
Since Ri = Re−α /2
= (1+o(1))R(1−αi /2) for i ≥ i0 ,
A3 (R`+1 ) \ A2 (Ri0 )
i i+1
− 12 R (1+o(1)) R (α )
X
+α
|θu − θum | ≤ (2 + o(1))e e 4

i≥i0
Figure 2: Illustration of sequence of exposures of
− 12 R 1
2 (C0 +C0 α)(1+o(1))
= O(e R ) procedure Expose.
1
= O(n−1 (log n) 2 C0 (1+α)(1+o(1)) )
= o(logC0 n/n), having no center path starting at v0 = vk0 := Q is
at most c < 1. If we find a center path starting at
where the last equality is a consequence of α < 1. vk0 , we have reached our goal and stop with success.
The statement follows.  Otherwise, we will expose regions around the location
of Q in different exposing phases. Specifically, fix
We are now ready for the main lemma. ε > 0 to be any arbitrary small constant. Then, for
0 ≤ j ≤ C 00 log n, where C 00 is a sufficiently large
Lemma 3.6. Let Q ∈ Q be a vertex added. With constant, define
probability at least 1 − o(n−3/2 ), Q either connects
to a vertex in P ∩ BO (Ri0 ) through a path of length Aj (ρ) := {P | rP > ρ, |θP − θQ | ≤ j+1 C0 +ε
C0 +1+o(1) n log n}
O(log n), or any path starting at Q has
length at most O(logC0 +1+o(1) n). and perform the Expose procedure whose pseu-
docode is given in Algorithm 1 (see illustration in
Proof. Suppose first that rQ ≤ Ri0 . By applying Figure 2).
iteratively either Lemma 3.3 and Lemma 3.4 (if Rj0 < If ’no path’ is reported in Line 8 of Expose
rQ ≤ Ri0 ), or only Lemma 3.4 (if rQ ≤ Rj0 ), w.e.p., during phase j + 1, then among the paths starting
Q is connected to a vertex in BO (R/2) through a path from v0 with all vertices vi satisfying rvi > Ri0 , there
of length O(log R) + O(T ) = O(log log n). is no path with its last vertex vkj+1 ∈ Aj+1 (Ri0 ) \
Otherwise, suppose now rQ > Ri0 . If rQ > R − ξ Aj (Ri0 ) satisfying rvkj+1 ≤ R − ξ. We will show now
for some ξ > 0, then by Lemma 3.1, with probability that the length of any such path is with probability
at least 1 − o(n−3/2 ), either all paths starting at at least 1 − o(n−3/2 ) bounded by O(j logC0 +ε n):
Q have length O(log n), or after at most O(log n) indeed, partition the disc of radius R centered at the
steps we reach a vertex v with rv ≤ R − ξ. Since origin into Θ(n/ log n) equal sized sectors of angle
an additional O(log n) steps will be negligible in the C 0 log n/n for C 0 > 0 sufficiently large. By Chernoff
following, we may thus assume rQ ≤ R − ξ. Let `0 bounds for Poisson variables (see for example [AS08,
be the smallest integer i such that Ri > R − ξ. Theorem A.1.15]) together with a union bound over
Let ` be such that R` < rQ ≤ R`+1 and note all Θ(n/ log n) sectors, with probability at least 1 −
that ` ≥ i0 . By Lemma 3.2, the probability of o(n−3/2 ), the number of vertices in each sector is

33 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 Algorithm 1 Sequence in which regions are exposed
1: procedure Expose(Q, ξ)
2: Let `0 be the smallest integer such that R`0 > R − ξ
3: for j = 0, . . . , C 00 log n do
4: Expose A0 (Ri0 )
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
5: ` ← `0 + 1
6: repeat
7: `←`−1
8: if ` = i0 then return ’no path’ and stop
9: Expose Aj+1 (R` ) \ Aj (Ri0 )
10: until ∃Q =: v0 , . . . , vkj+1 ∈ P a path, vi ∈ Aj (Ri0 ) ∪ A
j+1 (R` ) for i < kj+1 ,
vkj+1 ∈ Aj+1 (R` ) \ Aj (Ri0 ) and rvkj+1 ≤ R − ξ
11: if ∃ center path starting from vkj+1 then
12: return ’success’ and stop
13: return ’failure’ and stop
Θ(log n). Thus, with this probability, the number
of vertices in Aj (Ri0 ) is O(j logC0 +ε n). By (3.7)
any two adjacent vertices w, w0 with rw , rw 0
> Ri0
C0 +o(1)
satisfy |θw − θw0 | = O(log n/n). If a path
starting from v0 with all vertices vi satisfying rvi >
Ri0 ends with some vertex v ∈ Aj (Ri0 ), then the
path does not extend beyond Aj+1 (Ri0 ), and thus in
this case, with probability at least 1 − o(n−3/2 ), its
length is at most the number of vertices in Aj (Ri0 ),
i.e., O(j logC0 +ε n). If the path ends with a vertex
v ∈ Aj+1 (Ri0 ) \ Aj (Ri0 ), then note that rv > R − ξ
must hold. Consider then the last vertex w in this
path with w ∈ Aj (Ri0 ). Since the next vertex w0
satisfies w0 ∈ Aj+1 (Ri0 ) \ Aj (Ri0 ) and rw0
> R − ξ, by
calculations as in (3.7), it must hold that |θw −θw0 | =
O(logC0 /2+o(1) n/n). Moreover, by Corollary 3.1,
with probability at least 1−o(n−3/2 ), any two vertices
w00 , w000 on a path containing only vertices inside
Aj+1 (Ri0 )\Aj (Ri0 ) satisfy |θw00 −θw000 | = O(log n/n).
Thus, also |θw −θv | = O(logC0 /2+o(1) n/n). Since this
would also have to hold for another possible neighbor
z of v with rz > Ri0 , and z ∈ / Aj+1 (Ri0 ), starting
from v the path cannot extend to a vertex outside
Aj+1 (Ri0 ) as well. Hence, with probability at least
1 − o(n−3/2 ), the number of vertices of such a path is
bounded by O((j + 1) logC0 +ε n).
If on the other hand the condition of Line 10 of
Expose is satisfied during phase j, then for some
i0 ≤ `0 ≤ `0 there is a path v0 , . . . , vkj+1 ∈ P with
vi ∈ Aj (Ri0 ) ∪ Aj+1 (R`0 ) for all i < kj+1 , vkj+1 ∈
Aj+1 (R`0 ) \ Aj (Ri0 ). Recall that rvkj+1 ≤ R − ξ, and
observe that R`0 < rvkj+1 ≤ R`0 +1 . Consider then
the restriction of a center path starting from vkj+1 to
9
34
. Start of (j + 1)-th phase
. (`0 − ` + 1)-th sub-phase of the (j + 1)-th phase
the region
Sj+1 :=
{P | j+1
n logC0 +ε n < |θP − θQ | ≤ j+2
n logC0 +ε n}.
Note first that since R`0 < rvkj+1 ≤ R`0 +1 , the first
region to be exposed on the center path starting from
vkj+1 contains vertices v with R`0 −1 < rv ≤ R`0 ,
and this region when restricted to Sj+1 has not been
exposed before. In particular, the regions exposed by
the center paths of vki and vkj+1 , whenever restricted
to Si and Sj+1 , respectively, for i < j + 1 are
disjoint. Suppose without loss of generality that
|θQ − θvkj+1 | ≤ (j + 32 ) logC0 +ε n/n. Observe that
by Lemma 3.5 the angle between any vertex v with
rv > Ri0 and the terminal vertex of the center path
starting at v or the longest subpath in an attempt
of building a center path starting at v, in case a
center path is not found, is bounded by o(logC0 n/n).
Therefore, at least half of the region to be exposed by
a center path in each step is retained in Sj+1 : more
precisely (see Figure 3), in the first step the region
containing all points P that were to be exposed in
a center path without restrictions and that satisfy
|θQ − θP | ≥ |θQ − θvkj+1 | is included in Sj+1 , and
at every subsequent step on the center path starting
from a vertex u, the region of all points P that were to
be exposed in a center path without restrictions and
that satisfy |θQ − θP | ≥ |θQ − θu | is always included
in Sj+1 . Lemma 3.2 can still be applied, and the
probability of having no center path starting at vkj+1
is at most c0 < 1, and by disjointness, this is clearly
independent of all previous events.
The probability that the exposing process returns
’failure’ is the probability that during C 00 log n phases
Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 O Proof. Let Q ∈ Q be a vertex added in the Pois-
sonized model. By Lemma 3.6, with probability at
least 1 − o(n−3/2 ), either Q connects to a vertex in
BO (R/2) using O(logC0 +1+o(1) n) steps, or all paths
R starting from Q have length O(logC0 +1+o(1) n). Note
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

that all vertices inside BO (R/2) form a clique, and

by Lemma 2.2 together with Chernoff bounds there
are w.e.p. Θ(n1−α ) vertices inside BO (R/2). By de-
Poissonizing the model, using (2.6), the same state-
ment holds with probability at least 1−o(n−1 ) for our
vk3
A particular choice of Q in the uniform model. Using
Ri0 a union bound over all n vertices, with probability
Ri0 +1 1 − o(1) the same holds for all vertices simultane-
B
R`0
ously. Thus, the maximum graph distance between
Q R`0 +1
any two vertices of the same connected component is
O(logC0 +1+o(1) n), and the statement follows. 

By the result of Bode, Fountoulakis and

Müller [BFM13], a.a.s., all vertices belonging to
Figure 3: The region that has already been exposed is BO (R/2) are part of a giant component. We call this
lightly shaded. The AOB slice of BO (R) corresponds the center giant component, and we will show that
to Sj+1 . The darkly shaded area contains the region there is no other giant component.
to be exposed in a center path without restrictions
Corollary 3.2. A.a.s., the size of the second
starting at vk3 .
largest component of Gα,C (n) is O(log2C0 +1+o(1) n).

no center path is found. By independence, for Proof. Having de-Poissonized some results in the
previous proof, we may now work directly in the
C 00 (c) sufficiently large, this probability is at most
C 00 log n model Gα,C (n) during this proof. As in the proof of
c0 = o(n−3/2 ). Thus, with probability at
Lemma 3.6, partition the disc of radius R centered at
least 1 − o(n−3/2 ), the exposing process ends with
the origin into Θ(n/ log n) equal sized sectors of angle
’no path’ in some phase j, or with ’success’ in
C 0 log n/n for some large constant C 0 > 0 and order
some phase j 0 . If the exposing process ends with
them counterclockwise. The distance between sectors
’no path’ in some phase j, with probability 1 −
is measured according to this ordering: two sectors
o(n−3/2 ) any path starting at v0 containing only
are at distance D from each other, if the shorter of the
vertices v with rv > Ri0 has length at most O((j +
two routes around the clock between the two sectors
1) logC0 +ε n) = O(logC0 +1+ε n). If the exposing
has exactly D − 1 sectors in between.
process ends with ’success’ in some phase j 0 , a
Now, by the proof of Theorem 3.1, a.a.s., ev-
center path of length O(log R) starting from vkj0 is
ery vertex v with rv ≤ Ri0 connects to BO (R/2),
found, and as in the case of ’no path’, the total
so a.a.s., every such vertex is part of the center
length of the path starting at v0 together with the
giant component. We may assume that all other
center path is at most O(logC0 +1+ε n). Thus, with
vertices v in other components satisfy rv > Ri0 ,
probability at least 1 − o(n−3/2 ), v0 either connects
and consider only these components from now on.
to a vertex in P ∩ BO (Ri0 ) through a path of length
By Theorem 3.1, a.a.s., for any two vertices of the
O(logC0 +1+ε n), or any path starting at v0 has length
same component, their graph distance is bounded by
at most O(logC0 +1+ε n), and the statement follows,
O(logC0 +1+o(1) n). Recall that for any two vertices
since ε can be chosen arbitrarily small. 
vi , vj with d(vi , vj ) ≤ R and rvi , rvj > Ri0 , by (3.7),
C0 +o(1)
The upper bound for the diameter now follows easily. |θ vi − θvj | = O((log n) /n), and therefore, by
the triangle inequality for angles, for any two ver-
tices u and v belonging to the same component, a.a.s.
Theorem 3.1. A.a.s., any two vertices u and v we have |θu − θv | = O((log n)2C0 +1+o(1) /n). Since
belonging to the same connected component satisfy for any two vertices u and v belonging to sectors at
dG (u, v) = O(logC0 +1+o(1) n). distance 2 ≤ D = o(n/ log n) from each other we

35 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 have |θu − θv | = Θ(D log n/n), it must hold that be- get that the second expression equals
tween any two vertices u and v of the same compo- Z R−c1 Z θR (r,rA )
nent, there are at most O(log2C0 +o(1) n) sectors. By f (r)
2 dθdr
Chernoff bounds for binomial random variables to- R−c2 θR−c3 (r,rA ) 2π
gether with a union bound over all sectors, for C 0 > 0 e(R−rA )/2 (1 − e−c3 /2 )
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

large enough, a.a.s., in any sector there are at most =2

πC(α, R)
O(log n) vertices. Thus, a.a.s., the number of ver- Z R−c1  
tices of the second largest component is bounded by × e−r/2 α sinh(αr) 1 + O(e−r ) dr
O(log2C0 +1+o(1) n).  R−c2
Ω(1)  (α− 1 )R 
4 The lower bound on the diameter = e 2 + O(e(α−3/2)R ) .
C(α, R)
In this section we show that the diameter of a random The desired conclusion follows recalling that
hyperbolic graph is Ω(log n). In fact, we do more, C(α, R) = cosh(αR) − 1 = 12 eαR + O(e−αR ). 
since we actually establish that the same can be said
about the existence of a component forming a path Lemma 4.2. Let 0 < c3 < c1 < c2 be small positive
of length Θ(log n). constants so that 2ec1 −c2 > ec3 /2 holds. Suppose
To achieve this section’s main objective we rely R − c2 ≤ rA , rB ≤ R − c1 and R − c3 ≤ d(A, B) ≤ R.
on a couple of bounds concerning the measure of Then
regions defined in terms of set algebraic manipulation
µ [(BB (R)\BB (R−c3 ))
of distinct balls. In the following section we derive

these bounds. The reader, however, might prefer to ∩ (BO (R−c1 )\BO (R−c2 ))]\BA (R)
skip their proofs upon first reading, and come back = Ω(e−R/2 ) = Ω(1/n).
to them once their intended use is understood.
Proof. By Lemma 4.1, Part 2, we know that BB (R) \
4.1 Useful bounds. The vertices of the path com- BB (R − c3 ) intersects the band at distance between
ponent whose existent we will establish will be shown R − c2 and R − c1 , i.e. BO (R − c1 ) \ BO (R − c2 ),
−R/2
to be at distance R − Ω(1) of the origin O. This in a region of measure Ω(e ). Note that the
explains why below we focus on approximating the intersection comprises two disconnected regions of
0
measure of regions which are close to the boundary. equal measure, say D and D . We may assume that
A ∈ D and also that for all points P ∈ D and P 0 ∈ D0
we have θP < θB < θP 0 . We will show that BA (R)
Lemma 4.1. The following statements hold: does not intersect D0 : suppose for contradiction that
P ∈ D and P 0 ∈ D0 are within distance R: then, by
1. If 0 ≤ ρ0 ≤ ρ < R, then Remark 2.1, we would have
O O
1
(4.11) |θP − θP 0 | ≤ (2 + o(1))e 2 (R−2(R−c2 ))
µ(BO (ρO ) \ BO (ρ0O )) = (2 + o(1))e−R/2 ec2 .
0
= e−α(R−ρO ) (1 − e−α(ρO −ρO ) + o(1)). On the other hand, for any P ∈ D, since d(P, B) ≥
R − c3 > R − c1 > R − c2 , we have again by
1
Remark 2.1, |θB −θP | ≥ (2+o(1))e 2 (R−c3 −2(R−c1 )) =
c3
(2 + o(1))e−R/2 ec1 − 2 , and the same bound holds for
2. If 0 < c1 < c2 are two positive constants, R − c2 ≤ |θB − θP 0 |. Since P and P 0 satisfy θP < θB < θP 0 ,
rA ≤ R − c1 , and c3 = Ω(1), then we have |θP − θP 0 | = |θB − θP | + |θB − θP 0 |, and
c3
thus |θP − θP 0 | ≥ (4 + o(1))e−R/2 ec1 − 2 . Since by
c3
assumption 2ec1 − 2 > ec2 , this contradicts (4.11).
µ((BA (R)\BA (R−c3 )) ∩ (BO (R−c1 )\BO (R−c2 )))
The lemma follows. 
= Ω(e−R/2 ).
4.2 The existence of a path component of
length Θ(log n). The main objective of this section
is to show that inside a band at constant radial
Proof. The first part of the lemma follows directly distance from the boundary we find Θ(log n) vertices
from Lemma 2.2. Again, relying on Lemma 2.1, we forming a path.
11

36 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 Theorem 4.1. A.a.s., there exists a component for 1 ≤ j ≤ m, the event Ej that occurs when
forming a path of length Θ(log n).
Proof. As before, we work first in the Poissonized proceed with the next j; also stop if all conditions
model and derive in the end from it the result in the hold for one j):
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy
uniform model. Fix throughout the proof c1 , c2 , c3
three positive constants such that c3 < c1 < c2 and
2ec1 −c2 < e−c3 /2 . First, by Lemma 4.1, Part 1,
applied with ρ0 = R − c1 and ρ00 = R − c2 we have
µ(BO (ρO ) \ BO (ρ0O )) = e−αc1 (1 − e−α(c2 −c1 ) + o(1))
= Θ(1).
Let Θ ⊆ [0, 2π) be a set of forbidden angles such
that µ(RΘ ) < 1, where RΘ := {(rP , θP ) : 0 ≤ rP <
R, θP ∈ Θ} (for a geometric interpretation of RΘ ,
note that when Θ is an interval, RΘ is a cone with
vertex O). As a constant fraction of the angles is still
allowed, clearly,
µ( (BO (ρO ) \ BO (ρ0O )) \ RΘ = Θ(1)


(4.12)
still holds. For any vertex A with R − c2 ≤ rA ≤
R − c1 , by Lemma 2.3 (applied with ρA = R and
ρO = R)
(4.13)
2α 1
µ(BA (R) ∩ BO (R)) = e− 2 rA + O(e−αrA )
π(α − 12 )
= Θ(1/n),
which together with Lemma 4.1, Part 2, yields
(4.14)
µ( BA (R)\ BA (R−c3 ) ∩ BO (ρO )\ BO (ρ0O ) )



= Θ(e−R/2 ) = Θ(1/n).
For two vertices A, B with R−c3 ≤ d(A, B) ≤ R that
satisfy ρ0O ≤ rA , rB ≤ ρO , by Lemma 4.2, together
with (4.13),
(4.15) µ [(BB (R)\BB (R−c3 ))
∩ (BO (ρO )\BO (ρ0O ))]\BA (R)


= Θ(e−R/2 ) = Θ(1/n).
Let ε = ε(α) be a constant chosen small enough so
1
that 1 − 2α − ε > 0. Let ν > 0 be a sufficiently
1
small constant. Let m := νn1− 2α −ε , and note
that since α < 1, we have m = o(n1/2 ). We
will add in the following, if necessary, up to m
vertices Q1 , . . . , Qm ∈ Q to P, all of them chosen
independently and following the same distribution
as in Gα,C (n). In order to be more precise, define
following conditions hold (if one condition fails, then
stop exposing and checking further conditions and
• Condition 1: Add the j-th vertex Qj ∈ Q and
let Aj0 := Qj . We require ρ0O ≤ rQj ≤ ρO , and
for j ≥ 2, we additionally require the coordinates
of Aj0 and Ak0 for any 1 ≤ k < j to be sufficiently
different in terms of their angles, i.e., letting
1
Θk = {θ : |θAk0 − θ| ≤ C 00 n−(1− 2α −ε) } for some
large constant C 0 , we require that θAj 6∈ ∪j−1
k=1 Θk .
0
• Condition 2: Expose the region BAj (R) \
0
BAj (R − c3 ) ∩ BO (ρO ) \ BO (ρ0O ) . We require



0
exactly one vertex in this region, call it Aj1 .
Let L := ν 0 log n with ν 0 > 0 being a small
constant. Then, inductively, for 1 ≤ i <
L,
expose the region BAj (R) \ BAj (R − c3 ) ∩
i i
BO (ρO ) \ BO (ρ0O ) \ BAj (R). We require also


i−1
exactly one vertex in this region, and name it
inductively Aji+1 . In other words, Aji+1 belongs
to the (R−c3 , R)-band centered at Aji and also to
the (ρ0O , ρO )-band centered at the origin O. For
i = L, expose the region BAj (R) \ BAj (R −
L L
c3 ) ∩ BO (ρO ) \ BO (ρ0O ) \ BAj (R), and we



L−1
require that there is no more vertex in this
region.
• Condition 3: For any 0 ≤ i ≤ L, expose
1
BAj (R) ∩ (BO (R) \ (BO (R(1 − 2α − ε))), and
i
we require that Aji has no other neighbor inside
this region except the one(s) from the previous
condition.
We will now bound from above the probability that
for all 1 ≤ j ≤ m the events Ej fail. Note first
that for ν sufficiently small, the set of forbidden
angles Θj := ∪j−1
k=1 Θk still is such that µ(RΘj ) < 1,
and (4.12) can be applied with RΘ = RΘj for any
1 ≤ j ≤ m. Hence, for any j, independently of
the outcomes of previous events, there is an absolute
constant c > 0 such that for this j the probability
that Condition 1 holds is at least c. Given that
Condition 1 holds, by (4.14) applied to Aj0 and (4.15)
applied successively to Aj1 , . . . , AjL we get that the
probability that Condition 2 holds is at least c0−L for
some fixed 0 < c0 < 1. Suppose then that Condition 2
37 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 also holds. By a union bound and by (4.13), we have for some positive ξ > 0. Thus, w.e.p. there exists one
L j, for which the event Ej holds.
In order to de-Poissonize, let EPj be the event
[
1
µ( [BAj (R) ∩ (BO (R) \ BO (R(1 − 2α − ε))])
i
i=0 that in the Poissonized model there exists some j
L for which the event Ej holds, and similarly EUj the
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

X
≤ µ(BAj (R) ∩ BO (R)) = O(L/n), corresponding event for the uniform model. Since we
i ξ
i=0 have shown that P(EPj ) = 1 − e−Ω(n ) , and since m =
−1/2
and thus for ν 0 sufficiently small the probability that o(n ), using (2.6), we also have for some function
j
Condition 3 holds is at least n −η
for some fixed ω n tending to infinity arbitrarily slowly P(EU ) ≥ 1 −
√ ξ ξ
value η which can be made small by choosing ν 0 ωn ne−Ω(n ) = 1 − e−Ω(n ) . Let ρO = R(1 − 2α 1
− ε).
small (in fact, part of the region might have been Denote by NU the random variable counting the
already exposed in Condition 2, but since we know number of vertices inside BO (ρO ) in Gα,C (n). By
there are no other vertices in there, this only helps). Lemma 2.2, µ(BO (ρO )) = O(e−R/2−αεR ) = o(1/n),
Altogether, we obtain and thus ENU = o(1), and by Markov’s inequality,
0 P(NU = 0) = 1 − o(1). Since
P(Ej ) ≥ n−η
for some η 0 > 0. Again, η 0 can be made sufficiently 1 − o(1) = P(NU = 0)
c c
small by making the constant ν 0 (and thus L = = P(NU = 0|EUj )P(EUj ) + P(NU = 0|EUj )P(EUj )
ν 0 log n) sufficiently small.
Now, since d(Aji−1 , Aji ) ≤ R and rAj , rAj ≥ = (1 − o(1))P(NU = 0|EUj ) + o(1),
i−1 i
R − c2 , by Remark 2.1, we have |θAj − θAj | = we have P(NU = 0|EUj ) = 1 − o(1). Thus,
i−1 i

O(e−R/2 ), and therefore |θAj − θAj | = O(log n/n). P(NU = 0, EUj ) = 1 − o(1). If this event holds, this
0 L
Since by construction for j 6= j 0 we have |θAj − means that there is no vertex inside BO (ρO ), and the
0 j j
θAj0 | = ω(log n/n), for j 6= j 0 the regions exposed vertices A0 , . . . , AL form an induced path, yielding
0 the desired result. 
in Condition 2 are disjoint. Also, for any j 6= j 0 ,
if the region in Condition 3 containing one of the
0 5 Conclusion
vertices Aji or Aji0 is exposed, then BAj (R)∩(BO (R)\
1
i We have shown that in random hyperbolic graphs
(BO (R(1 − 2α − ε))) is disjoint from BAj0 (R) ∩ a.a.s. the diameter of the giant component is
i0
1 O(polylog(n)), the size of the second largest com-
(BO (R) \ (BO (R(1 − 2α − ε))): indeed, note that for
1 ponent O(polylog(n)), and at the same time there
any point P ∈ BAj (R)∩(BO (R)\(BO (R(1− 2α −ε)))
i
we have by Remark 2.1 exists a path component of length Θ(log n). It is an
1 1 interesting and challenging problem to tighten these
|θAj − θP | ≤ (2 + o(1))e 2 (R−(R−c2 )−R(1− 2α −ε)) bounds by improving the exponents of the polylog(n)
i
1 1
= (2 + o(1))e 2 c2 n−(1− 2α −ε) . terms established in this work.
1
By construction, we have |θAj −θAj0 | ≥ C 0 n−(1− 2α −ε) References
0 0
for some large enough C 0 > 0, and thus by the trian-
gle inequality for angles there exists C 00 := C 00 (C 0 ) >
1 [AS08] N. Alon and J. Spencer. The Probabilistic
0 such that |θAj − θAj0 | ≥ (1 + o(1))C 00 n−(1− 2α −ε) Method, Third Edition. John Wiley & Sons, 2008.
i i0
holds for any i, i0 and any j 6= j 0 . Hence, the regions [BFM] M. Bode, N. Fountoulakis, and T. Müller.
exposed in Condition 3 are disjoint, and by the same The probability that the hyperbolic random
reason, the region exposed in Condition 2 for some j graph is connected. Preprint available at
and the one exposed in Condition 3 for some j 6= j 0 www.math.uu.nl/∼Muell001/Papers/BFM.pdf.
[BFM13] M. Bode, N. Fountoulakis, and T. Müller.
are disjoint as well. It follows that the probabilities
On the giant component of random hyperbolic
of the corresponding conditions to hold are thus inde- graphs. In Proceedings of the 7th European Con-
0 0
pendent. Hence, for ν sufficiently small, η is small ference on Combinatorics, Graph Theory and Ap-
enough such that −η 0 + 1 − 2α 1
− ε > 0, and plications, EUROCOMB’13, pages 425–429, 2013.
m [BnPK10] M. Boguñá, F. Papadopoulos, and D. Kri-
Y 0 ξ
P Ej c ≤ (1 − n−η )m = e−Ω(n ) ,


oukov. Sustaining the internet with hyperbolic map-
j=1 ping. Nature Communications, 1:62, 2010.
13

38 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.
 [CF13] E. Candellero and N. Fountoulakis. Clustering in
random geometric graphs on the hyperbolic plane.
ArXiv e-prints, September 2013. Preprint available
at arxiv.org/abs/1309.0459.
[Fou12] N. Fountoulakis. On the evolution of random
Downloaded 05/09/23 to 179.7.65.23 . Redistribution subject to SIAM license or copyright; see https://epubs.siam.org/terms-privacy

graphs on spaces with negative curvature. ArXiv

e-prints, May 2012.
[GPP12] L. Gugelmann, K. Panagiotou, and U. Peter.
Random hyperbolic graphs: Degree sequence and
clustering. In Proceedings of the 39th International
Colloquium Conference on Automata, Languages,
and Programming - Volume Part II, ICALP’12,
pages 573–585. Springer-Verlag, 2012.
[KPK+ 10] D. Krioukov, F. Papadopoulos, M. Kitsak,
A. Vahdat, and M. Boguñá. Hyperbolic geometry of
complex networks. Phys. Rev. E, 82:036106, 2010.
[Pen03] M. Penrose. Random Geometric Graphs. Oxford
Studies in Probability, Oxford U. P., 2003.
[PKBnV10] F. Papadopoulos, D. Krioukov, M. Boguñá,
and A. Vahdat. Greedy forwarding in dynamic
scale-free networks embedded in hyperbolic metric
spaces. In Proceedings of the 29th Conference on
Information Communications, INFOCOM’10, pages
2973–2981. IEEE Press, 2010.

39 Copyright © 2015.
by the Society for Industrial and Applied Mathematics.