Contents lists available atScienceDirect
Statistics and Probability Letters
journal homepage:www.elsevier.com/locate/stapro
The almost-sure population growth rate in branching Brownian motion with a quadratic breeding potential
J. Berestycki
a, É. Brunet
b, J.W. Harris
c,∗, S.C. Harris
daLaboratoire de Probabilités et Modèles Aléatoires, Université Pierre et Marie Curie (UPMC Paris VI), CNRS UMR 7599, 175 rue du Chevaleret, 75013 Paris, France
bLaboratoire de Physique Statistique, École Normale Supérieure, UPMC Université Paris 6, Université Paris Diderot, CNRS, 24 rue Lhomond, 75005 Paris, France
cDepartment of Mathematics, University of Bristol, University Walk, Bristol, BS8 1TW, UK
dDepartment of Mathematical Sciences, University of Bath, Claverton Down, Bath, BA2 7AY, UK
a r t i c l e i n f o
Article history:
Received 10 February 2010
Received in revised form 14 May 2010 Accepted 17 May 2010
Available online 25 May 2010 MSC:
60J80 Keyword:
Branching Brownian motion
a b s t r a c t
In this note we consider a branching Brownian motion (BBM) onRin which a particle at spatial positionysplits into two at rateβy2, whereβ > 0 is a constant. This is a critical breeding rate for BBM in the sense that the expected population size blows up in finite time while the population size remains finite, almost surely, for all time. We find an asymptotic for the almost-sure rate of growth of the population.
©2010 Elsevier B.V. All rights reserved.
1. Introduction
We consider a branching Brownian motion with a quadratic breeding potential. Each particle diffuses as a driftless Brownian motion onR, and splits into two particles at rate
β
y2, whereβ >
0 andyis the spatial position of the particle.We letNt be the set of particles alive at timet, and then, for eachu
∈
Nt,Yu(
t)
is the spatial position of particleuat time t(and for 0≤
s<
t,
Yu(
s)
is the spatial position of the unique ancestor ofualive at times). We will call this process the(β
y2;
R)
-BBM.It is known that quadratic breeding is a critical rate for population explosions. If the breeding rate were instead
β |
y|
pfor p>
2, the population size would almost surely explode in a finite time. However, for the(β
y2;
R)
-BBM theexpectednumber of particles blows up in a finite time while the total number of particles alive remains finite almost surely, for all time. For p∈ [
0,
2)
the expected population size remains finite for all time. SeeItô and McKean(1965,pp. 200–211) for a proof of these facts using solutions to related differential equations. In this note we prove the following result on the almost-sure rate of growth of|
Nt|
.Theorem 1. Suppose that the initial configuration consists of a finite number of particles at arbitrary positions inR. Then, almost surely,
lim
t→∞
ln ln
|
Nt|
t=
2p2
β.
∗Corresponding author. Tel.: +44 0 7812 107318.
E-mail addresses:julien.berestycki@upmc.fr(J. Berestycki),eric.brunet@lps.ens.fr(É. Brunet),john.harris@bristol.ac.uk(J.W. Harris), s.c.harris@bath.ac.uk(S.C. Harris).
0167-7152/$ – see front matter©2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.spl.2010.05.011
We defineRt
:=
maxu∈NtYu(
t)
to be the right-most particle in the(β
y2;
R)
-BBM. InHarris and Harris(2009) it was shown thattlim→∞
lnRt t
=
p2
β
(1)almost surely, and this result is crucial in our proof ofTheorem 1because it allows us both some control over the maximum breeding rate in the BBM, and also to show that the growth rate of
|
Nt|
is dominated by particles with spatial positions near Rt. Indeed, a BBM in which every particle has branching rateβ
R2t would see its population grow as inTheorem 1, and this point will provide our upper bound. Furthermore we shall see that, for anyδ >
0, at all sufficiently large timest, a single particle located nearRtwill single-handedly build during time interval[
t,
t+ δ ]
a progeny so large that it is again of the magnitude given byTheorem 1, which will prove the lower bound.2. Proof of the growth rate
To proveTheorem 1, it is sufficient to consider the case of the initial configuration of particles being a single particle at positionx
∈
R.Upper bound. Since the breeding potential is symmetric about the origin, we have from Eq.(1)that for
ε >
0 fixed, there exists, almost surely, a random timeτ < ∞
such that for allu∈
Nt,ln
|
Yu(
t) | <
p 2β + ε
t
,
for allt> τ.
As a consequence of this, the breeding rate of any particle in the population is bounded by
β ¯
t:= β
e2(
√2β+ε)
tfor allt> τ
. We now introduce a coupled branching process(
N¯
t
,
t≥
0)
as follows. For eacht≥
0 it consists in a population of particles{
u:
u∈ ¯
Nt}
whose positions are denoted by{ ¯
Yu(
t) :
u∈ ¯
Nt}
. Until timeτ
the two processesNt andN¯
tcoincide and for allt
≥ τ
we will haveNt⊆ ¯
Nt. Furthermore, we want that after timeτ
all particles inN¯
tbranch at rate
β ¯
t. In order for this to make sense we must construct(
N¯
t
,
t≥
0)
conditionally on(
Nt,
t≥
0)
. Given(
Nt,
t≥
0)
, we let(
N¯
t
,
t≤ τ) = (
Nt,
t≤ τ)
. Fort≥ τ
, each particleu∈
Ntgives birth to an extra particlev ∈ ¯
Nt\
Ntat rateβ ¯
t− β
Yu(
t)
2(note that ifu∈
NtthenYu(
t) = ¯
Yu(
t)
). Each particle thus created starts an independent BBM in(
N¯
t
,
t≥
0)
with time-dependent branching rateβ ¯
tat timet. Thus it is clear that all particles inN¯
tbranch at rate
β ¯
tfort≥ τ
. It is furthermore clear that|
Nt| ≤ | ¯
Nt|
for allt
≥
0. The upshot is that after timeτ
,N¯
tis a pure birth process, and as such is very well studied.
Consider
(
Zt,
t≥ τ)
a pure birth process starting with a single particle at timeτ
with inhomogeneous rate( β ¯
t,
t≥ τ)
, and defineVt
:=
infs
≥ τ :
Z sτ
β ¯
udu≥
t.
It is easily seen that
(
ZVt,
t≥
0)
is simply a Yule process (i.e. a pure birth process where particles split into two at rate 1) and so we know thatZVte−t
→
W∞ ast→ ∞
almost surely, whereW∞has an exponential mean 1 distribution. However, sinceRVt
τ
β ¯
udu=
twe see that this implies Ztexp
−
Z tτ
β ¯
sds→
W∞ ast→ ∞ .
AsN¯
tfort
≥ τ
is composed of the union of the offspring of the|
Nτ|
particles present at timeτ
, we can write Mt:= | ¯
Nt|
exp
−
Z tτ
β ¯
sds=
|Nτ|
X
i=1
Zt(i)exp
−
Z tτ
β ¯
sdswhere theZ(i)are independent, identically distributed, copies of a pure birth process with birth rate
β ¯
t. By conditioning on the value of|
Nτ|
(which is almost surely finite), we see thatMt
→ ¯
M∞∈ (
0, ∞ )
ast→ ∞
where, more precisely,M¯
∞is distributed as the sum of
|
Nτ|
independent exponential variables with mean 1.Thus
ln
| ¯
Nt| −
Z tτ
β ¯
sds→
lnM¯
∞ ast
→ ∞
and sinceRt
τ
β ¯
sds=
1c
[
exp(
ct) −
exp(
cτ) ]
withc=
2√
2β + ε
, it follows that
tlim→∞
ln ln
| ¯
Nt|
t
=
2p2
β + ε ,
almost surely. Finally, since
ε
may be arbitrarily small, using the coupling constructed above, we have that lim supt→∞
ln ln
|
Nt|
t≤
2p2
β,
almost surely, as required.Lower bound. Fix
ε >
0. From(1), we know that there exists, almost surely, a random timeτ
0< ∞
such that for allt> τ
0,Rt
>
e(
√2β−ε)
t.
(2)Fix
δ >
0 and lettn:=
nδ,
n∈ {
0,
1,
2, . . . }
. We will show that it is sufficient to consider only the offspring of the rightmost particlesRtnto prove the lower bound for the global population growth rate. There are two steps to the argument.First, we will couple the sub-population descended fromRtnduring the time interval
[
tn,
tn+1]
to a BBM in which the expected population size remains finite. Second, we will show that, asntends to infinity, the population sizes of the coupled processes grow sufficiently quickly. For this, we use Eq.(2)to give us a lower bound on the breeding rate in the coupled processes.LetNs(n)be the set of descendants ofRtnat times
∈ [
tn,
tn+1]
. As for the upper bound we introduce a coupled process(
eNs(n),
s∈ [
tn,
tn+1] )
. This time we have that for alls∈ [
tn,
tn+1] ,
eNs(n)⊆
Ns(n)⊆
Ns. More precisely,(
eNs(n),
s∈ [
tn,
tn+1] )
is obtained from(
Ns(n),
s∈ [
tn,
tn+1] )
by cancelling some of the split events inNs(n), in the following way: if at timesa particle u∈
eNs(n)⊆
Ns(n)splits in the original processNs(n), it also splits ineNs(n)with probabilitye
β
n(
Yu(
s))
β
Yu(
s)
2∈ [
0,
1] ,
wheree
β
n(
x) :=
β
e2(
√2β−2ε)
tn ifx≥
e(
√2β−2ε)
tn,
0 otherwise
.
If the split event is rejected, one of the two offspring inNs(n)is chosen at random to be the one which we keep ineNs(n). The process
(
eNs(n),
tn≤
s≤
tn+1)
is thus a BBM started at timetn from a single particle at positionRtn with space- dependent branching rateeβ
n. Observe that we trivially haveeNs(n)⊆
Ns(n)for alls∈ [
tn,
tn+1]
, as announced, since we obtain eNs(n)fromNs(n)merely by erasing some particles fromN(n).We also require another process on the same probability space, coupled to
(
eNs(n),
s∈ [
tn,
tn+1] )
. This process is denoted(
bNs(n),
s∈ [
tn,
tn+1] )
, and is defined by adding particles to(
eNs(n),
s∈ [
tn,
tn+1] )
in such a way that every particle inbN(n)breeds at constant rateβ
e2(
√2β−2ε)
tn, irrespective of its spatial position.More specifically, fors
∈ [
tn,
tn+1]
, every particleu∈
eNs(n)gives birth to an extra particlev ∈
bNs(n)\
eNs(n)at rateβ
e2(
√2β−2ε)
tn1Yu(s)<e
(
√2β−2ε)
tn,
and each particle thus created initiates an independent BBM in
(
bNs(n),
s∈ [
tn,
tn+1] )
with constant breeding rateβ
e2(
√2β−2ε)
tn. Note that we haveeNs(n)⊆
bNs(n)for alls∈ [
tn,
tn+1]
.Now for
v ∈
Nt(nn+)1define the event An(v) :=
min
s∈[tn,tn+1]Yv
(
s) <
e(
√2β−2ε)
tn,
and setAn= ∪
v∈eN(n) tn+1
An
(v)
, i.e., the event that there exists a descendantv
ofRtn(in the modified processeN(n)) such that Yv(
s) <
e(
√2β−2ε)
tnfor somes∈ [
tn,
tn+1]
. Observe thatn
eNt(nn+)1
6=
bNt(nn+)1 o⊆
An.
(3)We also define the eventBnas Bn
:=
n
Rtn
>
e(
√2β−ε)
tno,
and let
{
Ft}
t≥0be the natural filtration for the(β
y2;
R)
-BBM. Our aim is to show that, almost surely, only finitely many of the eventsAnoccur, from which it follows that the populations of the coupled subprocesseseN(n)are equal to the populationsof thebN(n)processes for all sufficiently largen, almost surely. Then the final step is to show that the populationsbN(n)grow sufficiently quickly to imply the desired lower bound on the size of the original population.
We start by writing the event of interest as An
= (
An∩
Bn) ∪ (
An∩
Bcn),
and we recall from Eq.(1)thatP
(
lim supBcn) =
0, and hence only finitely many of the eventsAn∩
Bcnoccur. We now use a standard ‘many-to-one’ argument (see, for example,Hardy and Harris, 2009) to bound the probabilitiesP(
An∩
Bn|
Ftn)
.LetPxbe the law of a driftless Brownian motionYstarted at the pointx
∈
R(andExthe expectation with respect to this law). Observing thatBn∈
Ftn, we have1BnP
(
An|
Ftn) ≤
1BnE
X
v∈eN(n) tn+1
1An(v)
Ftn
=
1BnERtnexp Z δ
0
e
β
n(
Ys)
ds;
mins∈[0,δ]Ys
<
e(
√2β−2ε)
tn≤
expβ
e2(
√2β−2ε)
tnδ
Pe
(
√2β−ε)
tn mins∈[0,δ]Ys
<
e(
√2β−2ε)
tn.
Using the reflection principle, we obtain that there existsC>
0 such thatPe
(
√2β−ε)
tn mins∈[0,δ]Ys
<
e(
√2β−2ε)
tn=
2P0Yδ
<
e(
√2β−ε)
tn(
e−εtn−
1)
≤
Cexp−
1 2δ
e
(
√2β−ε)
tn e−εtn−
12.
Combining this series of inequalities gives, almost surely, a faster than exponentially decaying upper bound onP
(
An∩
Bn|
Ftn)
, and so we have shown thatX
n≥0
P
(
An∩
Bn|
Ftn) < ∞
almost surely. SinceAn
∩
Bn∈
Ftn+1, Lévy’s extension of the Borel–Cantelli lemmas (seeWilliams(1991,Theorem 12.15)) lets us conclude that, almost surely, only finitely many of the eventsAn∩
Bnoccur. Thus only finitely many of the eventsAn occur, almost surely.Recalling Eq.(3)we see that, almost surely, there exists a random integern1
< ∞
such that for alln>
n1, we have eNs(n)=
bNs(n)for alls∈ [
tn,
tn+1]
. Recall also that the process(
bNs(n),
s∈ [
tn,
tn+1] )
is initiated with a single particle at time tn, and every particle in this process breeds at constant rateβ
e2(
√2β−2ε)
tn. Hence the population size|b
Ns(n)|
is a Yule process with constant breeding rateβ
e2(
√2β−2ε)
tn, fors∈ [
tn,
tn+1]
. Therefore|b
Nt(nn+)1|
has a geometric distribution with parameterpn
:=
exp− δβ
e2(
√2β−2ε)
tn.
If we defineqn
:=
l exp
e2
(
√2β−3ε)
tnm,
then the probability that the number of particles inbNt(nn+)1is smaller thanqnis P
( |b
Nt(nn+)1| ≤
qn) =
1− (
1−
pn)
qn=
pnqn+
o(
pnqn).
Using the Borel–Cantelli lemmas again, we see that there exists almost surely a random integern2
< ∞
such that|b
Nt(nn+)1| >
qnfor alln>
n2.Finally, for alln
>
max{
n1,
n2}
and allt∈ [
tn+1,
tn+2]
, we have that there are at least as many particles inNtas there are descendants ofRtnat timetn+1, which is to say that|
Nt| ≥ |e
Nt(nn+)1|
. Hence, fornsufficiently large,ln ln
|
Nt|
t
≥
ln ln|e
Nt(nn+)1|
tntn t
≥
2p2
β −
7ε,
(4)almost surely. (Certainly we must haven
>
max{
n1,
n2}
, but we may require thatnbe larger still in order that the multiplicative factortn/
tis close enough to 1 for Eq.(4)to hold.)We can take
ε
to be arbitrarily small, and so obtain lim inft→∞
ln ln
|
Nt|
t≥
2p2
β,
almost surely, which completes the proof.
Acknowledgement
The third author was supported by the Heilbronn Institute for Mathematical Research.
References
Hardy, R., Harris, S.C., 2009. A spine approach to branching diffusions with applications toLp-convergence of martingales. In: Séminaire de Probabilités XLII. In: Lecture Notes in Math., vol. 1979. Springer, Berlin, pp. 281–330.
Harris, J.W., Harris, S.C., 2009. Branching Brownian motion with an inhomogeneous breeding potential. Ann. Inst. H. Poincaré Probab. Statist. 45 (3), 793–801.
Itô, K., McKean, H.P., 1965. Diffusion Processes and their Sample Paths. In: Die Grundlehren der Mathematischen Wissenschaften, Band 125, Academic Press Inc., Publishers, New York.
Williams, D., 1991. Probability with Martingales. Cambridge University Press.