Asymptotic Behaviour of Continuous Time, Continuous State-Space Branching Processes

Author(s): D. R. Grey
Source: Journal of Applied Probability , Dec., 1974, Vol. 11, No. 4 (Dec., 1974), pp. 669-
677
Published by: Applied Probability Trust

Stable URL: https://www.jstor.org/stable/3212550

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Applied Probability Trust is collaborating with JSTOR to digitize, preserve and extend access to
Journal of Applied Probability

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 J. Appl. Prob. 11, 669-677 (1974)
Printed in Israel

? Applied Probability Trust 1974

ASYMPTOTIC BEHAVIOUR OF CONTINUOUS TIME,

CONTINUOUS STATE-SPACE BRANCHING PROCESSES

D. R. GREY, University of Sheffield

Abstract

Results on the behaviour of Markov branching processes as time goes to

infinity, hitherto obtained for models which assume a discrete state-space or
discrete time or both, are here generalised to a model with both state-space
and time continuous. The results are similar but the methods not always so.
BRANCHING PROCESSES; CONTINUOUS STATE-SPACE; CONTINUOUS TIME; LIMIT LAW;
LAPLACE TRANSFORM; FUNCTIONAL ITERATION

1. Introduction

Jifina [4] introduced the extension of the branching property to Markov

processes with state-space the non-negative reals, in both discrete and continuous
time. Lamperti characterised the continuous-time processes as weak limits of
sequences of scaled Galton-Watson processes in [5], and as processes obtained
by a random time-change on processes with stationary, independent increments
in [6]. Watanabe ([11] and [12]) considers more general continuous state-spaces.
The object of this paper is to use the equation describing the evolution of the
continuous-time process on the non-negative reals to study its asymptotic behaviour.
The results generalise those of Seneta [7] for Galton-Watson processes. It will
be shown that, subject only to finiteness of mean, a supercritical process may al-
ways be scaled so as to converge almost surely to a non-degenerate random variable;
in addition, this is also true of a large class of subcritical processes. Necessary
and sufficient conditions are given for the scaling constants to be asymptotically
exponential. Hitherto, it seems that only convergence in law in discrete time has
been studied: by Seneta and Vere-Jones in [9].
For the general background, the reader is referred to Athreya and Ney [1].

2. Preliminaries

A Markov process {Z,; t > 0} with state space [0, oo) and stationary transition
probabilities is called a (continuous state) branching process if it satisfies

Received in revised form 19 November 1973.

669

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 670 D. R. GREY

(1) E(e-Z',IZo
for all 0, x, y ?
nitely divisible

distribution
lationship of Zt (0,with
(1) together t _ the
0), Chapman-Kolmogorov
and the process is started at Zo
equations = 1,that
imply then the re-

t(,,(0)) = t+I,,(O) for O, t, u 0,

i.e., the family {('; t ! O} forms a semi-group under the operation of functional
composition. If we assume the regularity condition that

lim (,f(0) - O)/u = F(O) exists for 0 ? 0

ulO0

then for any t > 0,

(,t+.(0) - f,(0))/u =F(t,(O))

(0,(f,(0))as- u,(0))/u
, 0O.
Hence for any fixed 0, 0,(0) is the solution of the differential equation

(2) oT,(O) = F(f,(O))

with initial condition 0o(0) = 0 which has solution, uni

(3) F()
fo'(O) d _
Silverstein [10] prove
finite time (i.e., expl
then a semi-group {(,}

(4) F(O) = eO2

where e, P > 0 amd a
and
Since
f(o.)x2p(dx) < o

P[Zt < o0] = lim Ee-OZ'

010o = lim e-''(O)
0o10

it follows that the process is con

0 , 0 for all t, and by expressing (3) as

dI d t
e F() ,() F(r)
for sufficiently small 6 > 0, we see that this is true iff f[F(6)]-Yd( diverge
as 040. For this condition it is necessary that e = 0 and sufficient that e =

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
Asymptotic behaviour of continuous time, continuous state-space branching processes 671

and F'(0) < oo, or equivalently that e = 0 and ftflo)xjY(dx) < 0o. We shall
henceforth only consider conservative processes.
It will be seen that F is the exponent function of an infinitely divisible distri-
bution whose Levy measure has no mass on (- co, 0). Its interpretation in terms
of Lamperti's corresponding process with stationary independent increments in
[6] is obvious.

3. Behaviour of F and extinction probability

F will readily be seen from (4) to satisfy F(O) = 0 and have derivatives of all
orders for 0 > 0. It is also concave since F" < 0. The overall shape of F will be

known once its behaviour as 0 -+ 0 and 0 --+ is known. But

F'(O) = a + f(o,o)x(1 -e-x)p(dx)

which may be positive, negative or + cox. At infinity, let us study the behaviour
of F(0)/0 and F()/02 , and write

v = x(1 - e-')lp(dx) oo,

K = f xe-xp(dx) ? 00 .

From (4) we have that

F)= 0 - 0i + ( - e-Cx- Oxxlp(dx) + x(1 - e-x)p(dx)

0 (o, 1) Ox 0.1)
+ I- Cx )p(dx) - xe-x
f 1.0) 0 [1,00)
Of these terms, the fifth is dominated
to zero; and the third tends to - f(o0,)xPl(d

happens to be finite; and clearly it ten

we conclude, collecting terms, that F(0)/
otherwise, F(0)/8 -+ - oo as 0 -, co. In th
tion by 0, we have

F(O) 1 - e- x - Ox )
02 + f(.1)( 02 2 x2u(dx)
which tends to -Pfl as 0--+ oo, by bounded convergence. So we have the fol-
lowing summary of the behaviour of F.
Lemma A.

(i) F'(0) = c~ + v co, and so F(8)> O for small 0 > 0 iff + v > 0.
(ii) F(8)> O for all 8 > 0iff f = 0 and - K ?0.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 672 D. R. GREY

(iii) If K < 00 t
(iv) If K = oo a
(v) If # > 0 th
We now turn to
of absorption at

q = lim P[Z
t-#o0 t-4o0

But

P[Z, = 0] = lim Ee-ez'

0-- o0

- lim e-O )
= e-M) say.
Therefore

qt = e-

Now by letting 0 -, 00 in (3) and using Lemma A, we conclude that

for t > 0 iff F(O) < 0 and f$' [-F(?)] -d? < co for large enough 0,
is a sufficient, and # > 0 or K = co a necessary, condition for this t
gives the first part of the following theorem.

Theorem 1. For a continuous state-space branching process, q, > 0

(and then for all) t > 0 iff F(O) <0 and f' [ - F(?)] -ld < co for large

In this case,
F(O)=0, and qso= qlimt- q, = a+v>0.
< 1 iff e-= O where 0o is the largest root of the equation
The second statement is proved as follows.
Equation (3) gives us when 0 - oo that

rd=
and so as t -+ oo, (- log qt,) decreases to the point where the integral first becomes
infinite, i.e., 00. So -log q = 00. By part (i) of Lemma A,

q<1 iff 0o>O iff c+v>0.

If we differentiate (2) partially w.r.t. 0 and set 0 = 0, we find that

M, = EZt = e" ,
where m = F'(0) = x + v, and so, if we call a process subcritical, critical or super-
critical according as m < 0, m = 0 or m > 0, then Theorem 1 has the following
immediate corollary.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 Asymptotic behaviour of continuous time, continuous state-space branching processes 673

Corollary 1.1.
(i) All processes with P = 0, except those with F(O) < 0 and f0[ - F(_)]' d?
<oo for large 0, have q =0.
(ii) All other supercritical processes have 0 < q < 1.
(iii) All other critical and subcritical processes have q = 1.
The existence of subcritical processes with no probability of extinction provides
the main divergence from the theory of discrete state-space branching processes.

4. Almost sure convergence

Seneta [7] has shown that it is always possible to scale a supercritical Galton-
Watson process so that it converges with probability one to a non-degenerate
random variable, provided its mean is finite; for infinite mean, this is no longer
true (see also Seneta [8]). He has also given a necessary and sufficient condition
for the scaling constants to be asymptotically equivalent to the means of the
process, i.e., asymptotically geometric. We shall generalise these results to con-
tinuous time and state-space.
We consider two classes of process.
Class I: supercritical processes with n < co .
Class II: critical and subcritical processes with f = 0 and K < O .
Supercritical processes with ni = co, and critical and subcritical processes
with f > 0 or K = co, would not seem to lend themselves to suitable scaling, even
in the latter case when there is no probability of extinction.
We start by observing that for each t > 0, ', as a function of 0 is strictly
increasing from 0 to (-log q), and so has an inverse function It, defined on
[0, -logqt), which is strictly increasing from 0 to oo; moreover, from (3), t,
satisfies

(5) = t for
f.) GF(?) 0 A'< -logq,.
AlsoItobviously rt(u(A)) = I,,+u(A) for t, u > 0, O 2< -log q+,,.
turns out that {q,(A)Z,} for some fixed 2 > 0 is correctly scaled for almost
sure convergence, provided that in Class I, in case f3 > 0 or - K < 0, A is chosen
so that F(A)> 0, i.e., between 0 and 0o, the positive root of F(O) = 0. This is
proved in the following theorems.

Theorem 2. Let W, = I,(A)Z, for each t ? 0, some fixed A > 0 (with F(;.) > 0
in Class I). Then Y, = e-w' is a martingale and converges almost surely to a
random variable Y as t -+ xo; and EYo -t EYe for all 0 > 0.
Proof. The almost sure convergence is that of a non-negative, separable,
continuous time martingale (see, e.g., Doob [3]); and the convergence of moments
is by bounded convergence. It remains to prove that { Y,} is a martingale.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 674 D. R. GREY

For t? 0, u>0
E[Y,+, IY,; s t
(since clearl

= exp{- -,(q
(by branch

= exp{ - ,
= exp{ -r l
To ensure that th
in distribution,
done by studyin
Theorem 3.

(i) Let 0(0) = -logEYO for 0 > 0. Then 0 satisfies

6()) d _ 1
(6) - -log 0 , F(?) k
where k = a + v = m for Class I and k = a - K for Class II.

(ii) P[Y= 0] = 0 and P[Y = 1] = q, and so W, -+ Wa.s. where Wis a finite

random variable with P [W = 0] = q.

Proof. Let b,(0) = -logEYI for t > 0, 0 > 0. Then

,(O) = logE exp{ - t,(A)Z,0}

Therefore by Theorem 2,

(0) = lim /,[0r?,(1)].

t- a00

But using (3) and (5) we have that

d + reG) d -
( F(?) f F( ) F()
fen,(> dP
St + f -t,

((7) ) F()"

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
Asymptotic behaviour of continuous time, continuous state-space branching processes 675

But as t -+ oo, we have by (5) that

O for Class I

00 for Class II

and so since F(?) ~ k as

{0 for Class I

oo for Class II
we have that as t- oo,

F log O .
o0es(A) d _1 f01.,(A) d 1log
fq,(A) F( ) k a,, (, x k
Thus, letting t -- oo in (7), we obtain (6).
(ii) P [Y = 0] = 1 -lim EY = 1 - lim e -(0)
0olo0 01o

and

P [Y= 1] = lim EYe =

Now using (6), we have that, fo

(which may be + oo) as 0 -0 oo; w
0(0) -. 0 as 0 1, 0 and 0(0) - 0oo a
It remains to determine condit
are asymptotically exponential f
stated separately for Class I and
as that in the latter case follows similar lines.

Theorem 4a. For Class I,

-, e-mu

as t - oo, for u > 0, and so if r4,(A)--1 cekt as t -+ oo for some constants

c > 0 and k, then necessarily k = m. For this to happen it is necessary and suf-

ficient that f,) x logx p(dx)< co.

Theorem 4b. For Class II,

rht +,,(A))rl( -+e(- e-"-""

as t - oo, for u > 0, and so if r(A)- ~ cek' as t 0 oo for some constants

c > 0 and k, then necessarily k = - Kc. For this to happen it is necessary and
sufficient that f<0 1)x(-logx)p(dx) < 00.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
 676 D. R. GREY

Proof of Theore

log +() m = -mu

as t - co, by (5). Hence the resul

For the second part, again using

1 d( m)
= t - --log
m

1 ,log[.
= -log e(e"

Now the integrand is non-negative since F() ? m? by concav

,(A) 0 as t -+ o, t,(A)em" increases with t and remains bounded if

F(1) -m ;?Z
which, since F(?) ~ m?, is equivalent to

But from the definition of F in (4), dom () < oo.

jmo - F(G) x= IP+f (dx) f (X - 1 + e-X) d

xy' d fAX(0 - 1 +e-0o
= 2 d IA+ x f(dx) 02 dO
(O.00)

where the inner integral ~ log

It is interesting to note that

fied, then the norming constan

in Class II, except in the trivial

from the means.

5. A note on moments and tail behaviour

It is easy to show for Class I, using Equations (3), (4), (5) and (6), that the
following are all equivalent.

(i) F() m de < 00 .

(ii) EZtlogZ, < co for all t> 0.
(iii) EW < oo.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms
Asymptotic behaviour of continuous time, continuous state-space branching processes 677

(iv) rt,(A) ~ const. e-m' as t -* oo.

(v) fr Io,) x log xp(dx) < oo.
Many results of the type (ii) . (iii) relating tail behaviour of the distributions
of Z, and W have been proved for supercritical Galton-Watson processes by
Bingham and Doney [2]; these results apply equally well to our case as the proofs
use the functional equation analogous to our

4(em'O)= =,,(0(0))
which is obtained from (3) and (6).

However, perhaps more fundamental are theorems of the type (ii) . (v) re-
lating tail behaviour of the distribution of Z, with that of the measure p, and
this topic will be studied elsewhere.

Acknowledgements

Thanks are expressed to Professors J. F. C. Kingman and J. A. Williamson

for help during the course of this work.

References

[1] ATHREYA, K. B. AND NEY, P. E. (1972) Branching Processes. Springer-Verlag, Berlin.

[2] BINGHAM, N. H. AND DONEY, R. A. (1974) Asymptotic properties of supercritical branch-
ing processes, I: Galton-Watson processes. Adv. Appl. Prob. 6, 711-731.
[3] DooBn, J. L. (1953) Stochastic Processes. Wiley, New York.
[4] JIRINA, M. (1958) Stochastic branching processes with continuous state space. Czech.
Math. J. 8, 292-313.
[5] LAMPERTI, J. (1967) The limit of a sequence of branching processes. Z. Wahrscheinlich-
keitsth. 7, 271-288.
[6] LAMPERTI, J. (1967) Continuous state branching processes. Bull. Amer. Math. Soc.
73, 382-386.
[7] SENETA, E. (1969) Functional equations and the Galton-Watson process. Adv. Appl.
Prob. 1, 1-42.
[8], SENETA, E. (1973) The simple branching process with infinite mean, I. J. Appl. Prob. 10,
206-212.
[9] SENETA, E. AND VERE-JONES, D. (1969) On a problem of M. Jifina concerning continuous
state branching processes. Czech. Math. J. 19, 277-283.
[10] SILVERSTEIN, M. L. (1967-8). A new approach to local times. J. Math. Mech. 17, 1023-
1054.

[11] WATANABE, S. (1968) A limit theorem of branching processes and continuous state
branching processes. J. Math. Kyoto Univ. 8, 141-167.
[12] WATANABE, S. (1969) On two dimensional Markov processes with branching property.
Trans. Amer. Math. Soc. 136, 447-466.

This content downloaded from

134.157.146.115 on Wed, 10 Aug 2022 01:19:10 UTC
All use subject to https://about.jstor.org/terms