On the path to extinction

Peter Jagers*, Fima C. Klebaner, and Serik Sagitov*

*Department of Mathematical Sciences, Chalmers University of Technology, SE-412 96 Gothenburg, Sweden; and School of Mathematical Sciences, Monash University, Clayton, Victoria 3800, Australia Communicated by David Cox, University of Oxford, Oxford, United Kingdom, January 4, 2007 (received for review September 15, 2006)

Populations can die out in many ways. We investigate one basic form of extinction, stable or intrinsic extinction, caused by individuals on the average not being able to replace themselves through reproduction. The archetypical such population is a subcritical branching process, i.e., a population of independent, asexually reproducing individuals, for which the expected number of progeny per individual is less than one. The main purpose is to uncover a fundamental pattern of nature. Mathematically, this emerges in large systems, in our case subcritical populations, starting from a large number, x, of individuals. First we describe the behavior of the time to extinction T: as x grows to innity, it behaves like the logarithm of x, divided by r, where r is the absolute value of the Malthusian parameter. We give a more precise description in terms of extreme value distributions. Then we study population size partway (or u-way) to extinction, i.e., at times uT, for 0 < u < 1, e.g., u 12 gives halfway to extinction. (Note that mathematically this is no stopping time.) If the population starts from x individuals, then for large x, the proper scaling for the population size at time uT is x into the power u 1. Normed by this factor, the population u-way to extinction approaches a process, which involves constants that are determined by life span and reproduction distributions, and a random variable that follows the classical Gumbel distribution in the continuous time case. In the Markov case, where an explicit representation can be deduced, we also nd a description of the behavior immediately before extinction.

eada 1.

Z t

ru 0 e 1 Ludu ru 0 ue du

ert, t 3

eneral branching processes are defined through individuals who give birth independently of each other. We limit ourselves to single-type such populations, all individuals having the same reproduction and, more generally life-path, distribution. Benchmark cases of such processes are the simple Galton Watson, birth-and-death, and Markov branching processes. Even though such Markovian structures dominate the probabilistic literature, more general processes (allowing aging, various distributions of life spans and fertility periods, as well as repeated litters of varying sizes) are relevant for biological modelling (1). A presentation from a mathematical point of view can be found in ref. 2 (multi-type populations) or ref. 3 (single-type processes). A multi-type framework would have been still more general but also far less accessible. The not-so-mathematical reader will undoubtedly note that even this presentation is demanding at certain points. However, conclusions should be clear, and they, rather than the work to derive them, form the message of this paper: intrinsic stable extinction follows a simple and beautiful pattern. The purpose is thus general understanding, rather than paving the way for specific applications. For those, multi-type theory may be useful, but at the bitter end they will tend to need custom-made models of narrow relevance. A (single-type) branching process is subcritical if the expected number m of children per individual is less than one. Let ( a ) denote the expected number of children by age a . Clearly, m ( ). In the lattice case, births can occur only at multiples of some time unit, e.g., yearly. For simplicity, our focus is on the nonlattice case. We consider so-called Malthusian processes. These are defined by the requirement that there exists a number , the Malthusian parameter, such that
www.pnas.orgcgidoi10.1073pnas.0610816104

(ref. 3, p. 156). We write C for the proportionality constant in this, i.e., [ Z t] Ce t Ce rt and recall that in the lattice case the enumerator integral is replaced by a sum. (To the nonmathematical reader, the importance of this is not the special formula for C , but rather the fact that the constant can be calculated from demographic facts.) For GaltonWatson and Markov branching processes C 1, and expressions are exact, [ Z t] m t for integer t in the GaltonWatson case. The expected population size thus has the same asymptotic form in all three cases, , , 0. For variances, though, asymptotics take different forms in the sub- and supercritical cases. Assume that m 1 and that Malthusianness is strengthened to a second order property: let denote the reproduction point process, so that ][, and write

eada;

it is required that [ ()2] . Then, Var Z t Be rt, for some constant B . This can be checked from the convolution equation holding for second moments (ref. 3, p. 136). In the more often studied supercritical case, the leading term is proportional to e 2t. We impose second order Malthusianness throughout and assume that the life span distribution L and reproduction measure satisfy
Author contributions: P.J. designed research; P.J., F.C.K., and S.S. performed research; P.J., F.C.K., and S.S. contributed new reagentsanalytic tools; and P.J., F.C.K., and S.S. wrote the paper. The authors declare no conict of interest. Freely available online through the PNAS open access option.
To

whom correspondence should be addressed. E-mail: jagers@chalmers.se.

2007 by The National Academy of Sciences of the USA

PNAS April 10, 2007 vol. 104 no. 15 6107 6111

POPULATION BIOLOGY

APPLIED MATHEMATICS

This is always fulfilled in the critical ( m 1) and supercritical (1 m ), and under very mild conditions (always met with in the real world) also in subcritical cases. Then, 0 and since will play an important role, we shall write r 0. In the case of GaltonWatson processes, r ln m . Let Z x t be a shorthand for the number of individuals alive at time t , provided the population started from x individuals at time 0, taken as newborn then, for preciseness. Z t without a suffix, is Z1 t , the case of one ancestor. If L denotes the life span distribution, any nonlattice Malthusian process will then satisfy

erttLdt and

e rtt dt ,

[1]

where suptv c t c for any 0, given that v is sufficiently large. Thus it remains to observe that as x 3
x1

respectively. As a consequence, the so called x log x -condition holds, guaranteeing the Kolmogorov and Yaglom theorems on survival chances and conditional population size given nonextinction (see ref. 4, p. 170, for the case of age-dependent processes and ref. 3, p. 159, and ref. 5, for general processes). In the nonlattice formulation, they tell us that Zt 0 cert, t 3 for some 0 c 1,
t3

k0

cert1 certkdt

1 r

x1

k0

1 1 1 cervk1 k1

1 ln x ln c v o 1 , r

[2]

lim Z t k Z t 0 b k , k 1, 2, 3, . . .

[3]

exist, k b k 1, and b

where o (1), denotes a remainder term, vanishing as x 3 . Extreme value theory handily delivers distributional forms of this, though again distinction should be made between continuous-time and lattice cases. In the former, Kolmogorovs theorem yields the exponential tail that implies a classical Gumbel limit (ref. 8, p. 17). Indeed, for each initial number x , define x by Tx ln x ln c x r . [7]

kbk Cc.

[4]

Then, as x 3 , x y 3 expe y , y . [8]

It is worth noting that these two theorems do not require the full strength of branching process independence assumptions (see ref. 6 for population-size-dependent branching). Even though we shall not embark upon the subtleties of measuring the population in other ways than by just counting its individuals, it turns out that the proof of our results is considerably facilitated by use of counting with a particular random characteristic (ref. 3, p. 167 ff.). The characteristic we shall use will be one that records all events to come in an individuals life and daughter process. It is thus not individual like the ones treated in op. cit. but still well behaved (7). The first and second moment results quoted above for processes just counting the number of individuals alive, clearly extend to processes counted with characteristics such that ert([t)] is directly Riemann integrable, and so do the Kolmogorov-Yaglom theorems (5). The Time to Extinction Now consider processes with x ancestors, Z x t. By 2 the time to extinction, Tx inf t 0; Z x t 0 satisfies Tx t 1 T1 tx 1 Zt 0x 1 1 ct ertx [5] for some c t tending to c , as t 3 . As a consequence, T x

This extends Pakess (9) result for Markov branching processes. In Between Dawn and Demise x Let Z uTx denote the population size at some time uT x, 0 u 1, between its inception at size Z x 0 x and extinction. We consider nonlattice processes, and to ease exposition, write t x rt , as t 3 ln x , so that T x t x r , by 7 and 8. Since [ Z x t ] xCe x , also x u1 [ Z u(txt)/r] Ce ut, x 3 , and it is natural to guess that x 1u provides the right norming and that
x 3 C 1ub ue u xu1ZuT x

in distribution, as the initial population size x 3 , at least for fixed 0 u 1. A first check of this conjecture can be made by simulation. Consider a binary splitting Markov branching process, i.e., birth-and-death process, with the probability p 0 0.75 for zero and p 2 0.25 for two offspring, and expected life span 1. For birth-and-death, as pointed out, C 1 and further b can be shown to equal p 0 (2 p 0 1), which is 1.5 in the present case. Fig. 1 displays the simulation results. The two upper panels x show that the process Z uTx, 0 u 1 displays much more of x variation than does Z t . The lower panels indicate that the amount of variation is just right to allow nontrivial normalisation. They also point at convergence towards a process with a rather constant mean and a variance increasing with u . The expectation and variance of the proposed limiting population size are C1ubueu C1ubuu 1, Var C 1ub ue u C 1ub u 2 2 u 1 2 u 1 , 0 u 1, in terms of the classical Gamma function. Thus, for b C close to one (it cannot be smaller), the Gamma function exerts a strong influence on the behavior of expected size, and since (1) (2) 1 the overall picture is that of a fairly constant mean. For large b C it increases practically exponentially. This may seem intriguing, but is explained by the norming through x 1u. The limiting variance always increases with u , from 0 to b 2. Fig. 2 shows these expectations and standard deviations for different values of b assuming C 1. The situation for b 1.5 thus seems to fit simulations.
Jagers et al.

Tx tdt ln x ln c x r ,

[6]

where x 3 Eulers . To verify 6 note that for any fixed number v , 0 v

Tx tdt 3 v, x 3 .

On the other hand,

Tx tdt

x1

ct ert

k0

1 ct ertkdt ,

6108 www.pnas.orgcgidoi10.1073pnas.0610816104

1200 1000 800 600 400 200 0 6 5 4 3 2 1 0 0 0.2 0.4 0.6 0.8 1 (xu1ZuT , 0<u<1) 0 5 10 15 20 (Z , 0<t<T)
t

1200 1000 800 600 400 200 0 6 5 4 3 2 1 0 0 0.2 0.4 0.6 0.8 1
APPLIED MATHEMATICS POPULATION BIOLOGY

(ZuT , 0<u<1)

0.2

0.4

0.6

0.8

(xu1ZuT , 0<u<1)

Fig. 1. Markov branching with p0 0.75, p2 0.25, life expectancy 1. (Upper) Forty simulations with initial number x 1,000, population size on the vertical axis, time (Left) and proportion of time u to extinction (Right) horizontally. (Lower) Forty simulations of normed size with x 1,000 (Left) and x 10,000 (Right), proportion u of time to extinction on the horizontal axis.

The Path-to-Extinction Theorem and Its Proof Our proof of the convergence follows a three-step programme, which requires acquaintance with weak convergence theory (10). The reader interested in the basic pattern of extinction, rather than mathematical technique is thus advised to jump to Theorem 1. 1. For fixed 0 u 1, write x( t ) x u1 Z u(txt)/r and prove that
x

xt 3 Ceut, t ,
as x 3 , the arrow indicating weak convergence in the LindvallSkorohod topology extended to the whole real line. 2. Consider x at the random argument x : ln c x 3 : ln (C b ) , where is Gumbel by 7 and 8, and check conditions for x( x) 3 Ce u C 1u b u e u in distribution for fixed u .

b 1.5 1.4 1.2 1 0.8 0.6 0.4 0.2 0.2 0.4 0.6 0.8 1
0 0 0.2 0.4 0.6 0.8 1 2 4 6 b=1.5

b 1.1 5 1 0.8 0.6 0.4 0.2 0.2 0.4 0.6 0.8 1 4 3 2 1 0.2 0.4

b 5

0.6

0.8

Fig. 2. panel shows 40 simulations of (1.5)ue u, to be compared with the 40 simulations in Fig. 1.

Mean (upper line) and standard deviation (lower line) of the limit process bue u for b

1.1, 1.5, 5, 0 u 1 on the horizontal axis. The upper right

Jagers et al.

PNAS April 10, 2007 vol. 104 no. 15 6109

3. Study the asymptotics of x( x)( u ) x u1 Z uTx as a function of u varying inside the unit interval.
Step 1. We already noted that
x u u t x t Ceut, x 3 . xt xu1Zu txt/r x Ce

If we assume that reproduction is L 2-Lipschitz in the sense that the second moment of the number of births between ages t and t h , v h( t ), satisfies

vhtertdt Oh, h2 0,

[9]

Similarly,
x 2u1 Var x t Var x u1Z u Bx 1ue ut 3 0. t x t / r x

then a similar, but prohibitive, analysis of the convolution equation for second moments, shows that also
1 t2 ertOh. Dh

We can conclude that for any fixed t , x( t ) 3 Ce ut in mean square. The finite dimensional convergence xt1, xt2, . . . , xtk 3 Ceut1, Ceut2, . . . , Ceutk is obvious. Turning to tightness, we note that for any v and a 0 sup x t a
tv

x u1

sup
t u t x v / r

Zx t a

x u1

This Lipschitz-property is broadly satisfied, e.g., if the number of children in short intervals is bounded. Indeed, if B is a bound, then v h( t ) B ( ( t h ) ( t ), so that such qualities of the reproduction function transfer. By Chebyshevs inequality, therefore, for a suitable K ,

sup
t u t x v / r

Zx t a.

sup
tsth

xs xt

sup

tsth

x x 1u Zu txs/r Zutxt/r x

But the total progeny Y of a population from one newborn ancestor clearly dominates the maximum of the numbers ever simultaneously present, and the sum of the maxima majorises the supremum of the sum. Since [ Y ] 1 (1 m ), it follows that

x Dh utx tr x1u x x Dh utx tr Dh utx tr x1u xxu Kh x VarDh utx tr 21u x Kh2

sup
tutxv/r

x 1u Zx Ceuv1 m. t Zutxv/r1 m] x

eutKxu1h for h little enough and K some constant. Tightness follows along the lines of the Corollaries in ref. 10 (p. 83 and p. 142). The process x( t ) thus has the weak nonrandom continuous limit Ce ut, t , as x 3 . The modification from there is that those corollaries concern processes on the unit interval, whereas our processes are defined on the whole line, with the natural LindvallSkorohod topology of convergence on all finite subintervals.
Step 2. In the nonlattice case, the pair ( x, x) is also weakly

We can conclude that sup x t a K a

tv

for some constant K . Now consider the population counted with a characteristic (ref. 3, p. 167) that for any h 0 gives to each individual present the value of
Y Y

an indicator of dying the next h r time units plus the total number of progeny of the individual born within the same next time span of length h r .

convergent and by the continuity of the exponential limit function and the Gumbel limit, d x ln C b O C 1 u b ue u, xu1ZuT x x x x x in distribution, as claimed (see ref. 10, p. 151, or ref. 11, Section 2.2).
Step 3. The preceding two steps can now be repeated for any

x ( t ). Clearly, for any t , Denote the process thus counted by D h

sup
tsth

x Zu t x s / r

x Zu t x t / r

x Dh utx

tr.

Write h r , and denote the total progeny within t time units stemming from one newborn individual by Y t. Then, t Lt Lt

Ytudu.

linear combination

x i x u i 1 Z u iT x

For nonlattice processes it follows that, as t 3 ,

1 t Dh

yielding the finite dimensional distributional convergence fd x xu1ZuT , 0 u 1 O C 1ub ue u, 0 u 1 , x where is standard Gumbel. We summarize the situation for nonlattice populations:
Jagers et al.

ert
rt 0 te d t

rt t rt Ytududt 0 e L t L t dt 0 e t

[10]

ertOh, h 2 0.
6110 www.pnas.orgcgidoi10.1073pnas.0610816104

Theorem 1. Consider subcritical Malthusian general, single-type,

nonlattice branching processes with finite reproduction variances. Assume conditions 1 and 9. Then the finite dimensional distributional convergence 10 holds. It remains to note the discontinuity at the endpoints of the unit x interval in that x u1 Z uTx equals 1 at u 0 and 0 at u 1. The former disappears in the Markov case, where age distribution does not matter and C 1. The latter mirrors the drastic fall of population size from u close to but 1 to u 1, at which we proceed to have a closer look. Markov Branching Processes on the Eve of Extinction Markov branching processes, i.e., splitting processes (children are only born at their mothers death) where individuals have exponentially distributed life spans, have been much used in mathematical biology, even though they are of limited biological relevance, since exponential life spans imply the absence of aging. Being nonlattice, Markov branching processes are covered by our general results on extinction time and path to extinction. However, they also allow a direct approach. The latter is mathematically interesting, using a conditioning argument at a nonstopping time. For the Markov case, it recovers Theorem 1 under slightly relaxed conditions, and adds new knowledge because it renders it possible to study the process immediately before extinction, viewing the discontinuity from the left, as u 3 1, through a magnifying glass, as it were (Theorem 2). Now, let Z x t be a Markov branching process in which particles have exponentially distributed lifetimes with parameter a . Then
rt Zx , t xe

processes. What is special in the branching case is the simple relation T y t T 1 t y, valid for any positive integer y and helpful in calculating an integral representation of the probability generating function s Z vTx . Analysis of the latter for the case v ( t ) ut , 0 u 1 yields Theorem 1 for the Markov branching case, as mentioned under somewhat less restrictive conditions: the finite reproduction moment can be replaced by the x log x condition
x

k log k p k ,

x1

x Z x t y , y 1, 2, . . . ,

exactly. Hence, as pointed out, C 1; also r a (1 m ), as any textbook would tell you. Further, consider a function v ( t ) t . A total probability argument yields
x Zv T x y

Theorem 2. Consider a subcritical Markov branching process Z t

starting from x individuals, having parameters a and { p k} satisfying the x log x condition, as above. Then, as x 3 ,
x ZT 3 Yu , x u

x Zv Tx y , T x dt

0 x Zv t

y, Tx dt

0 x x Zv t y T x dt Z vt y

in distribution for fixed u 0, and finite-dimensionally. The limiting process is Markov. It starts from Y 0 1, stays at any position y 1, 2, 3, . . . an exponentially distributed time, with parameter ay , and then jumps to a position z with probability zz p . y y yz1 In the long run, it increases exponentially: as u 3 , Yu Xeru in distribution. Here X is an exponentially distributed random variable, with mean b .
We thank a referee for taking readability seriously. This work has been supported by the Swedish and Australian Research Councils.
8. Leadbetter MR, Lindgren G, Rootze n H (1983) Extremes and Related Properties of Random Sequences and Processes (Springer, New York). 9. Pakes AG (1989) Adv Appl Prob 21:243269. 10. Billingsley P (1999) Convergence of Probability Measures (Wiley, New York), 2nd Ed. 11. Silvestrov DS (2004) Limit Theorems for Randomly Stopped Stochastic Processes (Springer, London). 12. Harris TE (1963) The Theory of Branching Processes (Springer, Berlin).

0 x Zv t y T y v t dt .

x In words, the last equation holds because, conditionally on Z v (t) y , the probability of the original population, initiated by x ancestors, dying out at dt is the same as the probability that a population of y individuals at v ( t ) dies out after time t v ( t ). Note that this argument applies quite generally to Markov

1. Haccou P, Jagers P, Vatutin V (2005) Branching Processes: Variation, Growth, and Extinction of Populations (Cambridge Univ Press, Cambridge, UK). 2. Jagers P (1989) Stoch Proc Appl 12:183212. 3. Jagers P (1975) Branching Processes with Biological Applications (Wiley, New York). 4. Athreya K, Ney P (1972) Branching Processes (Springer, Berlin). 5. Green PJ (1977) J Appl Prob 14:451463. 6. Gosselin F (2001) Ann Appl Prob 11:261284. 7. Jagers P, Nerman O (1984) J Appl Prob 16:221-259.

Jagers et al.

PNAS April 10, 2007 vol. 104 no. 15 6111

POPULATION BIOLOGY

the stationary measure of the process (ref. 12, p. 24). Then, the following can be proved to hold:

APPLIED MATHEMATICS

where p k is the probability of splitting into k children. In the present context, it is more interesting that the case v ( t ) t u , 0 u t , yields distributional convergence of x Z Tx u , as x 3 , the limiting process Y u, thus displaying the properties of extinct populations on the eve of their disappearance. Indeed, write y, y 1, 2, . . . for the unique solution (up to multiplicative factors) of the system