r
X
i
v
:
p
h
y
s
i
c
s
/
0
3
0
5
0
2
2
v
1
[
p
h
y
s
i
c
s
.
b
i
o

p
h
]
7
M
a
y
2
0
0
3
Analytic solution of Hubbell’s
Model of Local Community Dynamics
Alan J. McKane
1
, David Alonso
2,3
and Ricard V. Sol´e
2,4
1
Department of Theoretical Physics, University of Manchester, Manchester M13 9PL, UK
2
Complex Systems Lab, Universitat Pompeu Fabra, Dr Aiguader 80, 08003 Barcelona, Spain
3
Department of Ecology, Facultat de Biologia, Universitat de Barcelona,
Diagonal 645, 08045 Barcelona, Spain
4
Santa Fe Institute, 1399 Hyde Park Road, New Mexico 87501, USA
Abstract
Recent theoretical approaches to community structure and dynamics reveal that
many largescale features of community structure (such as speciesrank distributions
and speciesarea relations) can be explained by a socalled neutral model. Using
this approach, species are taken to be equivalent and trophic relations are not taken
into account explicitly. Here we provide a general analytic solution to the local
community model of Hubbell’s neutral theory of biodiversity by recasting it as an urn
model i.e. a Markovian description of states and their transitions. Both stationary
and timedependent distributions are analysed. The stationary distribution — also
called the zerosum multinomial — is given in closed form. An approximate form
for the timedependence is obtained by using an expansion of the master equation.
The temporal evolution of the approximate distribution is shown to be a good
representation for the true temporal evolution for a large range of parameter values.
Keywords: Community dynamics, Hubbell’s neutral theory, abundance distribution,
zerosum multinomial, analytic solution.
Submitted to Theoretical Population Biology
1
1 Introduction
Understanding the global patterns of biodiversity and its dynamics at diﬀerent time scales
remains a great challenge for ecological science (Rosenzweig, 1995; Wilson, 2003). One
of the key features that deﬁnes community structure is the relation between range and
abundance. How the community structure develops in time and how species are spatially
distributed largely deﬁne the ﬁeld of macroecology (Brown, 1995). In this context, an
important step to unify biogeography and biodiversity has been achieved by Hubbell
(Hubbell, 2001; Bell, 2001) through the formulation of a neutral theory.
The mathematical framework employed by Hubbell allows for speciation processes to
be integrated with the MacArthurWilson theory of island biogeography. In this way, the
neutral theory predicts some universal features that can be tested by direct analysis of
speciesabundance distributions and other largescale measurements. In Hubbell’s theory,
two key quantities largely determine the steadystate distributions of species richness
(as well as relative species abundances on local and large geographic scales). These
two parameters are the socalled biodiversity number and the immigration (dispersal)
rate. Under Hubbell’s assumptions, the ecological properties of every individual in the
population are assumed to be identical.
In a neutral model of this type, individuals compete for the same pool of resources,
but chance events are responsible for the identity of the ﬁnal winner(s). The dynamics of
each species is thus pathdependent and a Markovian description of their time evolution
is appropriate. Under the assumption of a balance between birth, death and immigration
rates, the neutral theory is able to reproduce the quantitative patterns of species distri
butions that are well known from the ecological literature. It also permits the generation
of several nontrivial and testable quantitative predictions about biodiversity and biogeog
raphy. In particular, the theory predicts that rare species would typically be recent in
terms of their origination. In relation to conservation biology, a neutral community in
which species are essentially equal would be very ﬂuid, with frequent replacements. If
true, protected areas should be larger than those expected for stable communities with
species closely adapted to given niches.
Formally, Hubbell’s theory is the ecological analog to the neutral theory of genetic drift
in genetics (Kimura, 1983; Ewens, 1972; Karlin and McGregor, 1972). Early attempts
to incorporate the neutral approach from population genetics (Caswell, 1976; Hubbell,
1979) mainly highlighted the relevance of drift in community dynamics, providing evi
dence for a global view of ecosystems in which competitive forces, ecological niches, and
even trophic interactions could be ignored in the pursuit of a better understanding of
biodiversity dynamics. More recent work incorporated these ideas in an explicit way
(Hubbell, 1997; Sol´e and Alonso, 1998) and Hubbell’s recent book provides an extensive,
unifying account of these (Hubbell, 2001). The starting point of neutral models is a ran
dom community that evolves towards an equilibrium state under a stochastic birthand
death process incorporating dispersal. At high immigration rates, Hubbell’s theory pre
dicts a logseries distribution for the abundance of species in the local community, while
when the immigration coupling between the metacommunity and the local community is
lower, a lognormallike shape is obtained for this distribution. Within Hubbell’s approx
imation, these distributions are shown to be particular cases of what he denotes as the
zerosum multinomial (Hubbell, 2001).
Hubbell’s model for local communities is similar to that proposed in Sol´e et al., (2000)
and analysed in McKane et al, 2000. There we took advantage of a mean ﬁeld argument
to ﬁnd an analytical form for the stationary distribution for the probability of ﬁnding
species having an abundance of n individuals. In addition, we studied in detail its time
behaviour using diﬀerent approximations. Furthermore, our simpliﬁed approach based
on this mean ﬁeld argument allowed us to recover the scaling relationship between the
2
fraction of links actualised and the number of species in a community — the socalled
C
∗
S relation, and gave conditions in which such a relation arose.
Within Hubbell’s mathematical framework the dynamical stochastic models were nu
merically solved and the equilibrium properties analysed. In this paper we present an
analytic, general solution of Hubbell’s model for the local community dynamics, that
provides the stationary speciesabundance distributions together with the time evolution
from the initial state towards the stationary distribution.
2 Formulation of the theory
Hubbell’s theory concerns populations on two scales: local communities and regional
metacommunities. To explain the model and derive the equations in the simplest possible
way, we will use the language of urn models (Feller, 1968; Johnson and Kotz, 1977). This
is a natural description when the stochastic dynamics in one time step only depends on the
state of the system at the beginning of the time step (in other words is a Markov process).
It also provides us with a concrete picture of the process which aids the derivation of the
governing equation for the model.
We begin by considering the model in a limit where the two levels of description are
uncoupled. This allows us to focus only on the local community. We assume that there
are N
i
individuals of species i in the local community, with the total number of individuals
of all species being J, that is, J =
r
i=1
N
i
where r is the total number of species. The
model is deﬁned by picking one individual at random from the local community, killing
it, and then replacing it by an individual also drawn from the local community. In terms
of the associated model this corresponds to having N
i
balls of colour i (i = 1, . . . , r) in
the urn. If we focus on one particular colour, j, the probability that the number of balls
will decrease from N
j
to N
j
−1 during one time step is
W (N
j
−1N
j
) =
N
j
J
((J −1) −(N
j
−1))
J −1
, (1)
since a ball of colour j must be discarded and one of any other colour replaced for such
a transition to occur. On the other hand, the probability that the number of balls will
increase from N
j
to N
j
+1 requires that a ball of any other colour but j must be discarded,
and one of colour j be replaced. Therefore
W (N
j
+ 1N
j
) =
(J −N
j
)
J
N
j
J −1
. (2)
The whole point of the model, however, is to couple local communities and regional
metacommunities. This is achieved by choosing a replacement ball from the urn only (1−
m) of the time. For the rest of the time it is chosen from outside the urn. The probability
of picking a ball of colour j from this external source is deﬁned to be P
j
, and corresponds
to assuming that the replacement individual comes from the regional metacommunity
where species i has a relative abundance of P
i
. The transition probabilities (1) and (2)
now read
W (N
j
−1N
j
) = (1 −m)
N
j
J
(J −N
j
)
J −1
+ m
N
j
J
(1 −P
j
) (3)
and
W (N
j
+ 1N
j
) = (1 −m)
(J −N
j
)
J
N
j
J −1
+ m
(J −N
j
)
J
P
j
. (4)
The change in the probability that there are N
j
balls in the urn from time t to the time
after one time step has elapsed consists of four contributions. Two of these correspond
3
to an increase in this probability (due to transitions from (N
j
− 1) and (N
j
+ 1) to N
j
)
and two to a decrease (due to transitions from N
j
to (N
j
+1) and (N
j
−1)). The balance
equation showing this change is:
∆P(N
j
, t) = W(N
j
N
j
−1)P(N
j
−1, t) + W(N
j
N
j
+ 1)P(N
j
+ 1, t)
− {W(N
j
+ 1N
j
) + W(N
j
−1N
j
)} P(N
j
, t) . (5)
Compared with the long time scales we are interested in — during which many transi
tions will take place — the step size is very small, and we may take the limit in which
∆P(N
j
, t) →dP(N
j
, t)/dt. The resulting equation is a master equation for the probabil
ity P(N
j
, t) (Van Kampen, 1981; Gardiner, 1985). Some care is needed with the boundary
conditions on this equation: clearly the cases N
j
= 0 and N
j
= J are special cases since
there can be no transitions which reduce N
j
in the former case or which increase N
j
in
the latter case. One possibility is to write two separate equations for these special cases.
However there is no need for this if we ﬁrst observe that some of these conditions are nat
ural consequences of the form of the transition probabilities. For example, the expressions
in (3) and (4) are both zero if N
j
= 0 and N
j
= J respectively. So as long as we agree
to impose the formal deﬁnitions W(0 − 1) = 0 and W(JJ + 1) = 0 the same master
equation may be used for all states. In addition, an initial condition needs to be imposed
to complete the speciﬁcation of the problem. Typically, the number of individuals in the
local community at t = 0 will be given: P(N
j
, 0) = δ
N
j
,N
j,0
.
The mathematical formulation of Hubbell’s theory described above can be directly
mapped on to another dynamical model of a multispecies community which we introduced
a few years ago (Sol´e et al., 2000; McKane et al., 2000; Sol´e et al., 2002). In this case
though, the nature of the interaction depends on the “score” between one species and
another, and a form of mean ﬁeld theory had to be used in order to describe the dynamics
by such a straightforward dynamics. In terms of the notation we have used above — N
denoting the number of individuals of a particular species and J denoting the total number
of individuals of all species — the transition probabilities of this model are (Sol´e et al.,
2000; McKane et al., 2000):
W(N + 1N) = C
∗
(1 −µ)
N
J
_
J −N
J −1
_
+
µ
S
J −N
J
, (6)
and
W(N −1N) = C
∗
(1 −µ)
N
J
_
J −N
J −1
_
+
µ
S
(S −1)
N
J
. (7)
Here µ is the fraction of the time that replacing of one species by another can happen
by chance, and not because the replacement individual belongs to a species which has a
positive score against the ﬁrst. It clearly maps into m. The other constants are S, the
number of species, and C
∗
, a parameter related to the degree of connectivity of the matrix
of scores between the species. The precise form of the mapping is C
∗
= 1 and P
j
= S
−1
.
Since we have analysed this model extensively (McKane et al., 2000) we may simply
deduce expressions for quantities of interest in the Hubbell theory by setting C
∗
= 1, S =
P
−1
j
and µ = m.
3 Stationary state
The most straightforward questions we can investigate concern the nature of the stationary
state of the theory. Let us begin by introducing the abbreviations
r
N
j
≡ W (N
j
−1N
j
) =
N
j
J
_
(1 −m)
(J −N
j
)
J −1
+ m(1 −P
j
)
_
(8)
4
and
g
N
j
≡ W (N
j
+ 1N
j
) =
(J −N
j
)
J
_
(1 −m)
N
j
J −1
+ mP
j
_
. (9)
The master equation now reads
dP(N
j
, t)
dt
= r
N
j
+1
P(N
j
+ 1, t) + g
N
j
−1
P(N
j
−1, t) −
_
r
N
j
+ g
N
j
_
P(N
j
, t) . (10)
The stationary probability distribution, P
s
(N
j
), is determined by setting dP(N
j
)/dt = 0.
This gives
r
N
j
+1
P
s
(N
j
+ 1) −g
N
j
P
s
(N
j
) = r
N
j
P
s
(N
j
) −g
N
j
−1
P
s
(N
j
−1) . (11)
This is true for all N
j
, which implies that r
N
j
P
s
(N
j
) −g
N
j
−1
P
s
(N
j
−1) = I, where I is a
constant. Applying the boundary condition at N
j
= 0, we ﬁnd that I = 0 and therefore
r
N
j
+1
P
s
(N
j
+ 1) = g
N
j
P
s
(N
j
) ; N
j
= 0, 1, ..., J . (12)
To solve this equation, let us ﬁrst assume that m = 0. Then the r
N
j
and g
N
j
given by (8)
and (9) are all nonzero and we can solve (12) by iteration to obtain
P
s
(N
j
) =
g
N
j
−1
g
N
j
−2
... g
0
r
N
j
r
N
j
−1
. . . r
1
P
s
(0) ; N
j
= 1, ..., J . (13)
The constant P
s
(0) can be determined from the normalisation condition
J
N
j
=0
P
s
(N
j
) = P
s
(0) +
N
j
>0
P
s
(N
j
) = 1 . (14)
To simplify the algebra let us introduce some new notation for various combinations of
parameters which naturally appear in the solution of the model. We write the transition
probabilities as
r
N
j
=
(1 −m)
J(J −1)
N
j
(N
∗
j
−N
j
) , (15)
and
g
N
j
=
(1 −m)
J(J −1)
(J −N
j
)(N
j
+ P
∗
j
) , (16)
where
P
∗
j
=
m(J −1)
(1 −m)
P
j
and N
∗
j
=
_
J −m
1 −m
_
−P
∗
j
. (17)
Substituting the expressions (15) and (16) into (13) gives an explicit representation for the
P
s
(N
j
) in terms of P
s
(0). An expression for P
s
(0) itself can be obtained by performing the
ﬁnite sum which appears in (14). This sum can be performed analytically using properties
of Jacobi polynomials (Abramowitz and Stegun, 1965). Alternatively, the mapping into
the model deﬁned by (6) and (7) can be used since the result for the P
s
is known in this
case (Sol´e et al., 2000; McKane et al., 2000). One ﬁnds (see McKane et al., 2000, for
details of the derivation):
P
s
(N
j
) =
_
J
N
j
_
β(N
j
+ P
∗
j
, N
∗
j
−N
j
)
β(P
∗
j
, N
∗
j
−J)
, (18)
5
where β(a, b) = Γ(a)Γ(b)/Γ(a + b) is the betafunction.
It is interesting to note that in the case m = 0, where the local community is decoupled
from the regional metacommunity, g
0
= 0, and so from (12), since r
1
= 0, it follows that
P
s
(1) = 0. In fact, since r
N
j
= 0 for 0 < N
j
< J, we see from (12) that P
s
(N
j
) = 0 for all
0 < N
j
< J. So with no interaction with the regional metacommunity, species j either
disappears or becomes the only species there is in the local community. Therefore some
degree of coupling is vital for biodiversity.
In Fig 1, we have computed the stationary distribution for diﬀerent parameter values
and sizes of the system. The relative species abundance distribution predicted to occur
in local communities — the zerosum multinomial— by the uniﬁed theory of Hubbell can
be readily computed even for high community sizes using the analytic formula (18).
4 Time dependence
Together with universal features displayed by the stationary patterns observed in mature
communities, some common features are also observed when looking at how diversity de
velops in time. When an empty ﬁeld starts to be colonized by immigrant species a new
community gets formed and a pattern of species replacement develops. The transition
from abandoned ﬁeld to mature forest is one of the best known examples of ecological
sucession and is common in many places after the abandonment of agricultural land. In
temperate climates, a mature forest is the end point of sucession, in spite of the difer
ent potential initial conditions. The path towards the steady speciesranks distribution
seems to be common to many diﬀerent ecosystems (Hubbell, 2001). Furthermore, natural
systems are continuously perturbed; any disturbance resumes the process of ecological
succession. It is thus natural to ask: what predictions about this process can be made in
the context of Hubbell’s neutral theory?
In the last section it was shown that a closed form expression could be obtained for
the probability of ﬁnding N
j
individuals of species j in the local community when the
systems has reached the stationary state. In addition to this, just mentioned, we also
wish to know how the community is assembled from a given starting point. This requires
us to solve for the timedependence of the model. It is not possible, in general, to carry
this out exactly, since the transition probabilities (15) and (16) are nonlinear functions of
N
j
. It is nevertheless possible to get a very good approximation to P(N
j
, t) by using the
fact that in cases of interest J will be large. The approach which we will use, due to Van
Kampen (1981), is rather technical and has been discussed elsewhere in some detail (Van
Kampen, 1981; McKane et al., 2000), but the basic idea is quite simple. Therefore, we
will avoid these complications, and quote relevant results using the correspondence with
the transition probabilities (6) and (7).
The key idea is to expand about the deterministic version of the theory. In the limit
where the number of individuals becomes inﬁnite, all stochasticity is lost, and the system is
completely described by a deterministic equation. This equation is not known a priori, but
if it can be established, an expansion in powers of J
−1
could perhaps be set up to calculate
corrections to the deterministic result which would be valid for large, but ﬁnite, J. Quite
generally we would expect a plot of P(N
j
, t) against N
j
for ﬁxed t to be approximately
Gaussian for large J. The motion of the peak of this distribution would move with
t according to the deterministic equation. Van Kampen’s large J expansion gives the
deterministic equation as the zeroth order (J →∞) result, with the next to leading order
result giving a Gaussian distribution peaked at this value. Higher order contributions give
corrections to this distribution, but they are usually so small for large J that they are of
very little interest. Since a Gaussian centred on a given value is completely determined
6
by its width, there are only two things to ﬁnd: (i) the deterministic equation, (ii) the
width of the distribution.
In practice one writes N
j
= Jφ
j
(t) + J
1/2
x
j
, where φ
j
(t) = lim
J→∞
(N
j
/J) is the
fraction of j species which are present in the local community at time t in the deterministic
limit. The variable
x
j
=
1
√
J
(N
j
−Jφ
j
(t))
characterises the ﬂuctuations away from the deterministic theory. We require φ
j
(t) and
x
2
j
(x
j
= 0). Using the correspondence between the two models we obtain (McKane
et al., 2000)
dφ
j
dτ
= m(P
j
−φ
j
) , (19)
where τ = t/J is a rescaled time. This equation is easily understood: if φ
j
is less than the
abundance of species j in the regional metacommunity, then it increases. If it is more,
then it decreases. The equation is easily solved to give
φ
j
(τ) = φ
j
(0)e
−mτ
+ P
j
(1 −e
−mτ
) . (20)
Initially we ask that x
j
(0) = 0, which means that φ
j
(0) = N
j
(0)/J = N
j,0
/J. Going back
to the t variable gives
φ
j
(t) =
N
j,0
J
e
−mt/J
+ P
j
(1 −e
−mt/J
) . (21)
In Hubbell (2001, Chapter 4), an alternative discretetime formulation of this local
community model is given. Obviously, both time discrete and time continuous formula
tions give rise to the same equations for the deterministic model counterpart (Hubbell,
2001, page 110). However, he does not address the stochastic timecontinuous formula
tion. Here we show that insight can be gained by ﬁnding approximate solutions to the
timedependent model.
The width of the distribution is given by
x
2
j
τ
=
1
m
P
j
(1 −P
j
)
_
1 −e
−2mτ
_
+ A
j
2 −m
m
(1 −2P
j
) e
−mτ
_
1 −e
−mτ
_
−2(1 −m)A
2
j
τe
−2mτ
, (22)
where A
j
= (N
j,0
/J) −P
j
. We have already commented that the probability distribution
is a Gaussian to the order we have been working. Speciﬁcally, in terms of the quantities
calculated above,
P(N
j
, t) =
1
_
2πJ x
2
j
τ
exp
_
−
(N
j
−J φ
j
(t))
2
2J x
2
j
τ
_
, (23)
where φ
j
(t) and x
2
j
τ
are given by equations (21) and (22) respectively.
In Fig. 2, we show the temporal evolution for P(N
j
, t) computed both using a Gaussian
approximation (Eq. (23)) and the numerical integration of the master equation. The good
agreement which is obtained is a reﬂection of the fact that community sizes J are taken
to be large enough so that further terms in the large Jexpansion are negligible. However,
if the ﬁnal stationary distribution does not have a Gaussian shape, more terms should
be included in the expansion so as to capture the true temporal behaviour of P(N
j
, t).
Notice that, while the approximation given by Eq. (23) is always represented as dotted
or punctuated curves, in some cases these are not visible because they match the exact
distribution so completely.
7
5 Conclusion
The main aim of this paper has been to show that aspects of Hubbell’s neutral model
of local community biodiversity dynamics can be solved for exactly, and even if this is
not possible, calculational schemes are available which provide very good approximations
to the solution. Speciﬁcally, we have shown that the stationary properties of the model,
which can be obtained from the zerosum multinomial, can all be found exactly. So, for
instance, the mean value and variance of the number of individuals of species j, can be
obtained from this probability distribution. The nature of the time evolution cannot be
determined in closed form, but a controlled approximation based on assuming that the
total number of individuals of all species, J, is large, is possible. This is an excellent
approximation in most cases of interest, and we would expect that the results that we
have obtained will be relevant in these situations. The applicability of our approximation
scheme was checked by carrying out the numerical integration of the master equation (Eq.
10). The results, displayed in Fig. 2, conﬁrm our expectations.
While the results which we have reported describe the essential aspects of the solution
of Hubbell’s model, there are many other interesting features which are also amenable
to analysis and for which deﬁnite, and wellcontrolled, results may be obtained. The
structure of the metacommunity and the form of the colonisation curve are examples.
These, and related questions, are presently under study, and we hope to report our results
in a future publication.
Acknowledgements
DA would like to thank the MACSIN research group at the UFMG, Belo Horizonte,
Brazil for providing constant support and a nice working environment. This work has
been supported by a grant CIRIT FI00524 (DA) from the Catalan Government and by
the Santa Fe Institute.
8
References
Abramowitz, M. and Stegun, I. A., 1965. Handbook of mathematical functions. Dover,
New York.
Bell, G., 2001. Neutral macroecology. Science 293, 24132418.
Brown, J.H., 1995. Macroecology. The University of Chicago Press, Chicago.
Caswell, H., 1976. Community structure: a neutral model analysis. Ecol. Monogr. 46,
327354.
Ewens, W. J., 1972. The sampling theory of selectively neutral alleles. Theor. Popul.
Biol. 3, 87112.
Karlin, S. and McGregor, J., 1972. Addendum to a paper of W. Ewens. Theor. Popul.
Biol. 3, 113116.
Kimura, M., 1983. The neutral theory of molecular evolution. Cambridge University
Press, Cambridge.
Feller, W., 1968. An introduction to probability theory and its applications. Volume 1,
Third edition. Wiley, New York.
Gardiner, C. W., 1985. Handbook of stochastic methods. Springer, Berlin. 2nd ed.
Hubbell, S. P., 1997. A uniﬁed theory of biogeography and relative species abundance and
its application to tropical rain forests and coral reefs. Coral Reefs 16 (Suppl.), S9–S21.
Hubbell, S. P., 2001. The uniﬁed theory of biogeography and biogeography. Princeton
University Press. Princeton, NJ.
Johnson, N. L. and Kotz, S., 1977. Urn models and their applications. Wiley, New York.
Sol´e, R. and Alonso, D., 1998. Random walks, fractals and the origins of rainforest
diversity. Adv. Complex Syst. 1, 203220.
McKane, A. J., Alonso, D. and Sol´e, R., 2000. A mean ﬁeld stochastic theory for species
rich assembled communities. Phys. Rev. E 62, 8466–8484.
Rosenzweig, M. L., 1995. Species diversity in space and time. Cambridge University
Press, Cambridge, UK.
Sol´e, R., Alonso, D. and McKane, A. J., 2000. Scaling in a network model of a multispecies
ecosystem. Physica A 286, 337–344.
Sol´e, R., Alonso, D. and McKane, A. J., 2002. Selforganized instability in complex
ecosystems. Phil. Trans. R. Soc. Lond. B 357, 667–681.
Van Kampen, N. G., 1981. Stochastic processes in physics and chemistry. Elsevier,
Amsterdam.
Wilson, E. O., 2003. The encyclopedia of life. Trends in Ecology and Evolution 18, 7780.
9
Figure captions
1. Zerosum multinomial distribution. The analytic formula (18) has been used to
compute the stationary distribution, P
s
(N
j
), for diﬀerent values of the abundance
of species j in the metacommunity, the total number of individuals J and the prob
ability of immigration from the metacommunity, m. We have dropped the subscript
j, which labels a particular species, in the ﬁgure.
2. Temporal evolution of the probability, P(N
j
, t), of having the jth species repre
sented by N
j
individuals. The temporal evolution has been computed using both
the Gaussian approximation and the straightforward numerical integration of the
exact master equation. In both cases, the initial number of individuals of the focus
species was 0.8 × J. The relative abundance of the focus species in the metacom
munity was P
j
= 0.1 also in both cases. We have dropped the subscript j, which
labels a particular species, in the ﬁgure.
10
0 10 20 30 40 50 60
Abundance of species j
10
5
10
4
10
3
10
2
10
1
10
0
P
s
(
N
)
0 10 20 30 40 50 60
10
9
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
P
s
(
N
)
J = 64
0 20000 40000 60000 80000 100000
Abundance of species j
10
40
10
32
10
24
10
16
10
8
10
0
0 25000 50000 75000 100000
10
8
10
6
10
4
10
2
10
0
J = 100000
m = 0.05
P = 0.1
m = 0.05
P = 0.1
P=0.999
0.99
0.9
P=0.001
0.01
0.1
0.8 0.2
0.3 0.7
0.6 0.4
0.01
0.1
0.2
0.3
0.6 0.7 0.8
m=0.999
0.9
0.99
0.4 0.5
0.5 0.2 0.1
0.05
0.02
0.0001
0.0005
0.002
0.01
0.0001
0.0005
0.002
0.01
0.0002
0.001
0.005
0.02
m=0.5
Figure 1
11
0 200 400 600 800 1000
Abundance of species j
10
4
10
3
10
2
10
1
P
(
N
,
t
)
Stationary distribution
P(N,t), Exact
P(N,t), Gaussian approx.
600 800 1000
0
0.02
0.04
0.06
0.08
P
(
N
,
t
)
J = 1000 m = 0.001
0 2000 4000 6000
Abundance of species j
10
6
10
5
10
4
10
3
10
2
4000 6000 8000 10000
0
0.01
0.02
J = 10000 m = 0.05
t=100
1000
3000
6000
20000
300000
800000
6000
1000
20000
40000
60000
100000
100000
200000
2000000
2500000
302500
520000
Figure 2
12
death and immigration rates. we studied in detail its time behaviour using diﬀerent approximations. (2000) e and analysed in McKane et al. These two parameters are the socalled biodiversity number and the immigration (dispersal) rate. Hubbell’s theory is the ecological analog to the neutral theory of genetic drift in genetics (Kimura. our simpliﬁed approach based on this mean ﬁeld argument allowed us to recover the scaling relationship between the 2 . providing evidence for a global view of ecosystems in which competitive forces. and even trophic interactions could be ignored in the pursuit of a better understanding of biodiversity dynamics. In a neutral model of this type. In this context. Hubbell’s model for local communities is similar to that proposed in Sol´ et al. individuals compete for the same pool of resources. a lognormallike shape is obtained for this distribution. while when the immigration coupling between the metacommunity and the local community is lower. Formally. 2000. Karlin and McGregor. protected areas should be larger than those expected for stable communities with species closely adapted to given niches. Ewens. The dynamics of each species is thus pathdependent and a Markovian description of their time evolution is appropriate. a neutral community in which species are essentially equal would be very ﬂuid. 2001) through the formulation of a neutral theory. 2001). 1972. with frequent replacements. but chance events are responsible for the identity of the ﬁnal winner(s). 1976. Under Hubbell’s assumptions. How the community structure develops in time and how species are spatially distributed largely deﬁne the ﬁeld of macroecology (Brown. Within Hubbell’s approximation. 1979) mainly highlighted the relevance of drift in community dynamics. the ecological properties of every individual in the population are assumed to be identical. 1997. In Hubbell’s theory. these distributions are shown to be particular cases of what he denotes as the zerosum multinomial (Hubbell. The starting point of neutral models is a random community that evolves towards an equilibrium state under a stochastic birthand death process incorporating dispersal. the neutral theory is able to reproduce the quantitative patterns of species distributions that are well known from the ecological literature. More recent work incorporated these ideas in an explicit way (Hubbell. the theory predicts that rare species would typically be recent in terms of their origination. 1995. In addition. two key quantities largely determine the steadystate distributions of species richness (as well as relative species abundances on local and large geographic scales). Under the assumption of a balance between birth. At high immigration rates.. The mathematical framework employed by Hubbell allows for speciation processes to be integrated with the MacArthurWilson theory of island biogeography.1 Introduction Understanding the global patterns of biodiversity and its dynamics at diﬀerent time scales remains a great challenge for ecological science (Rosenzweig. There we took advantage of a mean ﬁeld argument to ﬁnd an analytical form for the stationary distribution for the probability of ﬁnding species having an abundance of n individuals. e unifying account of these (Hubbell. Early attempts to incorporate the neutral approach from population genetics (Caswell. Hubbell’s theory predicts a logseries distribution for the abundance of species in the local community. the neutral theory predicts some universal features that can be tested by direct analysis of speciesabundance distributions and other largescale measurements. an important step to unify biogeography and biodiversity has been achieved by Hubbell (Hubbell. 2003). One of the key features that deﬁnes community structure is the relation between range and abundance. 2001. In particular. 1983. It also permits the generation of several nontrivial and testable quantitative predictions about biodiversity and biogeography. ecological niches. If true. In relation to conservation biology. 1972). 2001). Wilson. Hubbell. Bell. In this way. Furthermore. 1998) and Hubbell’s recent book provides an extensive. 1995). Sol´ and Alonso.
the probability that the number of balls will increase from Nj to Nj +1 requires that a ball of any other colour but j must be discarded. In terms of the associated model this corresponds to having Ni balls of colour i (i = 1. On the other hand. . (4) W (Nj + 1Nj ) = (1 − m) J J −1 J The change in the probability that there are Nj balls in the urn from time t to the time after one time step has elapsed consists of four contributions. Two of these correspond 3 .fraction of links actualised and the number of species in a community — the socalled C ∗ S relation. is to couple local communities and regional metacommunities. r) in the urn. and one of colour j be replaced. general solution of Hubbell’s model for the local community dynamics. however. It also provides us with a concrete picture of the process which aids the derivation of the governing equation for the model. To explain the model and derive the equations in the simplest possible way. 1968. and then replacing it by an individual also drawn from the local community. J = r Ni where r is the total number of species. The transition probabilities (1) and (2) now read Nj Nj (J − Nj ) +m (1 − Pj ) (3) W (Nj − 1Nj ) = (1 − m) J J −1 J and (J − Nj ) (J − Nj ) Nj +m Pj . we will use the language of urn models (Feller. J J −1 (1) since a ball of colour j must be discarded and one of any other colour replaced for such a transition to occur. The i=1 model is deﬁned by picking one individual at random from the local community. . This is a natural description when the stochastic dynamics in one time step only depends on the state of the system at the beginning of the time step (in other words is a Markov process). Therefore W (Nj + 1Nj ) = (J − Nj ) Nj . J J −1 (2) The whole point of the model. The probability of picking a ball of colour j from this external source is deﬁned to be Pj . This is achieved by choosing a replacement ball from the urn only (1 − m) of the time. killing it. and corresponds to assuming that the replacement individual comes from the regional metacommunity where species i has a relative abundance of Pi . We begin by considering the model in a limit where the two levels of description are uncoupled. We assume that there are Ni individuals of species i in the local community. 2 Formulation of the theory Hubbell’s theory concerns populations on two scales: local communities and regional metacommunities. j. If we focus on one particular colour. In this paper we present an analytic. . This allows us to focus only on the local community. Johnson and Kotz. that provides the stationary speciesabundance distributions together with the time evolution from the initial state towards the stationary distribution. Within Hubbell’s mathematical framework the dynamical stochastic models were numerically solved and the equilibrium properties analysed. 1977). the probability that the number of balls will decrease from Nj to Nj − 1 during one time step is W (Nj − 1Nj ) = Nj ((J − 1) − (Nj − 1)) . For the rest of the time it is chosen from outside the urn. and gave conditions in which such a relation arose. with the total number of individuals of all species being J. . that is.
2002). However there is no need for this if we ﬁrst observe that some of these conditions are natural consequences of the form of the transition probabilities..Nj. S J (6) N J −N µ N + (S − 1) . 2000) we may simply deduce expressions for quantities of interest in the Hubbell theory by setting C ∗ = 1. For example. Some care is needed with the boundary conditions on this equation: clearly the cases Nj = 0 and Nj = J are special cases since there can be no transitions which reduce Nj in the former case or which increase Nj in the latter case. 0) = δNj . and we may take the limit in which ∆P (Nj . 2000): W (N + 1N) = C ∗ (1 − µ) and W (N − 1N) = C ∗ (1 − µ) N J J −N J −1 + µJ −N . Typically.0 . The precise form of the mapping is C ∗ = 1 and Pj = S −1 . an initial condition needs to be imposed to complete the speciﬁcation of the problem. Sol´ et al. S = Pj−1 and µ = m. and C ∗ .. the nature of the interaction depends on the “score” between one species and another. The mathematical formulation of Hubbell’s theory described above can be directly mapped on to another dynamical model of a multispecies community which we introduced a few years ago (Sol´ et al. the number of species. and a form of mean ﬁeld theory had to be used in order to describe the dynamics by such a straightforward dynamics. t)/dt. the expressions in (3) and (4) are both zero if Nj = 0 and Nj = J respectively. the number of individuals in the local community at t = 0 will be given: P (Nj . So as long as we agree to impose the formal deﬁnitions W (0 − 1) = 0 and W (JJ + 1) = 0 the same master equation may be used for all states. t) (Van Kampen. Gardiner. 3 Stationary state The most straightforward questions we can investigate concern the nature of the stationary state of the theory. McKane et al. (7) J J −1 S J Here µ is the fraction of the time that replacing of one species by another can happen by chance. t) − {W (Nj + 1Nj ) + W (Nj − 1Nj )} P (Nj . McKane et al. The balance equation showing this change is: ∆P (Nj . 1981. t) + W (Nj Nj + 1)P (Nj + 1.. 2000. t) → dP (Nj . (5) Compared with the long time scales we are interested in — during which many transitions will take place — the step size is very small. 1985)... It clearly maps into m. In terms of the notation we have used above — N denoting the number of individuals of a particular species and J denoting the total number of individuals of all species — the transition probabilities of this model are (Sol´ et al. One possibility is to write two separate equations for these special cases. t) = W (Nj Nj − 1)P (Nj − 1. e 2000. In addition. The resulting equation is a master equation for the probability P (Nj .to an increase in this probability (due to transitions from (Nj − 1) and (Nj + 1) to Nj ) and two to a decrease (due to transitions from Nj to (Nj + 1) and (Nj − 1)). Let us begin by introducing the abbreviations rNj ≡ W (Nj − 1Nj ) = (J − Nj ) Nj (1 − m) + m(1 − Pj ) J J −1 4 (8) .. In this case e e though. Since we have analysed this model extensively (McKane et al. and not because the replacement individual belongs to a species which has a positive score against the ﬁrst. The other constants are S. 2000. a parameter related to the degree of connectivity of the matrix of scores between the species. t) .
J J −1 (9) dP (Nj . .. Alternatively. Ps (Nj ). (15) J(J − 1) g Nj = where Pj∗ = (1 − m) (J − Nj )(Nj + Pj∗ ) . 2000. An expression for Ps (0) itself can be obtained by performing the ﬁnite sum which appears in (14). dt (10) The stationary probability distribution. 1965). . We write the transition probabilities as (1 − m) r Nj = Nj (Nj∗ − Nj ) . 2000). is determined by setting dP (Nj )/dt = 0. t) − rNj + gNj P (Nj . r1 (13) The constant Ps (0) can be determined from the normalisation condition J Ps (Nj ) = Ps (0) + Nj =0 Nj >0 Ps (Nj ) = 1 .. t) + gNj −1 P (Nj − 1. Nj = 0. (12) To solve this equation. Applying the boundary condition at Nj = 0.. Nj∗ − Nj ) . rNj rNj −1 . Nj∗ − J) Nj 5 (18) m(J − 1) J −m Pj and Nj∗ = − Pj∗ . t) . the mapping into the model deﬁned by (6) and (7) can be used since the result for the Ps is known in this case (Sol´ et al. McKane et al. for e details of the derivation): Ps (Nj ) = J β(Nj + Pj∗ . (1 − m) 1−m (17) . where I is a constant. we ﬁnd that I = 0 and therefore rNj +1 Ps (Nj + 1) = gNj Ps (Nj ) . 1. let us ﬁrst assume that m = 0... which implies that rNj Ps (Nj ) − gNj −1 Ps (Nj − 1) = I. J . Nj = 1. t) = rNj +1 P (Nj + 1.. . . 2000. (11) This is true for all Nj .. (14) and To simplify the algebra let us introduce some new notation for various combinations of parameters which naturally appear in the solution of the model.. Then the rNj and gNj given by (8) and (9) are all nonzero and we can solve (12) by iteration to obtain Ps (Nj ) = gNj −1 gNj −2 . J . This gives rNj +1 Ps (Nj + 1) − gNj Ps (Nj ) = rNj Ps (Nj ) − gNj −1 Ps (Nj − 1) . g0 Ps (0) . β(Pj∗.. One ﬁnds (see McKane et al...and gNj ≡ W (Nj + 1Nj ) = The master equation now reads Nj (J − Nj ) (1 − m) + mPj . J(J − 1) (16) Substituting the expressions (15) and (16) into (13) gives an explicit representation for the Ps (Nj ) in terms of Ps (0). This sum can be performed analytically using properties of Jacobi polynomials (Abramowitz and Stegun.
In addition to this. natural systems are continuously perturbed. It is nevertheless possible to get a very good approximation to P (Nj . Since a Gaussian centred on a given value is completely determined 6 . This requires us to solve for the timedependence of the model. t) by using the fact that in cases of interest J will be large. In fact. but if it can be established. Van Kampen’s large J expansion gives the deterministic equation as the zeroth order (J → ∞) result. g0 = 0. Therefore. The path towards the steady speciesranks distribution seems to be common to many diﬀerent ecosystems (Hubbell. McKane et al. Higher order contributions give corrections to this distribution. just mentioned. It is interesting to note that in the case m = 0. to carry this out exactly. b) = Γ(a)Γ(b)/Γ(a + b) is the betafunction. The transition from abandoned ﬁeld to mature forest is one of the best known examples of ecological sucession and is common in many places after the abandonment of agricultural land. any disturbance resumes the process of ecological succession. where the local community is decoupled from the regional metacommunity. and so from (12). Quite generally we would expect a plot of P (Nj . since rNj = 0 for 0 < Nj < J. and the system is completely described by a deterministic equation. it follows that Ps (1) = 0. t) against Nj for ﬁxed t to be approximately Gaussian for large J. 2000). but the basic idea is quite simple. since r1 = 0. due to Van Kampen (1981). Furthermore. Therefore some degree of coupling is vital for biodiversity. In Fig 1. 4 Time dependence Together with universal features displayed by the stationary patterns observed in mature communities. we will avoid these complications. some common features are also observed when looking at how diversity develops in time. The motion of the peak of this distribution would move with t according to the deterministic equation. The key idea is to expand about the deterministic version of the theory. 1981. This equation is not known a priori. we have computed the stationary distribution for diﬀerent parameter values and sizes of the system. with the next to leading order result giving a Gaussian distribution peaked at this value. 2001). In the limit where the number of individuals becomes inﬁnite. The relative species abundance distribution predicted to occur in local communities — the zerosum multinomial— by the uniﬁed theory of Hubbell can be readily computed even for high community sizes using the analytic formula (18). and quote relevant results using the correspondence with the transition probabilities (6) and (7). a mature forest is the end point of sucession. J. we see from (12) that Ps (Nj ) = 0 for all 0 < Nj < J. When an empty ﬁeld starts to be colonized by immigrant species a new community gets formed and a pattern of species replacement develops. but they are usually so small for large J that they are of very little interest. It is thus natural to ask: what predictions about this process can be made in the context of Hubbell’s neutral theory? In the last section it was shown that a closed form expression could be obtained for the probability of ﬁnding Nj individuals of species j in the local community when the systems has reached the stationary state. It is not possible. but ﬁnite.. we also wish to know how the community is assembled from a given starting point. all stochasticity is lost. In temperate climates. So with no interaction with the regional metacommunity. The approach which we will use. species j either disappears or becomes the only species there is in the local community. in general. an expansion in powers of J −1 could perhaps be set up to calculate corrections to the deterministic result which would be valid for large. is rather technical and has been discussed elsewhere in some detail (Van Kampen.where β(a. in spite of the diferent potential initial conditions. since the transition probabilities (15) and (16) are nonlinear functions of Nj .
Notice that.by its width. If it is more. (19) dτ where τ = t/J is a rescaled time. Speciﬁcally. 2001. 2000) dφj = m (Pj − φj ) . This equation is easily understood: if φj is less than the abundance of species j in the regional metacommunity. we show the temporal evolution for P (Nj . where φj (t) = limJ→∞ (Nj /J) is the fraction of j species which are present in the local community at time t in the deterministic limit. The variable 1 xj = √ (Nj − Jφj (t)) J characterises the ﬂuctuations away from the deterministic theory. then it decreases. in some cases these are not visible because they match the exact distribution so completely. The good agreement which is obtained is a reﬂection of the fact that community sizes J are taken to be large enough so that further terms in the large Jexpansion are negligible. (23)) and the numerical integration of the master equation. + Aj j m (22) where Aj = (Nj. more terms should be included in the expansion so as to capture the true temporal behaviour of P (Nj . both time discrete and time continuous formulations give rise to the same equations for the deterministic model counterpart (Hubbell. (ii) the width of the distribution. Here we show that insight can be gained by ﬁnding approximate solutions to the timedependent model. However. We require φj (t) and x2 ( xj = 0). We have already commented that the probability distribution is a Gaussian to the order we have been working. 7 . J (21) In Hubbell (2001. 2. (23) is always represented as dotted or punctuated curves. an alternative discretetime formulation of this local community model is given. which means that φj (0) = Nj (0)/J = Nj. t) computed both using a Gaussian approximation (Eq. if the ﬁnal stationary distribution does not have a Gaussian shape. Obviously. there are only two things to ﬁnd: (i) the deterministic equation.0 −mt/J e + Pj (1 − e−mt/J ) . while the approximation given by Eq. (23) where φj (t) and x2 τ are given by equations (21) and (22) respectively. The width of the distribution is given by x2 j τ = 1 Pj (1 − Pj ) 1 − e−2mτ m 2−m (1 − 2Pj ) e−mτ 1 − e−mτ − 2(1 − m)A2 τ e−2mτ . (20) Initially we ask that xj (0) = 0. However. in terms of the quantities calculated above. P (Nj . he does not address the stochastic timecontinuous formulation. Using the correspondence between the two models we obtain (McKane j et al. Going back to the t variable gives φj (t) = Nj.0 /J) − Pj .. The equation is easily solved to give φj (τ ) = φj (0)e−mτ + Pj (1 − e−mτ ) . j In Fig. t). page 110).0 /J. Chapter 4). then it increases. In practice one writes Nj = Jφj (t) + J 1/2 xj . t) = 1 2πJ x2 j τ exp − (Nj − J φj (t))2 2J x2 τ j .
This is an excellent approximation in most cases of interest. 2. there are many other interesting features which are also amenable to analysis and for which deﬁnite. Brazil for providing constant support and a nice working environment. and even if this is not possible. is large. are presently under study. results may be obtained. and related questions. for instance. Acknowledgements DA would like to thank the MACSIN research group at the UFMG. can be obtained from this probability distribution. which can be obtained from the zerosum multinomial.5 Conclusion The main aim of this paper has been to show that aspects of Hubbell’s neutral model of local community biodiversity dynamics can be solved for exactly. we have shown that the stationary properties of the model. Belo Horizonte. the mean value and variance of the number of individuals of species j. and we hope to report our results in a future publication. displayed in Fig. The structure of the metacommunity and the form of the colonisation curve are examples. conﬁrm our expectations. is possible. J. and wellcontrolled. 10). The results. The nature of the time evolution cannot be determined in closed form. While the results which we have reported describe the essential aspects of the solution of Hubbell’s model. 8 . These. but a controlled approximation based on assuming that the total number of individuals of all species. This work has been supported by a grant CIRIT FI00524 (DA) from the Catalan Government and by the Santa Fe Institute. So. The applicability of our approximation scheme was checked by carrying out the numerical integration of the master equation (Eq. and we would expect that the results that we have obtained will be relevant in these situations. can all be found exactly. Speciﬁcally. calculational schemes are available which provide very good approximations to the solution.
Ewens. Wilson. A uniﬁed theory of biogeography and relative species abundance and its application to tropical rain forests and coral reefs. Brown. Caswell.. J. Ewens. Popul.).. Alonso. A. Neutral macroecology. Biol. D. 1965. R. 3. W. Alonso.. 1968. 3.. 327354. Dover. 2000. 1995. and McKane. Stochastic processes in physics and chemistry. 1972. 2000. M. and McKane. 667–681.. A.. Wiley. 46.. and Alonso. New York. J. D. W. W. Trans. 7780. R. 2003. 2002. N. Monogr. Johnson.. 203220. C. Elsevier. Phys.. 1977. S. Berlin. L. 24132418. A mean ﬁeld stochastic theory for speciese rich assembled communities. Handbook of stochastic methods. Princeton. NJ. and Kotz. 1997. H. 9 . Rev. Cambridge University Press. R. New York.. Coral Reefs 16 (Suppl. New York. Theor. Scaling in a network model of a multispecies e ecosystem. S. E 62. The University of Chicago Press. 113116. S9–S21. M.H. Ecol. and McGregor. Science 293. R. Karlin. Cambridge. Complex Syst. Community structure: a neutral model analysis. Urn models and their applications. Sol´. 2001. An introduction to probability theory and its applications. J. Sol´. O. The sampling theory of selectively neutral alleles. Bell. G. 1983. Gardiner.. fractals and the origins of rainforest e diversity. Princeton University Press. Volume 1. Cambridge University Press.. Springer.. Trends in Ecology and Evolution 18. P. Hubbell. Physica A 286. A. Biol. D. 1976. Hubbell. Popul.. P. 337–344. 1. Kimura. N. 1985. and Stegun. Rosenzweig. 1972. Species diversity in space and time. 8466–8484. M.. S. B 357. 2001... UK. Adv. Phil... R. 2nd ed. Sol´. L. J. McKane. Addendum to a paper of W. 1995. A. and Sol´.. I. S. 1981. Handbook of mathematical functions. The uniﬁed theory of biogeography and biogeography. 87112. J.. Lond. Third edition. The neutral theory of molecular evolution.References Abramowitz. E. J. Wiley. The encyclopedia of life. G.. Selforganized instability in complex e ecosystems. 1998. Alonso. Macroecology. Van Kampen. Soc. Random walks. Theor. Cambridge. Feller. D. Chicago. Amsterdam.
t). The temporal evolution has been computed using both the Gaussian approximation and the straightforward numerical integration of the exact master equation. 2. the total number of individuals J and the probability of immigration from the metacommunity.8 × J. In both cases.Figure captions 1. in the ﬁgure. We have dropped the subscript j. for diﬀerent values of the abundance of species j in the metacommunity. We have dropped the subscript j.1 also in both cases. which labels a particular species. the initial number of individuals of the focus species was 0. Zerosum multinomial distribution. Ps (Nj ). which labels a particular species. The relative abundance of the focus species in the metacommunity was Pj = 0. 10 . P (Nj . Temporal evolution of the probability. in the ﬁgure. of having the jth species represented by Nj individuals. The analytic formula (18) has been used to compute the stationary distribution. m.
1 0.001 10 0 10 20 30 40 50 60 8 0 10 0 25000 m=0.J = 64 10 10 10 10 0 J = 100000 10 0 1 m = 0.9 0.01 0.002 0.02 10 40 0 10 20 30 40 50 60 0 20000 40000 60000 80000 100000 Abundance of species j Abundance of species j Figure 1 11 .1 1 10 10 0.0005 8 0.2 0.7 0.7 0.3 0.99 Ps(N) 10 10 4 4 0.0001 P = 0.999 P=0.005 4 32 0.0005 0.8 0.5 0.1 0.05 0.99 0.1 0.05 10 10 10 2 2 3 0.6 0.2 0.2 0.0001 5 0.5 50000 75000 100000 10 10 0 P = 0.05 0.002 0.4 0.02 0.01 10 10 10 10 5 6 7 8 9 6 P=0.001 Ps(N) 10 2 16 10 10 3 10 10 10 24 0.01 0.5 0.01 0.8 0.4 0.999 0.3 m = 0.0002 0.1 m=0.6 0.9 0.
t) 10 2 Stationary distribution P(N.t) 0. Gaussian approx.08 t=100 J = 10000 m = 0.05 0.t). Exact P(N.J = 1000 m = 0.t).04 0.001 0.02 1000 0. 300000 2000000 800000 2500000 520000 302500 200000 100000 10 10 3 4 10 3 10 800 5 10 4 0 200 400 600 10 1000 0 6 2000 4000 6000 Abundance of species j Abundance of species j Figure 2 12 .06 P(N.02 0 600 10 1 0.01 1000 3000 6000 6000 20000 40000 20000 100000 60000 800 1000 10 0 4000 2 6000 8000 10000 P(N.