You are on page 1of 23

Spatial structure and fluctuations in the

contact process and related models


Robin E. Snyder
Department of Physics
University of California
Santa Barbara, CA 93106, USA
Roger M. Nisbet
Department of Ecology, Evolution, and Marine Biology
University of California, Santa Barbara, CA 93106, USA

Abstract
The contact process is used as a simple spatial model in many disci-
plines, yet because of the buildup of spatial correlations, its dynamics
remain difficult to capture analytically. We introduce an empirically-
based, approximate method of characterizing the spatial correlations
with only a single adjustable parameter. This approximation allows
us to recast the contact process in terms of a stochastic birth-death
process, converting a spatiotemporal problem into a simpler temporal
one. We obtain considerably more accurate predictions of equilibrium
population than those given by pair approximations, as well as good
predictions of population variance and first passage time distributions
to a given (low) threshold. A similar approach is applicable to any
model with a combination of global and nearest-neighbor interactions.

1 Introduction
Spatial correlations can significantly affect population dynamics [1, 2]. Spa-
tial correlations can enable coexistence between competitors [3, 4], may sta-
bilize predator-prey systems [5, 6], or may destabilize predator-prey sys-
tems [7, 8]. Predictions of equilibrium population density and mean time

1
Bulletin of Mathematical Biology (2000) 62, 959–975 2

to extinction that do not consider spatial correlations can be grossly inaccu-


rate. Awareness of this has engendered strong, sustained interest in spatially
explicit models. Unfortunately, making a model spatially explicit greatly
complicates analysis because the dimensionality of the problem increases dra-
matically. Without space, one typically needs to account for the variation
of a few state variables with time. With space, one needs to account for the
variation of those state variables at every point in space as a function of time.
In order to study the effects of spatial correlations, researchers in many
disciplines turn to the contact process as a foundational, no-frills spatial
model. In this model space is divided into a grid, each site of which can
be occupied or vacant. Occupied sites become vacant at a constant rate.
Empty sites are recolonized at a rate proportional to the number of occupied
nearest neighbors. What an occupied site represents determines the model’s
application.
In a metapopulation context, each site represents a patch which can sup-
port a population. The local population dynamics are unstable, however,
so no one population persists. Instead, the metapopulation survives as local
populations wink in and out, perpetually undergoing local extinctions and
recolonizations. The contact process has also been used to model the spatial
spread of single populations, such as daffodils [9, 10]. In this case each occu-
pied site represents an individual, not a population. Simple epidemics have
been modeled with the contact process [11, 12], and it has also been incor-
porated into spatial models of predator-prey dynamics [13]. In all the above
applications, properties of interest include the equilibrium or time-averaged
occupancy, the size and temporal structure of fluctuations away from equi-
librium, and estimates of persistence time. This paper shows a way to find
all of these quantities, given knowledge of a single parameter.
Mathematically, the contact process is a continuous-time, spatially-explicit,
stochastic model with local dispersal. Other well-known models of spatially
distributed populations use different choices. For example, Levins’s metapop-
ulation model has similar rules but is deterministic and assumes global dis-
persal [14]. Coupled map lattices track the population within a site as well
as the migration between sites, and use discrete time [1].
What makes the contact process simultaneously interesting and difficult
to analyze is the buildup of spatial correlations due to the local nature of
the interactions. Some have tried to account for these spatial correlations by
making “pair approximations.” [15, 16, 17, 18, 19] The dynamics of single
site occupancy depends on the occupation probabilities for pairs of adjacent
Bulletin of Mathematical Biology (2000) 62, 959–975 3

sites. The equations for pairs of sites depend on triples, and so on in a rapidly
expanding hierarchy of moment equations. The pair approximation consists
of lopping off all but the first two levels of this hierarchy, approximating
probabilities for triples, for example, as the product of pair probabilities and
single site probabilities.
Unfortunately, the gains made in this way are modest. There is a simple
reason for this: the contact process has a critical point at zero occupancy,
and critical points are characterized by correlations which fall off as a power
law. Ignoring all but the correlations between nearest neighbors predicts spa-
tial correlations which fall off exponentially. Pair approximations therefore
underestimate correlation strength near the critical point. It is also difficult
to add noise terms to pair approximation equations, and so pair approxima-
tions tend to remain deterministic, improving the estimate of the mean but
offering no comment on how spatial correlations affect stochastic quantities,
such as variance.
Is there some other way we could predict dynamics near criticality? An
easy way seems unlikely. With considerable effort, physicists have developed
renormalization theory as a way to perform some calculations near critical-
ity. [20, 21]. However, renormalization theory has limitations. While it is
often used to calculate exponents for quantities which scale as power laws, it
is not clear that it could provide a sufficiently detailed description of dynam-
ics for those using the contact process to study spatial processes. All of this
is rather unfortunate, since it is precisely low densities, i.e. densities near the
critical point, with which ecologists are often most concerned. Whether we
interpret patches as containing a single individual or a population, it is at low
densities that demographic stochasticity is likely to lead to global extinction.
In any particular application, we could of course simulate the contact
process many times and measure all quantities of interest, or even derive
and solve equations for ever higher moments. However, in addition to be-
ing time-consuming, these exercises would not contribute much to general
understanding. The value of simplistic models such as the Lotka-Volterra
model, the Ricker model, and the contact process is that they allow us to
gain insight into the mechanisms behind what we observe and to apply that
new understanding to more complex models. What we want for the contact
process is an enlightening approximation. In this paper we shall present ev-
idence that pair probabilities can be written as functions of the equilibrium
occupancy using only a single parameter. This parameter is δc , the death
rate at which the contact process goes extinct. The relation between pair
Bulletin of Mathematical Biology (2000) 62, 959–975 4

probabilities and occupancy in turn completely characterizes the dynamics,


and we are able to calculate all quantities of interest.
By capturing the effects of spatial correlations in the relation between
pair probabilities and the equilibrium occupancy, we have reduced a high-
dimensional problem involving both time and space to a one-dimensional
problem involving merely time. This allows us to model the contact process
as a stochastic birth-death process and to avail ourselves of the analytic
techniques developed for stochastic birth-death processes. We do this in
section 2, obtaining expressions for mean occupancy, occupancy variance, and
first passage time distributions for the occupancy to reach a given threshold.
This approach is applicable not only to the contact process but to any single-
species model with only global and nearest neighbor interactions. In section 3
we demonstrate similar calculations for a version of the contact process with
long-range dispersal. Section 4 concludes with a discussion.

2 Approximate analysis of spatial correlations


in the contact process
2.1 Birth-death representation of the contact process
In the contact process space is discretized into a lattice. Every point on the
lattice can be either occupied or vacant. Occupied sites become vacant at
constant rate δ, irrespective of their local environment. By “rate δ,” we mean
that the probability for a vacancy to occur in an infinitesimal time h is hδ.
Vacant sites become occupied at basic rate r times the number of occupied
nearest neighbors. We can think of vacant sites as being colonized from
occupied nearest neighbors. The maximum rate of colonization is thus zr,
where z is the number of nearest neighbor sites. Without loss of generality,
we take the colonization rate as setting the timescale of the process, so that
r = 1.
We shall analyze the contact process as a stochastic birth-death process
(also known as a one-step process) [22, 23]. For this we need to write the
global “birth rate” (B) and “death rate” (D) as functions of p(t), the prob-
ability that a randomly chosen site will be occupied at time t. The death
rate is straightforward: D(p) = M δp, where M is the number of sites in the
lattice. However, the global birth rate is not a function of p, because the
birth rate at a vacant site depends on the states of its neighbors. Instead,
Bulletin of Mathematical Biology (2000) 62, 959–975 5

B = M zrp10 (t), where p10 (t) is the probability that, given two adjacent sites,
one site will be occupied and the other will be vacant at time t. We need a
way to write p10 as a function of p. That is, we need a way to describe local
correlations.
One measure of local correlation strength is p10 (the actual probability
for an occupied site and a vacant site to be adjacent) normalized by p(1 − p)
(the probability for an occupied site and a vacant site to be adjacent in
the absence of spatial correlations). In fig. 1 we plot instantaneous values of
p10 /(p(1−p)) versus p. Each cluster of points is taken from a single computer
run at a different value of δ. We discarded the beginning portion of each run
so as to avoid transient behavior. All computer realizations were made on a
128 × 128 lattice.
The plot of p10 /(p(1 − p)) versus p is very nearly a straight line. This
allows us to re-express p10 in terms of p and forms the basis of our approx-
imation. The idea of measuring spatial correlations instead of calculating
them came from a discussion with J. Bascompte about a manuscript of his
then in preparation [24]. There is only one parameter to adjust, since we
know that as p approaches 1, the lattice becomes more nearly random, and
p10 /(p(1 − p)) approaches 1. Thus,
p10
≈ α + (1 − α)p (1)
p(1 − p)
and

B(p) = M zrp10 ≈ M zr(α + (1 − α)p)p(1 − p). (2)


Indeed, were the line perfectly straight, there would be no adjustable
p10
parameters, for we also know the value of at p = 0. Setting B(p)
p(1 − p)
δp p10
equal to D(p) and solving for p10 , we find p10 = and thus =
zr p(1 − p)
δ δc
. The intercept is therefore , where δc is the death rate at which
zr(1 − p) zr
the contact process goes extinct. The value of δc has been estimated to be
2.424 [25], making the theoretical intercept 0.6060. Because the plot of local
correlation strength versus p (Fig. 1) does not follow precisely a straight line,
δc
a best-fit line yields an estimate of α slightly greater than (α = 0.6354).
zr
Nonetheless, the difference is small, and we see that knowledge of the critical
Bulletin of Mathematical Biology (2000) 62, 959–975 6

value of δ almost entirely specifies the local correlation structure.


We are not always blessed with knowledge of the critical point. However,
it is straightforward to estimate α from a linear regression, and as long as
the relation between local correlation strength and p is nearly linear, we can
even base that regression solely on measurements at high density. This is
convenient for models with critical points at p = 0 since systems settle to a
quasi-stationary state much faster when they are far from a critical point,
and so simulations are less time-consuming at high density. As an example,
fig. 1 shows regressions based on all of the data and on p > 0.6, yielding
α = 0.6354 and 0.6571, respectively. Both regressions are constrained to
pass through (1, 1).

2.2 Predictions
Having found an approximate restatement of the contact process in terms of
a stochastic birth-death process, we can now rapidly calculate the mean pop-
ulation density, the population variance, and first passage time distributions
for the population to reach low thresholds. (For concreteness, we interpret
the contact process as describing the spatial spread of a single population in
the remainder of this section and in the following section. Thus, “popula-
tion” refers to the number of occupied sites, and “population density” refers
to the fraction of occupied sites.) The mean population density can be well
approximated as the equilibrium density of a deterministic model with the
same birth and death rates [23]. Setting the global birthrate, B(p), equal to
the global death rate, D(p), we obtain
q
zr(2α − 1) − zr(zr − 4δ(1 − α))
p∗ = (3)
2zr(α − 1)
Eq. 3 also provides a quick means of estimating α. We simply solve for α in
terms of p and plug in the observed equilibrium population density and the
value of δ used. A plot of actual densities and predicted densities is shown
in fig. 2.
To calculate population variance, σn2 , we treat the number of occupied
sites, n = M p, as continuous and describe the dynamics with a stochastic
differential equation. This is acceptable as long as n is not too small [23]. The
most straightforward way of finding the variance involves linearizing B(n)
Bulletin of Mathematical Biology (2000) 62, 959–975 7

and D(n) about equilibrium and is detailed in a number of sources [23, 26, 27].
The general result is that
Q
σn2 = , (4)
2a

∗ ∗ ∂
where Q = B(n )+D(n ) and a = − (B(n) − D(n)) . The variance is
∂n
n=n∗
thus determined by a balance between transition rates (Q), which determine
how easily the system jumps away from equilibrium, and the response rate
to perturbations (a), which determines how rapidly the system is dragged
back to equilibrium. For our system

M [(α − 1)p2 + (1 − 2α)p + α]


σn2 = . (5)
2(1 − α)p + 2α − 1
A plot of variance vs. p is given in fig. 3.
Some readers might be concerned about the effects of linearizing and
treating the population as continuous. It is possible to avoid making these
simplifications by solving numerically for the population probability distri-
bution function conditional upon the system not going extinct (CPDF). The
master equation which defines our stochastic birth-death process is
dp(n = i, t)
= B(i − 1)p(i − 1, t) + D(i + 1)p(i + 1, t) (6)
dt
−(B(i) + D(i))p(i, t), i = 1, 2, . . .
dp(n = 0, t)
= D(n = 1)p(1, t). (7)
dt
Noting that B(n)Pc∗ (n) = D(n + 1)Pc∗ (n + 1), where Pc∗ (n) denotes the
stationary CPDF, we can derive an expression for Pc∗ (n) in terms of Pc∗ (1) by
applying this relation repeatedly. We proceed by guessing the value of Pc∗ (1),
building up the distribution Pc∗ (n), and using the normalization condition to
refine our guess for Pc∗ (1). Further details are given in [23, sect. 6.4.A].
While precise, this method adds little to our understanding, and a previous
study suggests that the linearized, continuous approximation is actually quite
good [28]. We have confirmed this in the present case (see figs. 2 and 3). The
equilibrium population found from the CPDF was indistinguishable from that
predicted by eq. 3. The variance found from the CPDF differed from eq. 4
only at very low populations, where treating n as continuous is presumably
no longer appropriate (see fig. 3).
Bulletin of Mathematical Biology (2000) 62, 959–975 8

When investigating a stochastic model, our most pressing question is often


how long will the system persist? While the ratio of variance to mean can
provide a rough estimate of viability (see for example [27]), a more precise
approach is to calculate a first passage time distribution (FPTD) – that is, a
probability distribution for the time at which the population first dips below
a given threshold. Having captured the dynamics of the contact process
with our expression for the birth rate, we use our approximation to present
a calculation of a first passage time distribution.
To calculate an FPTD to threshold S, we add to the equations defining
our stochastic birth-death process (eq. 6) an absorbing boundary at S:

dp(n = S, t)
= D(n = S + 1)p(S + 1, t) (8)
dt
For our initial condition, we set p(n = M p∗ , t = 0) = 1 and p(n 6= M p∗ , t =
0) = 0, representing an ensemble in which all replicates begin with population
equal to the equilibrium value.
Calculation of the FPTD is very sensitive to the value and slope of the
population distribution at S. Unfortunately, when we estimate α from a
linear regression, as described in section 2.1, the resulting prediction for the
population distribution is not sufficiently accurate to make a good prediction
of the FPTD. However, the population distribution is very nearly Gaussian as
long as δ is not too close to the critical point, and for a Gaussian distribution,
specifying the derivative or the value at any point fixes α. Therefore, we
have simulated the contact process for δ = 2 and fit the resulting population
distribution to a Gaussian curve. All runs were started with population equal
to the equilibrium value, so as to be consistent with the initial conditions
used in our differential equations (eq. 6). We found the derivative of our
fitted Gaussian at S = 669, a population three standard deviations below
the equilibrium value, and from this determined α. The change in α was
p10
not great: the value of α determined from a linear regression of on
p(1 − p)
p was 0.6354, while the value of α obtained in the present case was 0.6208.
We obtained similar results when we determined α by fixing the value of the
population distribution at S instead of its slope. The resulting prediction
of the FPTD is reasonably good given the high sensitivity to α. The true
FPTD declines exponentially with decay rate 0.00517. Our prediction has
decay rate 0.0073. A plot comparing our predictions with data is given in
fig. 4.
3 THE CONTACT PROCESS WITH DISPERSAL 9

3 The contact process with dispersal


So far, our model has only considered interactions between nearest neighbors.
However, many organisms disperse over larger distances, and the presence of
even a small level of long-range dispersal can significantly affect population
dynamics [29, 2, and references therein]. Our method of measuring near-
est neighbor correlations is not sufficient for studying full-blown dispersal
kernels. However, we consider a caricature, ignoring intermediate-range dis-
persal but adding global dispersal, as was studied by Harada and Iwasa [30].
The global dynamics of this model are also well approximated as a stochastic
birth-death process.
Suppose a fraction  of the colonists from the occupied sites are caught
by random breezes or passing animals and land on a random site. Then

B(p, p10 ) = M (zr(1 − )p10 + zrp(1 − p)) (9)


and
D(p) = M δp. (10)
Once again, we can measure p10 /(p(1 − p)) and note that it follows a straight
line so that p10 /(p(1 − p)) = α + (1 − α)p (see fig. 5). Regressing on all p,
we find that α = 0.8097.
Following the same procedures as in section 2.2, we find

q
zr(1 + 2α( − 1) − 2) + zr(zr + 4δ(−1 + α +  − α))
p∗ = (11)
2zr(1 − α)(1 − )

and

−M (1 − p)[ + α(1 − )(1 − p) + p(1 − )]


σn2 = . (12)
1 − 2α(1 − )(1 − p) − 2(1 − p) − 2p
Plots of equilibrium occupancy versus δ and variance versus p can be found
in figs. 6, 7. Looking at figures 6 and 7, we see that when dispersal is present,
the equilibrium population is higher and the variance is lower, making the
population less likely to fluctuate to low levels. A calculation of the FPTD for
S three standard deviations below the equilibrium population is compared
with data in fig. 8. The curve representing the data decays exponentially
with decay rate 0.0204 and our prediction has decay rate 0.0201. As in the
Bulletin of Mathematical Biology (2000) 62, 959–975 10

last section, we used a value of α designed to give us the correct derivative of


the population distribution with respect to n at the boundary. Comparing
with fig. 4, we see that the presence of dispersal significantly reduces first
passage times to population levels an equal number of standard deviations
below the equilibrium value.

4 Discussion
Spatial correlations can strongly influence population dynamics. However,
describing the effects of spatial correlations mathematically is difficult. Equa-
tions for first-order moments rely on second-order moments; equations for
second-order moments rely on third-order moments. The cascade of equa-
tions needs to be truncated at some point, but it is difficult to do so without
losing a lot of information.
We avoid ad hoc closure methods by measuring the relation between the
second-order moment p10 and the first-order moment p from simulations. By
more accurately estimating p10 , we are able to substantially improve upon the
predictions of moment closure schemes such as pair approximations. Since
the interaction rules for our models depend only on nearest neighbors, in-
formation about second-order moments fully specifies the dynamics. Having
captured the effects of space in the relation between p10 and p, we use this re-
lation to create a nonspatial description of the contact process as a stochastic
birth-death process. We are then able to apply the many techniques available
for analyzing stochastic birth-death processes to obtain predictions for equi-
librium occupancy, occupancy variance, and first passage time distributions.
Our method characterizes spatial correlations better than pair approxima-
tions or even “improved pair approximations” [15, 31]; however, our greatest
advantage over these methods is not our greater numerical accuracy but the
ease with which we can consider inherently stochastic quantities such as vari-
ance and first passage time distributions. It is possible to add appropriately
derived noise terms to pair approximation equations, but working with the
resulting equations is laborious, it is nearly impossible to derive simple ex-
pressions for quantities such as the variance, and improvements over methods
which ignore spatial correlations are disappointingly small [32]. In contrast,
our method yields insights into the mechanisms by which spatial correlations
affect stochastic quantities.
As an example, consider the variance of the contact process without dis-
Bulletin of Mathematical Biology (2000) 62, 959–975 11

persal. Ignoring spatial correlations causes us to greatly underestimate the


variance. (Compare the mean field prediction (no spatial correlations) to the
data in fig. 3.) However, mean field theory overestimates the birth rate, since
it ignores the tendency for occupied sites to cluster and increase crowding.
This would naively lead us to expect a variance that is too large, since the
system can jump away from equilibrium more rapidly. However, as we noted
in section 2.2, variance is determined by the balance between this turnover

rate and a response rate, represented by − (B(n) − D(n))
, which
∂n
n=n∗
measures how quickly the system returns to equilibrium after a fluctuation.
By comparing the mean field birth rate, M zrp(1 − p), and the birth rate we
derived by our method (eq. 2), we see that mean field theory overestimates
the response rate. That is, the mean field variance is too small because ignor-
ing spatial correlations causes us to overestimate the response rate even more
drastically than the turnover rate. Furthermore, we can see that the overly
large birth rate is ultimately the source of the error in the response rate. The
birth rate must of course be zero for a completely empty or completely full
lattice. Because mean field theory predicts too high a birth rate in between
these two extremes, the birth rate shoots up too rapidly as occupancy in-
creases from
zero and plummets too rapidly as occupancy approaches one,
∂B
making too large.

∂p
While we have focused on the contact process, this technique could be
applied to single-species models whose interactions depend on the state of
surrounding sites in more complicated ways. For example, Wilson et al.’s
mussel model [33] contains a “safety in numbers” term. Mussels with fewer
than three neighbors are deemed vulnerable to wave action and have a higher
mortality rate. In this case, we would need to measure not pair probabilities
but probabilities for various configurations of the entire neighborhood. Al-
though straightforward, this procedure can generate an unwieldy number of
parameters. For each new configuration that is associated with an interac-
tion, we need to measure the probability of that configuration as a function
of p, which adds parameters. Moreover, not all of these functions may be
well approximated by a line, which adds further parameters.
The technique we have used to analyze the contact process may or may
not be extendable to multi-species models with nearest neighbor interactions,
such as predator-prey models and susceptible-infected-recovered (SIR) epi-
demic models. The difficulty stems from the presence of multiple independent
Bulletin of Mathematical Biology (2000) 62, 959–975 12

variables. Instead of relating spatial correlations to equilibrium occupancy


and fitting a curve, we must relate spatial correlations to the equilibrium
occupancies of each species and fit a surface.
There remain a few caveats. Care needs to be taken in regions of metasta-
bility. If there are two stable states, then even if one is always the eventual
favorite, the probability distribution function will be bimodal. The system
trajectory will be subject to competing attractions to the two equilibria [23].
Predicting variance from linearized stochastic differential equations, on the
other hand, assumes that fluctuations are always drawn back to the same
equilibrium. This situation results in a Gaussian distribution (assuming lin-
earized interactions).
The method we have presented is also not appropriate for examining
metapopulations with only a few patches. In switching to differential equa-
tions, we assumed that we could treat the number of occupied sites as con-
tinuous. This assumption breaks down when the number of occupied sites
is small. Small systems also tend to be dominated by edge effects, whereas
we have assumed that the proportion of sites with fewer than z neighbors is
negligible.
Finally, the relation we have measured between p10 and p is valid near
equilibrium only. This means it is not appropriate for nonequilibrium situa-
tions such as invading waves. On the other hand, Ellner et al. have presented
an invasion speed prediction for the contact process that uses simple modi-
fications of equilibrium spatial correlations [34]. Their estimate of p10 comes
from a pair approximation. We have tried substituting ours instead but found
that it did not significantly improve their prediction. Evidently, the error in
p10 is less important than the error introduced by using modified equilibrium
values to describe spatial correlations in the wavefront.

5 Acknowledgments
We thank Jordi Bascompte, Benjamin Bolker, Douglas Donalson, Parviez
Hosseini, Timothy Keitt, Bruce Kendall, and Daniel Rabinowitz for helpful
discussions. One of us (RES) is grateful for financial support from NSF
grants BIR94-13141 and GER93-54870.
REFERENCES 13

References
[1] M. Keeling. Spatial models of interacting populations. In Jacqueline
McGlade, editor, Advanced Ecological Theory: Principles and Applica-
tions, pages 64–99. Blackwell Science, Oxford, 1999.
[2] P. Kareiva. Population dynamics in spatially complex environments:
Theory and data. Philosophical Transactions of the Royal Society of
London B Biological, 330(1257):175–190, 1990.
[3] I. Hanski. Coexistence of competitors in patchy environment. Ecology,
64:493–500, 1983.
[4] M. Slatkin. Competition and regional coexistence. Biometrika, 55:128–
134, 1974.
[5] G. Nachman. Systems analysis of acarine predator-prey interactions:
II. the role of spatial processes in system stability spatial processes in
system stability. Journal of Animal Ecology, 56(1):267–282, 1987.
[6] B. P. Zeigler. Persistence and patchiness of predator-prey systems in-
duced by discrete event population exchange mechanisms. Journal of
Theoretical Biology, 67:687–713, 1977.
[7] J. D. Reeve. Environmental variability, migration, and persistence in
host-parasitoid systems systems. American Naturalist, 132(6):810–836,
1988.
[8] J. Allen. Mathematical models of species interactions in time and space.
American Naturalist, 109:319–342, 1975.
[9] R. Durrett and S. A. Levin. Stochastic spatial models: A user’s guide to
ecological applications. Philosophical Transactions of the Royal Society
of London B Biological, 343(1305):329–350, 1994.
[10] J. P. Barkham and C. E. Hance. Population dynamics of the wild daffodil
(narcissus pseudonarcissus) III. implications of a computer model of
1000 years of population change. Journal of Ecology, 70:323–344, 1982.
[11] T. Ohtsuki and T. Keyes. Kinetic growth percolation: epidemic pro-
cesses with immunization. Physical Review A, 33(2):1223–32, February
1986.
REFERENCES 14

[12] Shane A. Richards, William G. Wilson, and Joshua E. S. Socolar. Se-


lection for short-lived and offspring-limited individuals in a spatially-
structured population which is subject to localized disturbances. Un-
published manuscript, 1999.

[13] E. McCauley, W. G. Wilson, and A. M. De Roos. Dynamics


of age-structured and spatially structured predator-prey interactions
individual-based models and population-level formulations. American
Naturalist, 142(3):412–442, 1993.

[14] R. Levins. Some demographic and genetic consequences of environmen-


tal heterogeneity for biological control. The Bulletin of the Entomological
Society of America, 15:237–240, 1969.

[15] K. Satō, H. Matsuda, and A. Sasaki. Pathogen invasion and host extinc-
tion in lattice structured populations. Journal of Mathematical Biology,
32(3):251–268, 1994.

[16] Y. Harada, H. Ezoe, Y. Iwasa, H. Matsuda, and K. Satō. Population


persistence and spatially limited social interaction. Theoretical Popula-
tion Biology, 48(1):65–91, 1995.

[17] D. A. Rand. Correlation equations and pair approximations for spatial


ecologies. In Jacqueline McGlade, editor, Advanced Ecological Theory:
Principles and Applications, pages 100–142. Blackwell Science, Oxford,
1999.

[18] R. Dickman. Kinetic phase transitions in a surface-reaction model:


mean-field theory. Physical Review A, 34(5):4246–50, November 1986.

[19] H. Matsuda, N. Ogita, A. Sasaki, and K. Satō. Statistical mechanics


of population-the lattice Lotka-Volterra model. Progress of Theoretical
Physics, 88(6):1035–49, December 1992.

[20] K. G. Wilson. The renormalization group and critical phenomena. Re-


views of Modern Physics, 55:583–600, 1983. Nobel Prize Lecture.

[21] M. E. Fisher. Renormalization group theory: Its basis and formulation


in statistical physics. Reviews of Modern Physics, 70(2):653–81, April
1998.
REFERENCES 15

[22] N. G. van Kampen. Stochastic Processes in Physics and Chemistry.


North-Holland, Amsterdam, 1981.
[23] R. M. Nisbet and W. S. C. Gurney. Modelling Fluctuating Populations.
John Wiley & Sons, New York, 1982.
[24] Jordi Bascompte. Aggregate statistical measures and metapopulation
dynamics. Submitted to Journal of Animal Ecology, 1999.
[25] M. Katori and N. Konno. Correlation inequalities and lower bounds for
the critical value lambda /sub c/ of contact processes. Journal of the
Physical Society of Japan, 59(3):877–87, March 1990.
[26] Eric Renshaw. Modelling Biological Populations in Space and Time.
Cambridge University Press, New York, 1991.
[27] W. S. C. Gurney and R. M. Nisbet. Single-species population fluctua-
tions in patchy environments. The American Naturalist, 112(988):1075–
1090, 1978.
[28] R. M. Nisbet, W. S. C. Gurney, and M. A. Pettipher. An evaluation of
linear models of population fluctuations. Journal of Theoretical Biology,
68:143–160, 1977.
[29] Nanako Shigesada and Kohkichi Kawasaki. Biological Invasions: Theory
and Practice. Oxford University Press, New York, 1997.
[30] Yuko Harada and Yoh Iwasa. Lattice population dynamics for plants
with dispersing seeds and vegetative propagation. Researches on Popu-
lation Ecology (Kyoto), 36(2):237–249, 1994.
[31] Minus van Baalen and David A Rand. The unit of selection in vis-
cous populations and the evolution of altruism. Journal of Theoretical
Biology, 193(4):631–648, August 1998.
[32] Robin E. Snyder. Effects of spatial correlations and fluctuations on
population dynamics. Ph.D. thesis, in prep., University of California,
Santa Barbara, 2000.
[33] W G Wilson, R M Nisbet, A H Ross, C Robles, and R A Desharnais.
Abrupt population changes along smooth environmental gradients. Bul-
letin of Mathematical Biology, 58(5):907–922, 1996.
Bulletin of Mathematical Biology (2000) 62, 959–975 16

[34] Stephen P. Ellner, Akira Sasaki, Yoshihiro Haraguchi, and Hirotsugu


Matsuda. Speed of invasion in lattice population models: Pair-edge
approximation. Journal of Mathematical Biology, 36(5):469–484, 1998.

1.05

1
local correlation strength

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0.6

0.55
0 0.2 0.4 0.6 0.8 1
p

p10
Figure 1: Local correlation strength, , as a function of population
p(1 − p)
density for the contact process. Each cross marks the instantaneous pop-
ulation density and local correlation strength for a snapshot of the lattice.
Each cluster of points is taken from a run with a different value of δ. The
long dashed line represents a linear regression on p > 0.6 and has intercept
α = 0.6571. The short dashed line represents a linear regression on all p and
has intercept α = 0.6354. Both regressions are constrained to pass through
the point (1, 1). Both p-values are zero to four decimal places.
Bulletin of Mathematical Biology (2000) 62, 959–975 17

1.1

0.9

0.8

0.7

0.6
p

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2

Figure 2: Mean population density as a function of the death rate, δ, for the
contact process. The crosses represent simulation results: for several runs
we averaged the fraction of occupied sites over a long time interval, then
averaged the time averages. The error bars are too small to be seen. The
dash-dot line gives the mean field (no spatial correlations) prediction, the
solid line uses our approximation with α = 0.6354, and the dotted line uses
δc
our approximation with α = . All simulations of the contact process were
zr
performed on a 128 × 128 lattice with the birthrate, r, set to 1. We discarded
the initial portion of each run so as to avoid transients.
Bulletin of Mathematical Biology (2000) 62, 959–975 18

50000

45000

40000
population variance

35000

30000

25000

20000

15000

10000

5000

0
0 0.2 0.4 0.6 0.8 1
p

Figure 3: Population variance as a function of the equilibrium population


density for the contact process. Equilibrium population density was found
as in fig. 2. The crosses represent simulation results with error bars set at
one standard error. Horizontal error bars are too small to see. The solid
line gives the mean field prediction using a linearized stochastic differential
equation, while the dotted line gives the prediction using the CPDF. The
δc
dashed line comes from our approximation with α = and the dash-dot
zr
line comes from our approximation with α = 0.6354.
Bulletin of Mathematical Biology (2000) 62, 959–975 19

1
Log(cumulative FPT probability)

0.5

0.25

0.125

0.0625

0.03125

0.015625

0.0078125

0.00390625

0.00195312
0 200 400 600 800 1000 1200

Time

Figure 4: Cumulative first passage time distribution to n = 669 for the


contact process, δ = 2.0. The solid line represents data, while the dotted
line represents our prediction. All runs were performed on a 50 × 50 lattice
with an initial population density equal to 0.3257, the equilibrium population
density. The absorbing boundary was placed at n = 669, a population level
three standard deviations below the equilibrium value. Note the logarithmic
scale on the vertical axis.
Bulletin of Mathematical Biology (2000) 62, 959–975 20

1.05
local correlation strength

0.95

0.9

0.85

0.8

0.75
0 0.2 0.4 0.6 0.8 1

p10
Figure 5: Local correlation strength, , as a function of population
p(1 − p)
density for the contact process with dispersal,  = 0.3. Each cross marks
the instantaneous population density and local correlation strength for a
snapshot of the lattice. Each cluster of points is taken from a run with a
different value of δ. The dashed line represents a linear regression on all p
and has intercept α = 0.8097. The regression is constrained to pass through
the point (1, 1). The p-value is zero to four decimal places.
Bulletin of Mathematical Biology (2000) 62, 959–975 21

0.8

0.6
p

0.4

0.2

0
0 1 2 3 4

Figure 6: Equilibrium population density as a function of the death rate,


δ, for the contact process with dispersal,  = 0.3. Equilibrium population
density was found as in fig. 2. The crosses represent simulation results. The
error bars are too small to be seen. The solid line gives the mean field (no
spatial correlations) prediction, the dashed line uses our approximation with
α = 0.8097. The CPDF prediction is almost the same as our approximation
and is shown by a dotted line. All simulations of the contact process with
dispersal were performed on a 50 × 50 lattice instead of a 128 × 128 lattice
because this model is more computationally intensive. The birthrate, r, was
set to 1. We discarded the initial portion of each run so as to avoid transients.
For comparison, results for the contact process without dispersal on the same
size lattice are shown with circles.
Bulletin of Mathematical Biology (2000) 62, 959–975 22

5000

4500
population variance

4000

3500

3000

2500

2000

1500

1000

500

0
0 0.2 0.4 0.6 0.8 1

Figure 7: Population variance as a function of the equilibrium population


density for the contact process with dispersal,  = 0.3. Equilibrium popu-
lation density was found as in fig. 2. The circles with error bars represent
simulation results with error bars set at one standard error. The solid line
gives the mean field prediction using a linearized stochastic differential equa-
tion. The dashed line comes from our approximation with α = 0.8097. The
CPDF prediction is shown by a dotted line. For comparison, results for the
contact process without dispersal on the same size lattice are shown with
stars and error bars connected by a dash-dot line.
Bulletin of Mathematical Biology (2000) 62, 959–975 23

1
Log(cumulative FPT probability)

0.5

0.25

0.125

0.0625

0.03125

0.015625

0.0078125
0 50 100 150 200 250

Time

Figure 8: Cumulative first passage time distribution to n = 669 for the


contact process with dispersal, δ = 2.0 and  = 0.3. The solid line represents
data, while the dotted line represents our prediction. All runs were performed
on a 50 × 50 lattice with an initial population density equal to 0.4519, the
equilibrium population density. The absorbing boundary was placed at n =
1016, a population level three standard deviations below the equilibrium
value. Note the logarithmic scale on the vertical axis.

You might also like