Preprint of article published 2007 in Biological Reviews 82:335–359.
doi:10.1111/j.1469-185X.2007.00014.x
CORRECTIONS TO FORMULA (23) AND TO CAPTIONS TO FIGS 1, 9 and 10 ADDED NOV 2008.
CORRECTION REGARDING FORMULA (15) ADDED AUG 2014: THOMAS CLERC SPOTTED THIS.
Use, misuse and extensions of “ideal gas”
models of animal encounter
John M. C. Hutchinson1* and Peter M. Waser2
1
Max-Planck-Institut für Bildungsforschung, Lentzeallee 94, 14195 Berlin, Germany and
School of Biological Sciences, University of Bristol, Bristol, BS8 1UG, UK
2
Department of Biological Sciences, Purdue University, W. Lafayette, IN 47907, USA
(Received 10 November 2005; revised 26 February 2007; accepted 12 March 2007)
* Address for correspondence: John Hutchinson, Privat, Staatliches Museum für
Naturkunder Görlitz, PF 300154, 02086 Görlitz, Germany. Tel: +49-3581-4760410 Fax:
+49-3581-4760499. Email: majmch@googlemail.com
ABSTRACT
Biologists have repeatedly rediscovered classical models from physics predicting collision
rates in an ideal gas. These models, and their two-dimensional analogues, have been used
to predict rates and durations of encounters among animals or social groups that move
randomly and independently, given population density, velocity, and distance at which an
encounter occurs. They have helped to separate cases of mixed-species association based
on behavioural attraction from those that simply reflect high population densities, and to
detect cases of attraction or avoidance among conspecifics. They have been used to
estimate the impact of population density, speeds of movement and size on rates of
encounter between members of the opposite sex, between gametes, between predators and
prey, and between observers and the individuals that they are counting. One limitation of
published models has been that they predict rates of encounter, but give no means of
determining whether observations differ significantly from predictions. Another
uncertainty is the robustness of the predictions when animal movements deviate from the
model’s assumptions in specific, biologically relevant ways. Here, we review applications
of the ideal gas model, derive extensions of the model to cover some more realistic
movement patterns, correct several errors that have arisen in the literature, and show how
to generate confidence limits for expected rates of encounter among independently
moving individuals. We illustrate these results using data from mangabey monkeys
originally used along with the ideal gas model to argue that groups avoid each other.
Although agent-based simulations provide a more flexible alternative approach, the ideal
gas model remains both a valuable null model and a useful, less onerous, approximation to
biological reality.
Key words: Cercocebus albigena, contact duration, encounter rate, fertilization kinetics,
home range, line transect, mangabey, predation rate, search theory, random walk.
1
CONTENTS
I.
II.
III.
IV.
V.
VI.
VII.
VIII.
IX.
X.
XI.
Introduction: a history of repeated rediscovery
Derivation of basic results
(1) The ideal gas model
(2) Associations versus encounters; counts and durations of associations
Analytically tractable complications: deviations from model assumptions about
detection distance, density, speed and direction
(1) Detection distance
(2) Density
(3) Speed
(4) Direction
Tracks that are not straight
(1) Monte-Carlo simulations
(2) Correlated random walks
(3) Rebounds
(4) Other forms of non-independence
Corrections of some errors in the literature
(1) Encounter duration
(2) D for spread-out groups
(3) Partial overlap of home ranges
(4) Infanticide rate
(5) Mating benefits of an increased home range
(6) Spiral trajectories of sperm
(7) Predation in the plankton
(8) Estimating seabird densities from ship- and aircraft-based counts
How to test the predictions?
An example: grey-cheeked mangabey intergroup encounters reanalysed
Discussion
Conclusions
Acknowledgements
References
I. INTRODUCTION: A HISTORY OF REPEATED REDISCOVERY
In 1860, the physicist James Clerk Maxwell worked out the expected rates of collision
among molecules in an ideal gas, given the concentration of molecules in the gas, their
mean speed, and their “cross section”, a measure of their size. His formula assumes that
the movements of the molecules are independent, equally likely in all directions, and with
speeds drawn from a “Maxwell-Boltzmann” distribution [i.e. the x and y components of
their velocities are normally distributed: e.g. Maxwell (1860); Kauzmann (1966, chapter
5)]. These assumptions have proven attractive to biologists, who have repeatedly
rediscovered Maxwell’s approach in constructing models of encounter rates between
moving animals. The image of moving animals as molecules has proven equally
appealing, and in the biological literature the term “ideal gas model” has come to be
attached to ideas that have developed from these assumptions.
One biological use of the ideal gas approach is as a null model. For example, Waser
(1975) used it to detect cryptic behavioural attraction between adults in a “solitary”
antelope species. Because observed rates of encounter among individual bushbuck
2
(Tragelaphus scriptus) were an order of magnitude higher than expected if bushbuck
behaved like gas molecules in a two-dimensional bottle, he argued that individual
movements were not independent. Waser (1976) subsequently applied the same null
model to detect cryptic cases of behavioural avoidance between social groups of greycheeked mangabeys (Cercocebus albigena). Mitani et al. (1991) used the approach to
show that encounter rates and association durations among some classes of orang-utans
(Pongo pygmaeus) were greater than expected by chance. Similarly Schülke & Kappeler
(2003) and Gursky (2005) working on prosimians (Tarsius spectrum and Phaner furcifer)
demonstrated associations between members of social pairs that had usually been
considered to forage solitarily. Barrett & Lowen (1998) and Sugiura et al. (2000) have
used the ideal gas model to ask why patterns of interanimal spacing sometimes differ
among primate populations at different densities. Do individuals change their behaviour
with density, or is it simply that encounter rates change with density, and the same
behavioural rules expressed at different rates of encounter produce different spacing
patterns?
Ideal gas models are readily generalized to investigate patterns of encounter between,
as well as within, species. Thus Crowley et al. (1991) used this approach twice in the same
model of mating choosiness, once for encountering predators and once for encountering
mates. Waser (1982, 1984, 1987) used the ideal gas model to generate null predictions
regarding the frequency and duration of mixed-species primate associations. Several other
investigators (Cords, 1987; Whitesides, 1989; Holenweg, Noë & Schabel, 1996) have
subsequently used this approach.
In most of the above examples, and others (e.g. De Vita, Kelly & Payne, 1982), the
ideal gas model provides a null hypothesis and interest focuses on deviations from its
predictions. In other cases, ideal gas assumptions are used to estimate encounter rates as a
component of some larger model. For example, Rowcliffe, Cowlishaw & Long (2003)
recently modelled the effect of human hunting on mammalian population density
assuming that prey individuals encounter snares at rates predicted using ideal gas logic.
Jetz et al. (2004) modelled how home range size scales with body size, assuming that the
proportion of resources lost to neighbours is related to encounter rate; the ideal gas model
enabled the scaling of encounter rate to be estimated from the known scaling relationships
of speed of movement, population density, and detection distance.
One context in which the ideal gas model has often been applied is mate finding.
Mosimann (1958) analysed the probability of a female encountering no males during the
breeding season in low-density populations. In effect, he quantified the suggestion of
Allee (1938) that populations below some minimum density would decline because of the
difficulty of finding mates. Katona (1973) modelled the effect of diffusible sex
pheromones in reducing the time taken for planktonic copepods to find mates. Manica &
Johnstone (2004) estimated the reduction in a male bug’s encounter rate with females if he
stopped moving around so as to be able to guard a brood. Similarly, a series of
primatology papers have applied the ideal gas model to compare mating success of a
roving polygamous strategy with that of monogamy (van Schaik & Dunbar, 1990; Dunbar,
1995, 2000). Another application was to estimate infanticide rates in great apes, under the
assumption that infanticide occurs if a male fails to encounter the female during oestrus,
but does encounter her while she is nursing (Harcourt & Greenberg, 2001).
In the field of fertilization kinetics, and particularly concerning the issue of the
evolution of anisogamy, Scudo (1967, p. 285) noted that “the adaptive value of various
modes of reproduction can be studied…using the same tools which classical statistical
mechanics applies to the collision of particles”. Research in this field continues to be
active (Randerson & Hurst, 2001; Dusenbery, 2002, 2006), with most models taking as
3
their starting point the ideal gas model of encounter rate (e.g. Vogel et al., 1982; Cox &
Sethian, 1985; Dusenbery, 2000). One issue of recent interest is the function and optimal
thickness of the jelly coat around some eggs; it increases target size but tends to decrease
sinking speed, which the gas model predicts to have opposite, although unequal, effects on
encounter rate with sperm (Farley & Levitan, 2001; Podolsky, 2002).
Lotka (1924, p. 358) derived the mean free path of a predator (the distance moved
between successive encounters with prey) from “the elementary kinetic theory of gases”.
He went on to show that the frequency of predator-prey encounters should be proportional
to predator speed and “size”, and to population densities of predator and prey, using this
result to justify the multiplicative form of the predator-prey encounter term in his famous
equation. Laing (1938), following Stanley (1932), used the same approach to show that
parasitoids found hosts at too high a rate to be explained by random movement, generating
one of the first quantitative arguments in evidence of area-restricted search. Gerritsen &
Strickler (1977) re-derived formulae for encounter rates in three dimensions and these
have had “…considerable influence in subsequent studies of plankton feeding” (Evans,
1989, p. 415). One instance is understanding rates and size-selectivity of predation on the
larvae of commercially significant fish (e.g. Fuiman & Gamble, 1989; Paradis & Pepin,
2001). Letcher et al. (1996) modelled predation and starvation in larval fish using the ideal
gas model twice, to calculate encounter rate with both its predators and prey.
In a quite different context, Yapp (1956) suggested that “…encounters between a
moving observer and the individuals of a mobile species could be likened to the collisions
between a molecule of one kind and molecules of another”. Accordingly, Yapp (1956) and
Skellam (1958) applied the encounter-rate equation in reverse, to estimate the density of a
target species from the number of times an observer walking a line transect encounters the
species, taking into account the observer’s speed, the average speed of target individuals,
and the distance at which the observer can detect them. Estimates of seabird numbers are
often made from such data collected from ships or aircraft, but curiously the literature
seems to have largely lost sight of the ideal gas approach. The modern theory of “distance
sampling” (Buckland et al., 2001) also estimates density from transect data and has
become a rather sophisticated technique, but it assumes that the targets are stationary (and
claims reasonable accuracy only if they move at less than half the speed of the observer:
Buckland et al., 2001, p. 31). Gaston & Smith (1984) and Spear, Nur & Ainley (1992)
independently redeveloped some aspects of the ideal gas model to correct for biases when
the birds predominantly fly in a particular direction.
“Search theory”, originally developed to optimize the search for submarines during
World War II, leads to the same encounter-rate predictions as ideal gas models in the null
case of randomly moving searchers and randomly moving targets (Koopman, 1956;
Dusenbery, 1992). This literature was the inspiration for biologists to apply the ideal gas
model in a variety of situations: encounters between zooplankton predators and
phytoplankton prey (Gerritsen & Strickler, 1977), the effect of prey aggregation on risk of
predation (Olson, 1964; Kiltie, 1980), encounters among gametes of aquatic plants (Cox,
1983), and the evolution of parthenogenesis in sparse populations (Gerritsen, 1980b).
During this long history of repeated rediscovery, two limitations of the ideal gas
approach have surfaced. First, while many authors have developed similar approaches to
estimating expected encounter rate, there has been little discussion of the confidence limits
on the estimates (but see Skellam, 1958). Thus it has not been the practice to compare
observations with predictions in a rigorous statistical way (Schülke & Kappeler, 2003,
provide an exception).
Second, the “random movement” assumption seems unrealistic in many biological
applications. This limitation was noted by Lotka (1924, p. 360) in what was apparently the
4
first discovery of the ideal gas model by a biologist: “…the type of motion presented by
living organisms [unlike that shown by ideal gas molecules]…can be regarded as
containing both a systematically directed and also a random element…mathematical
treatment…may appear to threaten formidable difficulties. It is to be hoped that this will
not altogether prevent its attack.”
Maxwell’s model of an ideal gas may fit the movements of dilute gas molecules
rather closely, but animals’ movements are almost certain to deviate from the assumptions
in some significant ways. In an ideal gas, molecules are randomly distributed, move long
distances in straight lines between encounters, and the size of the bottle that contains them
is relatively very large. By contrast, real animals may spend more time in some habitats
than others, deviate frequently from straight lines, and be restricted to partially
overlapping home ranges. Animal populations may further violate the assumptions by
showing anisotropy in the directions of movement, or by the speeds not following a
Maxwell-Boltzmann distribution. Biologists need to know whether such violations change
the rate of encounter substantially, and, if so, they might want to measure the relevant
characteristics of the paths of real animals so as to incorporate these parameters into more
sophisticated models.
The goals of this paper are to review the basic results of ideal gas models of animal
encounter, to examine the sensitivity of model predictions to some simple, but probably
common, deviations of real animals from model assumptions, and to illustrate how some
of these complications can be incorporated into the models. In addition, we correct a
number of errors in this literature, describe how to estimate confidence limits on encounter
rates, and illustrate our approaches by reanalyzing the data set used by Waser (1976) in his
early application of the ideal gas model to interactions among primate groups.
II. DERIVATION OF BASIC RESULTS
(1) The ideal gas model
In this section we state the standard analytic results from the ideal gas model. Strictly
speaking, the “ideal gas model” applies only to molecules moving in three dimensions
with a Maxwell-Boltzmann distribution of speeds and colliding inelastically, but we will
follow the biological literature in applying the phrase more generally. So assume initially
that individuals are moving independently at constant speed v, in a plane, in straight lines
oriented randomly with a uniform distribution. We focus on one individual and it is
helpful to consider the motion of all other individuals relative to that focal individual (Fig.
1); in this perspective, the focal individual appears stationary. By integrating over all
possible angles between the directions of movement, one can calculate (Fig. 1) that the
mean speed of the other individuals relative to the focal individual is
4v/.
(1)
We consider that two individuals encounter each other if their centres approach
within a detection distance D. We assume that contact does not change their speed or
direction. Each non-focal individual sweeps out a strip 2D wide (Fig. 1C) such that if the
focal individual lies within the strip an encounter occurs (remember that we are viewing
motion relative to the “stationary” focal individual). In time t the area swept out by each
individual is on average 2D 4v/ t. If the density of individuals is , an area A should
contain on average A individuals; these will have swept out strips of total area A
8Dvt/. The number of randomly positioned strips (= individuals) that cover the focal
individual follows a Poisson distribution, with mean (A8Dvt/)/A = 8Dvt/. Thus the
5
expected number of encounters after moving a distance x is 8Dx/, and an individual’s
encounter rate is
8Dv/.
(2)
The Poisson distribution of encounter number implies that the probability that an
individual has no encounters in time t is
exp(–8vDt/),
(3)
which can be subtracted from 1 to give the probability of at least one encounter. Since
encounters are random and independent, the time between encounters has an exponential
distribution, with the mean interval between initial contacts given by the reciprocal of the
encounter rate:
/(8Dv).
(4)
The mean distance between initial contacts (mean free path) is speed mean interval
= /(8D). If we are concerned with the total number of encounters in an area A, we can
multiply the individual encounter rate by the mean number of animals in the area (A)
divided by 2 (because each encounter is experienced by two individuals), giving the
expected number of encounters as
42ADvt/.
(5)
So far we have assumed that the focal individual is the same as any of the other
individuals that it encounters, but the ideal gas model has often been applied to encounters
between different classes (e.g. sexes) of individuals. In that case the encounter-rate
formula describes the number of encounters of an individual of one class if ρ refers to the
density of the other class. For the total number of interclass encounters within an area A,
simply replace the term 2 in (5) by ab, the product of the densities of each class (cf. the
law of mass action from chemistry). When the classes move at different speeds, the 4v/π
term for mean relative speed must also be modified, as discussed in the next section.
We have been predicting encounter rate assuming knowledge of density, detection
distance and speed. Sometimes we might measure encounter rate and want to estimate one
of the other variables, such as when seeking to estimate population density of pelagic
seabirds from the number seen from a ship. Similar strategies could be used as rules-ofthumb by non-human animals. For instance, Pratt (2005) shows how ants (Temnothorax
albipennis) deciding between two nest sites change their behaviour when a critical quorum
of scouts agree on one site, and this is apparently sensed by encounter rate with other ants.
When the number of encounters counted is large, it is sufficient simply to rearrange the
above formulae to achieve the estimation of density from observed encounter rate. But
when few encounters have been observed one should specify a prior distribution of
encounter rates and use a Bayesian approach to estimate the true encounter rate. A
uniform prior may not be the most appropriate; in the absence of further knowledge, for a
Poisson-distributed variable (like number of encounters) with mean μ there are theoretical
arguments in favour of a prior distribution with density proportional to μ–0.5 (Jeffreys,
1961, p. 186). Then, if we have observed n encounters over a time t, an unbiased estimate
of density is calculated from an encounter rate of (n + 0.5)/t, rather than n/t.
It is straightforward to extend these lines of argument to apply to three dimensions. In
that case the mean relative speed is
4v/3
(6)
(Fig. 2). The trajectory of each individual relative to the focal individual sweeps out a
bullet-shaped solid with the same volume as a cylinder of radius D and length 4v/3 t. So
its volume is D2 4vt/3, and the focal individual’s expected encounter rate is
4D2v/3.
(7)
Physicists have also derived equations for the rate at which molecules of an ideal gas
6
hit the walls of a container (rate = v/4 per unit area; Kauzmann, 1966, p. 179); the twodimensional analogue giving hits per unit length is
v/,
(8)
which has been applied by Lowen & Dunbar (1994) to analyse the defendability of
primate territories of different sizes.
(2) Associations versus encounters; counts and durations of associations
The ideal gas model predicts the number of initiations of encounters in time t. Thus when
testing predictions one should ignore encounters that have already started at the beginning
of the period of observation. Fewer observed encounters than predicted could occur
because individuals either avoid each other or persistently stay within contact (Struhsaker,
1981). If, instead of following an individual animal continuously, one made instantaneous
observations of it at n intervals, the expected number of “associations” (i.e. ongoing
occurrences of other individuals within a distance D, rather than instances of them
crossing into this region) is not given by the ideal gas model. Rather, the appropriate
formula in the two-dimensional case is
nD2.
(9)
This is simply the sum of the areas of n disks of radius D (think of a series of n disks
replacing each of the 2D-wide strips considered above: Fig. 3A). The formula assumes
that if on n successive sampling occasions the same pair of individuals are together, that
counts as n associations.
Unlike with the rate of encounters, the formula for the number of associations is
unaffected by speeds or by non-uniform distributions of directions (cf. Section III).
Consequently there are times when this approach may be more appropriate: for example,
when estimating seabird densities from counts from ships, it may be simpler and more
reliable to make periodic instantaneous counts of numbers occupying an area around the
ship than to count the number of new encounters over a period and apply the ideal gas
model (e.g. Tasker et al., 1984).
It is important not to confuse the two approaches. Schülke & Kappeler (2003) made
counts of observation times (every 5 min) when individuals were in contact, where two
consecutive observations of the same individuals in contact were counted as two
“encounters”. They should have compared their observations against nD2, not the
8Dvt/ from the ideal gas model. Likewise, at any single moment of observation, or
summed over n such observations, the predicted number of individuals observed between
distances D – δD and D + δD from a focal individual is directly proportional to D for
small δD (Fig. 3B). By contrast, the ideal gas model predicts that, over a long period of
observation, the number whose closest approach lies within this range of distances is a
constant independent of D (Fig. 3C; cf. Gursky, 2005; different results hold for three
dimensions).
At any instant the mean number of other individuals with which an individual is in
contact is D2: in comparison with Fig. 3A, each non-focal individual now is represented
by only one disc. If each individual’s position is independent of the positions of others,
how many discs cover the focal individual is Poisson distributed. Thus the probability (or
proportion of time) that an individual is in contact with at least one other is
1 – exp(–D2);
(10)
this corrects the equations in Whitesides (1989) and Holenweg et al. (1996).
The mean duration of an association is the mean number of contacts at one moment
(D2) divided by the rate at which encounters occur (given by the ideal gas model). So
with constant speeds in two dimensions, mean duration is D2/(8Dv/), which
7
simplifies to
2D/(8v).
(11)
This formula corrects that derived by Waser (1984) and used by others (e.g. Holenweg et
al., 1996); see Section V(1).
In three dimensions the mean number of other individuals with which an individual is
in contact at any moment is
(4/3)D3
(12)
(i.e. the volume of a sphere of radius D around each non-focal individual their density),
and the mean duration of an encounter is this divided by encounter rate, yielding
D/v.
(13)
III. ANALYTICALLY TRACTABLE COMPLICATIONS: DEVIATIONS FROM
MODEL ASSUMPTIONS ABOUT DETECTION DISTANCE, DENSITY, SPEED
AND DIRECTION
(1) Detection distance
If the detection distance varies over the environment (e.g. because visibility depends on
the vegetation), it is still valid to use the mean detection distance in the standard
encounter-rate formulae for two dimensions; but obviously the variance in encounter rate
will now be higher (the distribution is no longer Poisson). In the three-dimensional case,
because of the D2 term in those formulae, variation in detection distance about the mean
does affect mean encounter rate. Calculating the mean square of detection distance
sidesteps this complication.
(2) Density
If individuals are distributed patchily so that density is consistently higher in some
regions, the mean density can be used in the standard formulae, but again the variance in
encounter rate will be higher than for the Poisson distribution. Note that mean density
must be calculated by integrating over the time spent in each region, not over area. For
instance, suppose that there are two habitats of equal area, and that the population is three
times as dense in one habitat, so the densities are 0.5 and 1.5. If the focal individual
spends equal time in the two habitats, the mean density that it experiences is , so the
formula for encounter rate is unchanged. If instead it spends three times as long in the
high-density patch (such a tendency in the population might itself have generated the
higher density), the mean density that it experiences is 1.25, so the expected encounter
rate is increased by a factor of 1.25. If the density is higher in one habitat simply because
the animals move more slowly there, then the mean encounter rates within the two habitats
are the same (the product of density and speed is what matters). Gordon, Paul & Thorpe
(1993), working on the ant Lasius fuliginosus, explained a lower than expected increase in
encounter rate with density as resulting from a tendency to aggregate at low densities.
Without knowing how the aggregation is produced—reduced speed in the area of
aggregation or a reduced tendency to head away from it—one cannot predict whether a
tendency to aggregate will affect encounter rate, but there was evidence for the latter
mechanism in this example.
The ideal gas model has often been applied when individuals are confined to home
ranges. In cases where two individuals share the same home range, their expected
encounter rate is obtained by setting ρ = 1/(home range) (e.g. Schülke & Kappeler, 2003).
8
If a male home range is considered to enclose the complete home ranges of several
females, his expected encounter rate with them is obtained by setting ρ = (number of
females)/(male home range) (e.g. van Schaik & Dunbar, 1990). If home ranges of two
individuals overlap only partially, the situation should be analysed as in Section V(3); it is
wrong simply to set ρ = 1/(mean home range size) (cf. Gursky, 2005).
In some two-class systems an individual disappears from the system after one
encounter, as when a sperm sticks to an egg or when a predator meets a prey item and the
latter either is eaten or hides. If these disappearing items do not replenish, their density
will decrease with time, so that encounter rate also decreases. Suppose that the mean
relative speed of predator and prey is w, that the initial density of prey is ρp and that of
predators ρP. The probability of a prey item meeting no predator in time t is
exp(–2wρPDt),
(14)
so that in a large area A the number of encounters would be Aρp[1 – exp(–2wρPDt)]
distributed randomly over AρP predators. Hence the mean rate of encounter
[CORRECTION: NUMBER OF ENCOUNTERS] for a single predator is
(ρp/ρP)[1 – exp(–2wρPDt)]
(15)
and the probability of a predator encountering no prey is
exp(–(ρp/ρP)[1 – exp(–2wρPDt)]).
(16)
This last formula is particularly relevant in calculating the probability of an egg not having
been fertilized by a sperm, an approach developed in Vogel et al.’s (1982) “Don Ottavio”
model.
A related complication modelled by Nicolis, Theraulaz & Deneubourg (2005) occurs
when individuals that have met stay together immobile while they interact, but after some
time split up. Such is what often happens in ants. The temporary immobility obviously
lowers the encounter rate but Nicolis et al. (2005) further showed that with long
interaction periods individual encounter rate tends towards being proportional to density1/2
rather than the linear relationship of the ideal gas model.
(3) Speed
If individuals vary their speed or if speed varies among individuals, the ideal gas formulae
must be amended. In the two-dimensional case, for instance, the term 4v/ must be
replaced by a revised mean relative speed between the focal individual and other
individuals from the population. To calculate this it is necessary to know the distribution
of absolute speeds, not just their means, and then to integrate over the range of speeds of
the focal individual, over the range of speeds of the other individuals, and over all possible
differences in their directions. The formula thus becomes
u 0
pu (u ) p v (v)
v 0
π
u 2 v 2 - 2uv cos θ dθ dv du/π ,
(17)
θ 0
where pu(u) = probability of the focal individual having speed u, pv(v) = probability of a
non-focal individual having speed v, and = the angle between trajectories. The central
term is the cosine formula to calculate the length of the third side of a triangle if two sides
and the angle between them are known. The extension to three dimensions requires only
replacing the factor 1/π by 0.5 sin within the integral, so as to weight the average
according to the non-uniform distribution of (Fig. 2). The integration must be done
numerically except in some special cases of analytic tractability.
One tractable case, of widespread applicability, occurs when speeds in the population
follow a Maxwell-Boltzmann distribution. In the two-dimensional case its probability
density function is
9
2v
exp (v / r ) 2 ,
(18)
2
r
and in the three-dimensional case is
27 3 2 3
r v exp (v / r ) 2 ,
(19)
2π
2
where r = root-mean-square speed. These skewed distributions are reasonable
approximations to several observed speed distributions (Okubo, 1980; De Vita et al.,
1982; Waser, 1984; Mitani et al., 1991), and are the speed distributions expected if
animals are following a fine-scale random walk but speeds are measured on a coarser
scale. In a population of animals with a Maxwell-Boltzmann distribution of speeds, the
mean relative speed is
(20)
2 v ,
where v is the mean speed.
This result assumes that each individual’s speed is independently drawn from the
distribution. Suppose instead that the variation in speed was to a large extent an effect of
time of day (for instance, mites running around on a paving stone might all speed up as the
temperature increases in the afternoon). In the extreme case of all individuals travelling at
the same speed v at any particular time of day, the mean relative speed would be 4v/ at
any one time and thus 4 v / averaged over the whole period.
If we know the speed of the focal individual, we can refine our predictions. This is
relevant when calculating encounter rates between two classes of individual, such as
predators and prey, that move at different speeds. In many examples one class of
individual is stationary (e.g. eggs, when predicting their rate of fertilization by more
motile sperm), which simplifies the encounter-rate formula to 2 ρ D v for two dimensions,
or πρ D 2 v for three dimensions. If the two classes of individual travel consistently at two
different non-zero speeds, in three dimensions the encounter rate for an individual of one
class is
πD2(u2 + v2/3)/u
(21)
(e.g. Gerritsen & Strickler, 1977), where is the density of the other class and v is the
speed of whichever class is slower. Unfortunately in two dimensions the term for mean
relative speed when u v involves an elliptic integral. But if both predator and prey have
Maxwell-Boltzmann distributions of speeds, with mean speeds u and v , then the mean
relative speed is simply
2
2
(22)
u v .
[Skellam (1958) proves this for the two-dimensional case and it also holds in three
dimensions; see also Evans (1989) for a neat extension to include turbulence in the
media.]
Small inter-individual differences in speed have rather little effect on encounter rate.
For instance, Gursky (2005) was interested in contact rates between male and female
partners; males moved up to 1.49 times as far as their partners. In that case Gursky’s use
of the mean speed of partners is liable to underestimate encounter rate by a factor of only
0.95 if each individual’s speed was constant, or 0.98 assuming Maxwell-Boltzmann
distributions.
An increase in a predator’s speed increases its rate of encountering prey; the
relationship is approximately linear when the predator moves faster than the prey, but
becomes more curved at slower speeds (Fig. 4). This pattern has biologically interesting
consequences. For example, in the two-dimensional case, if prey all move with speed v,
and the predator moves at 2v, the mean relative speed is 2.13v; but if the predator moves at
10
v/2, the mean relative speed is only 1.06v. So a predator moving at half the speed of its
prey gains only a 6% increase in encounter rate over that had it remained motionless,
which, supposing some costs to moving, suggests advantages of a sit-and-wait tactic.
Gerritsen & Strickler (1977) developed such optimality arguments quantitatively by
introducing trade-offs relating speed to power requirements and detection distance.
Similarly in the optimality model of Dusenbery (2006) larger gametes benefit from their
faster speed (as well as from increased encounter distance and longevity), but fewer can be
produced.
Sometimes what matters is the number of encounters not per time, but per distance
moved. For instance an auk carrying fish from its feeding grounds to its nest is seeking to
reduce the total number of encounters with kleptoparasites over the course of the journey,
not the rate at which they occur. The number of encounters per distance moved by the
focal individual is obtained by dividing formulae for rate of encounter by the speed of the
focal individual. The number of encounters is then a decreasing function of its speed.
(4) Direction
The formulae also must be amended if the directions of movement are not uniformly
distributed, again because this affects the mean relative speed. This might be relevant for
animals on a migration or for a population of animals that all tend to avoid moving
directly upwind or against a water current. Particular attention has been paid to this issue
by researchers estimating densities of seabirds from sightings seen from ships or aircraft,
because, for instance, birds commuting between feeding sites and nesting colonies are
often predominantly headed in a common direction. Spear et al. (1992) proposed
recording the heading and speed of each bird seen so as to correct density estimates
individually. The estimated density is then the observed encounter rate divided by
2 D u i2 vi2 - u i vi cos θ i , [CORRECTION: 2 D u i2 vi2 - 2 u i vi cos θ i ] (23)
where ui and vi are the speeds of observer and the ith bird, and θi is the angle between their
directions of travel. Spear et al. (1992) tabulated how the effect of θ on expected
encounter rate depends on the ratio of speeds.
As another example, consider a stream of migrating wildebeest within which a male
is attempting to intercept receptive females. Imagine that within the migrating column
females are moving in parallel, that their speeds follow a two-dimensional MaxwellBoltzmann distribution, and that the male travels consistently at their mean speed v ; then
his mean speed relative to females is 0.39 v if moving with the flow, 1.38 v if
perpendicular to the flow, and 1.89 v if against the flow. So he encounters more females
by moving perpendicular to the flow than moving alternately with and against it (even
ignoring the tendency to re-encounter the same females in the latter case). Gerritsen
(1980a) applied similar arguments to the three-dimensional world of plankton, for
instance predicting that predators should move horizontally in response to vertical
migrations or “hop and sink” locomotion of their prey. Anderson, Gurarie & Zabel (2005)
developed the ideal gas model in terms of two components to movement, a directed
component and a random component, and applied it to model predation of migrating
salmon.
IV. TRACKS THAT ARE NOT STRAIGHT
Many of these results break down if individuals do not always travel in straight lines, for
instance if they take some sort of random walk. The reason is that bent paths, unlike
11
straight ones, can cross one another more than once. Moreover, random-walking
individuals sometimes backtrack, so one encounter with another individual is relatively
likely to be followed by a second encounter with the same individual before they have had
a chance to move far apart. The number of encounters among random-walking individuals
no longer follows a Poisson distribution, with more chance of many encounters or very
few. This invalidates formulae of the form exp(–kwρDt) for the probability of
encountering no items [e.g. as applied when a single encounter suffices to remove a prey
item or gamete from the population: Section III(2)]. Nor do the intervals between
encounters remain exponentially distributed.
Perhaps surprisingly, expected encounter rates are identical whether or not
individuals move in straight lines. However, for random-walking individuals many
encounters are recontacts between the same individuals, separated by only brief periods
apart. Observers might overlook brief periods apart and thus underestimate encounter rate.
For instance, Holenweg et al. (1996) introduced a bias into their test of the ideal gas
model by explicitly ignoring any change of association status lasting less than 30 min.
Often of more concern than the number of encounters is the number of different
individuals encountered. Fewer individuals will be encountered if paths are not straight,
but we know of no analytic formulae to make quantitative predictions. However, so long
as individuals are moving independently, the number of different individuals encountered
will follow a Poisson distribution. Also, doubling the density of individuals still doubles
the expected number of different individuals encountered.
It may help to explain these results geometrically. When deriving the formula for
encounter rate, we envisaged a strip swept out around the trajectory of an individual
relative to the focal individual (Fig. 1); the area of the strip gave the probability of a
randomly placed strip covering the focal individual and thus of an encounter. Changes in
direction of the trajectory do not affect the area of the strip if areas in which the strip
overlaps itself are counted twice (hence the number of encounters is still as predicted by
the gas model). But one must count areas of overlap only once to calculate the probability
of two individuals ever encountering each other. If individuals move with a regularly
turning trajectory (e.g. the sinusoidal movement of fish or the spiral trajectory of sperm:
Rosenthal & Hempel, 1970; Farley, 2002) it may be feasible to calculate this area or
volume analytically, at least if the items encountered are stationary [see Section V(6)].
Otherwise we must turn to simulation.
(1) Monte-Carlo simulations
Almost fifty years ago, Skellam (1958, p. 398) noted, “Whereas the expected number of
encounters does not appear to depend on the shapes of the paths, the variance of the
number of encounters does…in order to provide concrete support for the theoretical
formulae given earlier and the conjectures outlined above…[one approach is] to set up
laboratory experiments on Monte Carlo lines”. Skellam then proceeded to simulate animal
movements by moving coloured pins on a triangular lattice on graph paper. Being in a
position to set up “experiments on Monte Carlo lines” with enormously more power than
Skellam, we have written a computer simulation to quantify the consequences of animal
paths deviating from straight lines.
In these simulations we consider the track of each individual as a series of short
straight-line steps of constant duration. This is a flexible and widely used approach to
model meandering tracks, and simplifies the calculation of whether two individuals come
within range. Since we count only encounters involving one focal individual, and since
each simulation run covers a finite time, we need consider only those individuals that
12
would have time to meet the focal individual if they headed straight towards each other at
the specified maximum speed. Such individuals lie within a circular arena centred on the
focal individual; as the simulation progresses, this circle gets smaller and many
individuals become safe to ignore. The initial number of individuals in the arena is
generated from a Poisson distribution with mean given by density arena area. Their
starting positions and directions are randomly allocated assuming uniform distributions.
(2) Correlated random walks
A correlated random walk (successive steps have similar directions) fits most real animal
trajectories better than an uncorrelated random walk (e.g. Bergman, Schaefer & Luttich,
2000); one can imagine many reasons why animals tend to avoid backtracking.
Unsurprisingly, as the correlation in direction between steps increases, the encounter rate
more closely matches predictions based on straight-line trajectories.
To quantify this effect, we modelled the change in direction each step as fitting a von
Mises distribution, which is roughly the equivalent of a Normal distribution for angular
data (Fisher, 1993). The parameter alters the dispersion of turning angle about a mean
change of direction of 0, but in Fig. 5 we measure this dispersion in more intuitive terms
as the percentage of times the direction changes by less than 90 (“50% steps forward”
represents an uncorrelated random walk and “100% steps forward” a straight line). Fig.
5A and B show that as individuals follow more convoluted paths the distribution of the
number of encounters changes from a Poisson distribution to one with the same mean but
a greater variance. Fig. 5C shows the accompanying decrease in the number of different
individuals encountered. Although the broad patterns shown in Fig. 5 are consistent, it is
apparent that turning angle distribution, detection distance, step length and step number all
have interacting non-linear effects, so quantitative predictions are possible only using
simulations.
(3) Rebounds
So far we have assumed that individuals move independently of each other. Another
possibility is that when they contact they “rebound” so as to avoid each other. For
simplicity, and to utilize results from physics, our assumption here is that the separation
distance that counts as an encounter is the same as the distance at which rebounds occur.
After a collision, individuals are moving in different directions and cannot contact
each other again without rebounding off further individuals, so one might suppose that the
ideal gas model would continue to fit well. In fact, when rebounds occur the ideal gas
model underestimates encounter rate (Fig. 6). Physicists modelling dense gases have
concentrated on two effects (Chapman & Cowling, 1952, p. 274). One is that each particle
excludes others from its immediate surroundings, which has an effect of decreasing the
area available for movement and thus increasing effective density. The other effect is that
when two particles lie close together each shields the other from contact with a third
particle; later models have considered clusters of more than two particles screening each
other. In two dimensions, based on modelling clusters of ≤ 4 individuals, the predicted rate
of encounter 8Dv/ must be multiplied by a polynomial function of D2:
1 + 0.7820D2/2 (1 + 0.5322D2/2 (1 + 0.3336D2/2))
(24)
[see van Rensburg (1993) for higher-order “virial coefficients” and those for three as well
as two dimensions].
Even formulae based on this approach still underestimate encounter rates at very high
densities. At the extreme, individuals are packed in a crystalline fashion; the only
13
movement possible is a slight jostling back and forth from one neighbour to the other; as
density tends towards 2/(3D2) the collision rate tends to infinity. We produced estimates
of encounter rate at high densities using a refined version of our Monte Carlo simulation
in which individuals encountering each other immediately head directly away, but
otherwise move in straight lines at constant speed [Fig. 6A; cf. the simulation by Luding
(2001) in which speeds were Maxwell-Boltzmann distributed and collisions preserved
momentum and energy].
Fortunately, biological situations to which the ideal gas model has been applied are
rarely so extreme. Suppose that an encounter occurs when individuals are 100 m apart;
then it requires a density as high as 7.6 per km2 for the number of encounters to exceed
gas-model predictions by 10%. But clearly the ideal gas model would be inadequate at the
densities of ants in a nest, for instance. For example Gordon et al. (1993) studied
encounter rates in ants kept at densities up to 0.72 per cm2. Even were detection distance
as small as 0.6 cm (the length of these ants), the expected number of encounters is 40%
more than predicted by the ideal gas model.
However, ant encounters are often of interest because they enable the transfer of
information (Adler & Gordon, 1992). In this context what matters are contacts between
individuals that have not recently met one another. At high densities our simulations
demonstrate that the rate of novel contacts is decreased by each individual becoming
“boxed in” by neighbours (Fig. 6B).
(4) Other forms of non-independence
In nature individuals are often aggregated and even simple predators consequently adjust
their foraging paths to tend to remain within prey aggregations (area-restricted search).
Encounters with prey may stimulate adjustments of turning angles, speed or both. This
kind of non-independence invalidates the ideal gas model; for instance we no longer
expect a linear relationship between density and encounter rate (e.g. Travis & Palmer,
2005). The ideal gas model is most useful in such situations as a null model of
independent movement against which to compare the improvements in encounter rate
achieved by more sophisticated behaviours. The tendency for seabirds either to follow
ships or to turn away from them at distances greater than observers can identify them also
invalidates the ideal gas model, which creates problems in calculating densities from shipbased counts.
V. CORRECTIONS OF SOME ERRORS IN THE LITERATURE
The papers that have made the errors considered below have often also identified novel
situations to which the ideal gas model can be applied. It is thus worthwhile not only to
point out the mistakes, and to quantify how much they matter, but to show how they might
be corrected.
(1) Encounter duration
Waser (1984, 1987) calculated the expected duration of an encounter under the
assumptions of the ideal gas model. This provided a null hypothesis against which to
compare observed durations, to test whether avoidance or other interaction occurred. He
calculated the probability distribution of the relative speeds between random pairs of
particles and then integrated the product of probability and the reciprocal of relative speed.
What Waser neglected was that encounters occur more often when relative speeds are
14
high. As we saw in Section II(2), the correct calculation is to divide D2 by encounter
rate. In the case of Maxwell-Boltzmann distributions of speeds, the correct formula is
(25)
D/(2 u 2 v 2 ).
Waser’s formula overestimates this by a factor of /2 ~ 1.57, which neatly explains why
Waser’s (1984) predictions exceeded his observations by about this factor. Other papers
involving this error are Cords (1987), Mitani et al. (1991), Holenweg et al. (1996) and van
Schaik (1999), although the error is insufficient to have changed their qualitative
conclusions. Whitesides (1989) also followed Waser (1984) in using the wrong formula,
but relied more on the results of a simulation which were immune to the error (he
observed, but did not follow up, the discrepancy between the two methods of prediction).
Calculations of the proportion of time spent in contact with others that rely on
multiplying encounter frequency by Waser’s formula for encounter duration are also
overestimates by the same factor of 1.57 (Waser, 1987; Whitesides, 1989; Mitani et al.,
1991; Holenweg et al., 1996).
(2) D for spread-out groups
Waser (1976) realized that the ideal gas model might usefully be extended to encounters
between social groups, but unfortunately introduced an error into the literature. If the
diameter of each group is s, and the maximum detection distance between individuals is d,
peripheral members of the groups may detect each other when the group centres are d + s
apart; this corresponds to D in our equations (Fig. 7). However, instead of the 2(d + s) that
should have appeared in his encounter-rate equation, Waser (1976) wrote 2d + s. His later
papers applying the same technique (Waser, 1982, 1984, 1987) did not repeat the error,
but it has propagated in several doctoral theses (Bennett, 1984; Hill, 1991; Ham, 1994; see
also Schülke & Kappeler, 2003). Also, a review paper by Dunbar (2002) erroneously
implies that D should be taken as the distance between the edges of the groups (d),
ignoring group diameter.
Barrett & Lowen (1998) were the first to point out that the 2d + s term used by Waser
(1976) was wrong. Unfortunately they corrected the encounter-rate formula to
4v(s + d)/; this needs to be multiplied by a factor of 2.
(3) Partial overlap of home ranges
Barrett & Lowen (1998), following van Schaik, van Amerongen & Mouton (1985), noted
that the ideal gas approach might introduce errors when home range overlap between
neighbouring individuals (or, in this case, primate social groups) is not complete. They
therefore introduced a further factor x into the encounter-rate formula, which represents
the proportion of the group’s time spent in the part of their range shared with other groups.
They made the additional assumption that the group spent half as much time in the shared
part as expected on the basis of its area, because resources there were depleted by their
neighbours. However, this and similar analyses (e.g. Kinnaird & O’Brien, 2000) are
subject to several pitfalls.
There is a typographical error in the equation Barrett & Lowen (1998) printed to
make their intended adjustment. Their equation A7 should read x = (c/2)/(b + c/2) which
indeed yields the x value that they report. More fundamental a mistake is that multiplying
by x fails to take into account that the neighbours, like the focal group, will spend only
some of their time in the shared area. Kinnaird & O’Brien (2000), in a similar analysis,
apparently realised that it was the probability of groups being simultaneously in the shared
area that mattered, but did not allow for the increased density when this occurs.
15
For simplicity suppose that suitable habitat is distributed along a narrow strip 1 unit
wide and that a large number N of groups are spaced evenly along the strip (Fig. 8A). If
the strip is N units long, the density is one group per unit area, but we specify that each
group occupies a range of length and area h, resulting in each group sharing a strip of
width and area h – 1 with each neighbour. If density is not unity, the results derived below
still hold if h is defined as area of home range density (i.e. as actual home range area
divided by the area if home ranges were reduced to cover the region without overlapping).
The proportion of time that each group spends in each of its shared areas is (h – 1)/h
if it visits all parts of its range equally often. The probability that both the focal group and
the neighbour are in their shared area is (h – 1)2/h2. When that happens the density of the
neighbour within the shared area is one per h – 1 of area, = 1/(h – 1). Note that to calculate
encounter rate of a focal group requires local density to be calculated ignoring the focal
group itself (cf. Jolly et al., 1993). So the predicted rate of encounter between two
particular neighbours = [(h – 1)2/ h2] 1/(h – 1) = (h – 1)/h2 times that given by the usual
encounter-rate formula. (Unfortunately, even this correction factor is an underestimate,
especially with large D, because the centre of one group can be just outside the area of
shared use, but still lie within distance D of a neighbouring group whose centre is within
it.) Because each group has neighbours on either side, the rate must be doubled to give a
group’s overall rate of encounter:
(26)
2(h – 1)/h2.
This lies between 0 and 0.5 if no area is to be shared by more than two groups. If we
instead assume that groups visit shared resources half as frequently as resources in their
unshared home range centre, the correction factor becomes
(h – 1)/2
(27)
[whereas Barrett & Lowen’s (1998) correction factor of x corresponds to h – 1; it happens
that this error cancels out their earlier error pointed out in Section V(2)].
These formulae still hold if the area of overlap with each neighbour varies and
regardless of the number of neighbours with which each group shares its home range, but
only so long as every group still has an equal sized home range, with the same proportion
of it shared, and there are no areas shared by three or more individuals (Fig. 8B, C); these
seem reasonable rough approximations for many systems. Fig. 8D illustrates a
configuration in which some areas are shared by four groups; the correction factor,
assuming that each group utilises all parts of its range equally, is then increased to 4(h –
1)(2h – 1)/h2, which gets as high as 0.76 when h = 2.
Thus when home ranges do not overlap completely, expected rates of encounter will
be decreased by an amount that depends not only on how much home ranges overlap, and
on relative usage of those areas of overlap, but also on some aspects of overlap
configuration. Where the geometry of home range overlap is more complex, the
investigator will need information on the ranges of neighbours as well as that of the focal
group, and accurate null predictions will require simulation.
(4) Infanticide rate
The issue of how to correct for overlapping home ranges also arises in a model of
infanticide in the great apes (Harcourt & Greenberg, 2001), but it is exacerbated by further
complications that we now discuss. To predict the probability of infanticide, Harcourt &
Greenberg (2001) used the ideal gas model to calculate both the proportion of males that
would not meet a female when she was in oestrus, and the proportion that would meet her
later when she was nursing (only males that do not mate with a female but do encounter
her while nursing are assumed to kill infants). They took the product of these proportions
16
to calculate the probability of an infanticide: their formula is of the form exp(–2wDρtc)
(1 – exp(–2wDρtn)), where w is the mean relative speed, D is the distance at which a male
would detect a female, ρ is the male density, and tc and tn the periods of oestrus and
nursing respectively (here we have corrected the omission of a factor 2 in the encounterrate formula).
Unfortunately, their formula becomes inappropriate when more than a single male is
present in the female’s range. To understand why, consider the situation when the time
spent nursing is very long (as with the great apes), so that exp(–2wDρtn) ~ 0 and Harcourt
& Greenberg’s (2001) formula reduces to exp(–2wDρtc). This is the probability of no
encounters with any of the males during oestrus. But all it takes for an infanticide is for
one male not to have met the female: it is unnecessary for all not to have met her. The
greatest danger are males who rarely encounter the female, those whose home ranges
overlap the female’s range only a little. It should now be apparent why it becomes
necessary to consider range overlaps with each male individually. The probability that
each male commits infanticide can be calculated as Harcourt & Greenberg (2001) did, but
then one must calculate the product of the probabilities that each male does not do so to
give the overall probability of no infanticide. The correct formula for the probability of
infanticide is thus of the form
M
1–
[1 – exp(–2wDρitc) (1 – exp(–2wDρitn)].
(28)
i 1
Here M is the number of males with ranges overlapping a female’s home range and is
usually greater than obtained by multiplying male density by the area of female home
range [the apparent basis of Harcourt & Greenberg’s (2001) calculation]. Crucially, now ρi
is not male density ρ, but a quantity reflecting pairwise male-female range overlap as well
as density.
To illustrate this, we use Harcourt & Greenberg’s (2001) parameters for the Virunga
gorilla (Gorilla gorilla), under the fictional supposition that females were solitary. Home
range = 7.5 km2 and adult male density = 0.25 km–2, so h = 7.5 0.25 = 1.875. Let us
suppose that the topology of male home ranges is like Fig. 8C, and that the female’s home
range exactly matches the male home range marked with spots. For that home range ρi =
1/7.5 km–2. For each of the four other overlapping male home ranges ρi = ρ(h – 1)/(2h2) =
0.031 km–2 [derived similarly to formulae in Section V(3)]. The probability of infanticide
is then 1 – [1 – exp(–2 0.71 0.5 0.133 6.25)] [1 – exp(–
2 0.71 0.5 0.031 6.25)]4 = 1 – 0.45 0.134 = 0.9999. This is much higher than
Harcourt & Greenberg’s (2001) value. But note that the value would differ were the
female home range to be superimposed differently on the male home ranges or if the
males had a different configuration of home ranges; much more information has to be
specified for a reliable prediction, particularly about males that enter the female’s home
range only occasionally. One complication not covered by the formula is that, if during the
few days of her oestrous the female is in the corner of her home range where she is liable
to encounter one male, she is unlikely to wander into all the other corners where she could
meet the other males whose home ranges overlap hers only slightly. This, as well as the
typical tendency to spend disproportionately less time in non-exclusive peripheral parts of
the range, will tend to increase further the probability of infanticide.
(5) Mating benefits of an increased home range
There is another problem with Harcourt & Greenberg’s (2001) study, and indeed with our
corrected formula. It arises also in several papers using the ideal gas model to examine
whether males, rather than being monogamous, would profit from defending an enlarged
17
home range containing the home ranges of several females (van Schaik & Dunbar, 1990;
Dunbar, 1995, 2000). What is critical in these latter papers is the number of different
fertile females that a polygamous male would encounter. This was estimated as n(1 – e–m),
where n is the number of females in the male’s territory and m is the expected number of
encounters with each female within three fertile periods, as calculated from the ideal gas
model. However, with non-straight trajectories the number of encounters is not Poisson
distributed; therefore e–m (from the Poisson distribution) accurately estimates how often no
encounters occur only if m equals the mean number of different individuals encountered,
not the mean number of encounters. The predicted advantage in being polygamous is
potentially much diminished.
To gauge the magnitude of the error, we consider a correlated random walk consisting
of straight-line steps each lasting 30 min and turning angles described by a von Mises
distribution with κ = 0.69, which is based on observed movements of grey-cheeked
mangabey groups analysed in Section VII. We combine this with one set of van Schaik &
Dunbar’s (1990) figures for the gibbon Hylobates lar: speed = 0.108 km h–1, detection
distance = 25 m, density of a particular female within the enlarged male territory = 0.75
km–2, and a female is fertile for three consecutive days ( = 72 steps of our random walk).
Simulations predict that the male would encounter a mean of 1.00 different females,
compared with 1.52 total encounters, so the probability of not meeting a particular female
in his territory is (1 – e–1.00) = 0.63. If, following van Schaik & Dunbar (1990), we allow
three cycles for a successful copulation, the probability of fertilizing a particular female is
1 – 0.633 = 0.75. If a mean of 4.5 females remained within a male’s territory, the expected
number of females that he fertilizes is 0.75 4.5 = 3.3 (assuming that, if females are
territorial, they do not all become fertile at the same time, so that encountering one fertile
female does not mean that the male is less likely to encounter another). The original
prediction was 4.5 (all females mated); the corrected estimate remains greater than the 1
expected under monogamy, but remember that we used an estimate for the convolution of
the path based on a quite different species.
Note also that van Schaik & Dunbar (1990), as well as Dunbar (1988, p. 309; 1995;
2000), apply the version of the ideal gas model appropriate for interactions between
moving and stationary individuals to situations where all individuals are moving, thus
underestimating encounter rate by a factor of at least 4/π = 1.27.
(6) Spiral trajectories of sperm
Errors in application of the ideal gas model are not confined to primatology; here is an
example from work on fertilization kinetics. Usually sperm are so small relative to the egg
that the critical distance D is set simply as the radius of the egg re. The term πD2 term in
the three-dimensional encounter-rate formula is thus πre2, which equals the cross-sectional
area of the egg. However, Farley (2002) pointed out that sperm trajectories are typically
helical, so that it is reasonable to consider the sperm as having a diameter that of the
outside of the helix (with a correspondingly slower forward speed). In that case D should
be set as the radius of the egg plus the radius of the helix (rs), and πD2 becomes π(re + rs)2.
Instead Farley used π(re2 + rs2), believing mistakenly that adding the cross-sectional areas
of the egg and the helix was the appropriate procedure; his predictions of initial encounter
rate should be 1.5–1.8 higher.
The corrected predictions turn out to exceed observed encounter rates. One potential
reason is that the revised formula calculates the volume of a cylinder of radius re + rs,
whereas the real probability of encounter is given by the volume of a helix fitting exactly
within this cylinder but not fully filling it; the helix is constructed by dragging a sphere of
18
radius re along the sperm’s trajectory relative to the egg. Because the egg is stationary, this
relative trajectory is a regular helix and thus the volume is feasible to calculate. Sperm
stop moving after encountering an egg, so the relevant volume should be calculated by
counting volumes of overlap between consecutive turns of the helix once only. With the
parameter values of this example, the volume turns out to be exceedingly close to that of
the enclosing cylinder. So the revised version of Farley’s (2002) calculation is an excellent
approximation and other explanations must be sought for the disagreement with
experiment. If we had assumed that the sperm’s track was straight instead of helical, the
volume, and the predicted rate of encounter, would have been a factor of 3.3–4.7 too high.
We recommend trying an analogous treatment of the wide sinusoidal sweeping
movements of the heads of some predatory fish (Rosenthal & Hempel, 1970).
(7) Predation in the plankton
An example from this latter field concerns rates at which jellyfish (Aurelia aurita) capture
herring larvae (Clupea harengus). To replace the assumption that the herring are points or
spheres, Bailey & Batty (1983) supposed that they were lines of length L. The detection
distance between the centre of the jellyfish (taken as a sphere of radius rj) and the centre of
the herring then depends on the orientation of the herring relative to its direction of
movement towards the jellyfish. Bailey & Batty (1983) used an incorrect formula for this,
but more fundamentally they overlooked that—because the absolute velocity of the
herring is oriented along its body, and the absolute velocity affects the orientation of the
relative velocity—the herring tends to be oriented close to the direction of the relative
velocity. The correct formula for encounter rate is
π
ρ
u
2
v 2 2uv cos θ πrj2 rj Lv sin θ sin θ dθ / 2 ,
(29)
0
where is the angle between the absolute velocities of herring and jellyfish, and u and v
are their respective speeds (see Fig. 9 for derivation). Over the size range of jellyfish
considered, Bailey & Batty’s (1983) formula overestimates the correct predictions of
encounter rate by factors of 1.17–1.55; the corrected predictions worsen the fit to their
observed rates of predation in the laboratory. Unfortunately the incorrect formula has been
used by a number of other authors (e.g. Cowan & Houde, 1992; Letcher et al., 1996).
Incidentally, it is easy to make the model a little more realistic by replacing the line with a
cylinder of radius rh and length L capped with hemispheres (so total length = L + 2rh).
Then simply replace rj in (29) with rj + rh.
Several papers copying the erroneous formula use it to assess how predation affects
the size-distribution of fish larvae. Paradis, Pépin & Pepin (1999) and Paradis & Pepin
(2001) considered a cohort of larvae that as they grow are predated by predators of a range
of sizes. For each predator species they apparently used its modal size to calculate the
probability of N encounters with a prey of each size. They then randomly selected one
predator size from a truncated normal distribution (± c. 2 S.D.) about this mode, and the
ratio of prey to predator sizes was used to calculate the probability that each encounter led
to predation. The potential problem is with the use of the mean predator size to calculate
the number of encounters. Predators of above-average size are disproportionately likely to
encounter a prey item, and their encounter rate is affected by prey size to a different
degree. We assessed how important this was by using the parameter values of Paradis et
al. (1999) but considering predation only by crustaceans (115 per m3). Our recalculations
dividing the predator population into multiple size classes increase mortality over the first
day from 2.8% to 5.0%, but the difference diminishes as the prey grow larger than their
predators, so mortality after 30 days increases only from 60% to 75%. (We failed to
19
resolve the discrepancy between the 60% figure and the considerably lower mortality
levels calculated by Paradis et al., 1999) Reassuringly, the recalculations turn out scarcely
to alter the size distribution of survivors. Repeating the recalculations using our corrected
version of Bailey & Batty’s (1983) formula decreases mortality to 48% (3.8% on the first
day), and results in a larger mean size of survivors, but by only a third of a daily growth
increment.
(8) Estimating seabird densities from ship- and aircraft-based counts
Gaston & Smith (1984) and Gaston, Collins & Diamond (1987) sought to quantify the
error in estimates of seabird density derived from counts of individuals flying within a
prespecified distance of a ship or aircraft. In particular they examined the effect of the
direction in which birds were flying relative to the path of the observer. The effect has two
components, both of which Gaston & Smith (1984) and Gaston et al. (1987)
miscalculated. One component is the increased relative speed when observer and birds are
travelling in opposite directions [see Section III(4)]. An additional effect occurs if the area
monitored is not circular. Gaston & Smith (1984) considered a rectangular area aligned
with the observer’s absolute direction of movement (Fig. 10); let the width be 2D and the
length a. Consider the bird as the focal individual and the ship as laying down a strip
along its trajectory relative to the focal individual. The width of this strip depends on the
direction between the relative trajectory and the absolute trajectory, which depends on the
angle between the absolute trajectories of ship and bird (θ), and on their speeds (u and v
respectively):
(4 D 2 a 2 ) sin tan 1 (2 D / a ) tan 1 ((v sin θ ) /(u v cos θ ) .
(30)
This must be multiplied by the relative speed to generate the expected rate of encounter
divided by bird density.
Different shapes of area monitored require different calculations. The algebra for the
case of an ellipse aligned along the direction of movement turns out particularly neatly:
the rate of encounter is
(31)
2ρ (av sin θ ) 2 b 2 (u v cos θ ) 2
(where u, v and θ are as above, ρ is target density, the ellipse has length 2a along the
direction of movement and width 2b perpendicular to it). An ellipse seems an appropriate
way to model the greater sensitivity of many organisms in a forward direction (the algebra
is unchanged if the observer is positioned behind the ellipse centre). However, if the field
of sensitivity is not circular the rate of encounter increases whenever the observer changes
direction (or even turns its head), so the method may not be readily applicable. See
Skellam (1958) for a discussion of this point.
One relevant situation largely avoiding this rotation issue is of a primatologist or
ornithologist conducting a census along a path through thick vegetation. The area
effectively monitored may then often be keyhole shaped: a long narrow rectangular area of
clear visibility along the path in front, combined with a circular area centred on the
observer. One approximation to modelling this would be to calculate the number of
encounters at the periphery of the circular area (by the standard formula) and then to add
the number of encounters expected for the rectangular area sticking out beyond this circle
(calculated by the formula developed in Fig. 10). Animals first crossing the rectangular
area and then the circular one would be counted twice by such a calculation, but they
might also often be in the field (except for those that remained in view on the path).
If the path were narrow and the animals small, one might reasonably consider the
region monitored as one dimensional, setting D to 0 in the formula derived from Fig. 10.
20
The expected number of animals crossing the path ahead within a distance a of the
observer is then
π
ρ avt sin θ dθ / π = 2ρavt/π.
(32)
0
Note that this is independent of the speed of the observer. The same formula is also
applicable to counts of seabirds crossing the bows of a ship within a specified distance in
front (assuming a uniform distribution of their directions of flight, and that the ship is not
turning).
Another procedure when estimating seabird densities is to count only birds on one
side of the ship, so that the area monitored is half a circle. In that case it makes the
calculations much simpler to count only birds entering the area across the curved side (i.e.
ignoring those entering the semicircle from the other side of the ship); the expected
number is then simply half that if one were looking round the full 360° (assuming a
uniform distribution of directions, but otherwise the side of the ship monitored could be
alternated).
VI. HOW TO TEST THE PREDICTIONS?
In most of the earlier work that tested whether the ideal gas model adequately described
animal behaviour, conclusions were based only on whether the predicted and observed
number of contacts seemed to differ substantially. More rigorous statistical tests might be
based on three different approaches.
One approach relies on the assumption that the number of encounters is Poisson
distributed, which we know not to be exactly true unless the trajectories are straight, but
which is a reasonable approximation in particular cases. Non-straight trajectories increase
the variance in the number of encounters, so assuming the Poisson distribution will reject
the null hypothesis too often. When the predicted number of encounters is small,
confidence limits are calculated from the usual formula for a Poisson distribution:
j i
j e
probability of i encounters =
.
(33)
j!
j 0
When is larger it will be necessary to approximate this by a normal distribution with
variance .
The second approach, originally suggested by Skellam (1958), is to make several
independent observations of encounter rate and then make use of the variance in these
observations, allowing a t-test to compare the mean of these observations with the
prediction. The distribution of the observations can be more skewed than a Poisson
distribution (Fig. 5A), but, as a result of the central-limit theorem, the distribution of the
mean of m observations should usually be well approximated by a t-distribution. A squareroot transformation will tend to improve the approximation, especially when the number
of encounters is small. The advantage of the t-test approach is that we need not know
whether the individuals are taking straight or convoluted paths—the predicted number of
encounters is unaffected, and the test does not rely on assumptions of a Poisson
distribution. Note, however, that the assumption of independent observations is violated if
we use consecutive periods of observation on the same animal or use simultaneous
observations of neighbouring animals. The latter is a problem with the statistical test of
Mitani et al. (1991), which treats the encounter rates of each of a group of neighbouring
orang-utan with the others as independent, but not with that of Schülke & Kappeler
(2003), which is based on encounter rates only within different male-female pairs. Note
21
also that the use of a χ2 test to compare observed and predicted number of encounters
(Jolly et al., 1993; Gursky, 2005) is inappropriate.
Simulation provides a third way to test predictions about encounter rate. We advocate
building tailored simulation models based on the movement patterns of the species of
interest. The proportion of simulation runs in which the number of encounters is equal to
or more extreme than the number observed provides directly a one-tailed p-value, which
can be doubled to provide a two-tailed value (Manly, 1997, p. 72). Or the quantiles of the
distribution of encounter number may be used as confidence intervals.
A simulation is easier to write and faster to execute if we assume that trajectories are
composed of straight-line steps and that the changes in direction occur simultaneously in
all individuals. This may affect the best way to gather data: for instance Waser’s (1976)
recording of the position of mangabey groups every half hour produces data that can be
put directly into such a model. By contrast, Barrett & Lowen (1998) recorded time and
position whenever the group moved more than 10 m, which gives a more accurate
description of the trajectory but is more complicated to incorporate in a simulation.
The simulation will select step lengths and turning angles at random from the
observed distributions; there is little computational cost to incorporating any observed
correlations between these variables, or any autocorrelation. Such modifications will not
alter the mean number of encounters from that predicted by the ideal gas model, but will
broaden the confidence limits. Hence, if our first approach of relying on the Poisson
distribution already fails to reject the null hypothesis, there would be no need to develop
such a simulation. However, at some greater cost in computational time one might
incorporate further details of the biology as part of the null hypothesis, and these may
affect predicted encounter rate.
Restricted home ranges are one important violation of the assumptions of the ideal
gas model. The simple analytic model in Section V(3) showed that the extent of home
range overlap affects encounter rate considerably. Two ways to set up home ranges in a
simulation are to restrict movements to an area defined by the observed home ranges or to
engineer a biased random walk in which the directions of movement are increasingly
likely to lead back towards a central point the further away the animals drift. Either could
be biologically reasonable, and both could be incorporated in the same model. The choice
does matter because it affects whether peripheral zones of a home range, which are likely
places for encounters, are less frequently visited than elsewhere. Other possibilities that
could be readily incorporated into a simulation model are a tendency to visit a shared area
less frequently because of the greater competition for food there, or a tendency to spend
more time patrolling the edge of an exclusive territory.
The more such biological details that are incorporated into a model the more
confidence one might have that any disagreement of observations and prediction is due to
interactions with other individuals, rather than to other unconsidered complexities of the
movement patterns. One might even then incorporate hypothesised responses to nearby
individuals to test whether these are sufficient to explain the discrepancies. However, it is
unlikely that the information for a full model will be available for anything but a
laboratory system. Such factors as difficulty of moving through scrub may dominate an
animal’s choice of path, yet be hard to incorporate into a model.
One appealing alternative to simulating tracks is to use real tracks made by animals in
the same locality. The idea is that animals following two sets of tracks made at different
times cannot have responded to each other, and thus the number of fictional encounters
between them provides a prediction under the null hypothesis of no interaction. This is a
randomization procedure (Manly, 1997). The two sets of tracks might even be from the
same animal at different times, an approach taken below in our worked example. The time
22
separation should be sufficient that the animal could have moved anywhere else in its
range within that time, and ideally should be long enough to preclude indirect interactions
through scent marking or local food depletion. In species in which individuals set up
stable territories that largely avoid the territories of others, the procedure cannot exclude
these long-term interactions from the null hypothesis, unless the habitat is sufficiently
uniform that it would be reasonable to shift tracks in space as well as time. However, it
may often be a strength of the approach to be able to test whether unexpectedly more or
fewer encounters occur between particular neighbours given their observed home ranges.
This is the approach of Doncaster (1990) who compared the observed separation distances
of neighbouring foxes with those generated by taking distances between random points
(i.e. times) on their two tracks.
VII. AN EXAMPLE: GREY-CHEEKED MANGABEY INTERGROUP
ENCOUNTERS REANALYSED
We have discussed a variety of ways to predict the number of encounters and to test
whether the observed number differs statistically from the prediction. Here we illustrate
some of these approaches using data on grey-cheeked mangabeys (Cercocebus albigena)
in Kibale Forest, Uganda (Waser, 1976). The location of the centre of a focal group was
recorded every 30 min, typically for 9–12 h each day, for a block of 10 days each month,
over a total of 12 months. We use Waser’s original estimates of a density of 0.25 groups
km–2, and a group diameter of 90 m. Waser observed four encounters within 200 m, and
11 within 500 m. These encounters, rather than having already started when observations
began, all started during the course of the day.
The total period of observation was 1130.5 h, during which time the group moved
133.5 km. The simplest version of the ideal gas model thus predicts
(4/π) × 0.25 × 133.5 × 2(0.2 + 0.09) = 24.6 encounters within 200 m and 50.1 encounters
within 500 m (Table 1).
However, the distance moved in each half-hour (i.e. speed) varied considerably. If we
assume a Maxwell-Boltzmann distribution of speeds, the predictions increase by a factor
of 1.11. But the Maxwell-Boltzmann distribution is not a particularly close fit to the
observed speed distribution. Instead a x0.3 transformation produces a reasonable fit to a
normal distribution. If we truncate this at the observed minimum and maximum and
integrate numerically, the mean relative speed increases further from 1.41 to 1.53 times
the mean speed, predicting correspondingly more encounters (Table 1).
An alternative to fitting a distribution and integrating is systematically to pair all
observed step lengths (including with themselves), and calculate the mean relative speed
for each pair by integrating (numerically or using elliptic functions) over all possible
angles between them:
π
xi2 x 2j 2 xi x j cos θ dθ / π .
(34)
0
The mean over all possible pairs is 1.54 times the mean speed, very close to the figure
using the transformed normal distribution.
A closer look at the distribution of step lengths shows that there is no systematic
variation during most of the day, but that the mean is lower in the last two evening
periods, because sometimes the group had already settled down for the night. We repeated
the systematic pairing of all step lengths, but this time pairing step lengths recorded at
each of the latest two times of day only with others recorded at the same time. In this
23
example this refinement made very little difference.
These calculations have all assumed that groups do not have home ranges. In fact the
mangabey group had a roughly rectangular home range, bounded on its east and west sides
by unsuitable habitat. Very roughly, a central area of 0.3 the length of the rectangle was
used exclusively by this group and each of the ends was overlapped by the home range of
another group. If we assume that usage over the home range was even and that the areas of
overlap occupied a similar proportion of each neighbour’s home range, then we can use
formula (27) to revise our prediction of encounter rate: 2(h – 1)/h2, where h = 1/(0.3 +
0.35), implies a correction factor of 0.46, considerably reducing the predicted number of
encounters (Table 1). A better prediction would incorporate the observed proportions of
time that the focal group spent in the two areas of overlap: they were 0.30 and 0.37. If we
continue to assume that each neighbouring group spent the same proportion of its time in
each area of overlap as did the focal group, the probability of being in an area of overlap at
the same time as that neighbour is now (0.302 + 0.372), compared with the earlier
calculation of 2 0.352. This slightly reduces the correction factor to 0.42.
Now we consider how to test whether the observed number of encounters differs
significantly from these predictions, which requires estimating the variance in the number
of encounters. We first illustrate the second of our three proposed approaches, using the
observed variation from month to month. There were too few encounters within 200 m to
provide a reliable estimate of the variance. With the 500 m detection distance, there were
5 months of 0 encounters, 4 months of 1 encounter, 2 of 2, and 1 of 3. Very different
distances were moved in some months, in large part due to differing lengths of
observation; therefore we computed mean rate of encounter each month. A square-root
transformation served to make the distribution more normally distributed. The standard
error of the mean was multiplied by the 95% confidence limit for the t11 distribution,
added and subtracted from the mean, and back-transformed to generate confidence limits
of 2–18 encounters over the 2261 half-hours of observation. So predictions outside this
range should be judged as differing significantly from observations.
The other approaches to statistical testing estimate the variance from the model rather
than the data. Assuming a Poisson distribution of encounters and a mean number of
encounters of 29.7 or 60.5, yields 95% confidence limits of 21–38 and 48–73 for
encounters at 200 and 500 m respectively (Table 1).
Unfortunately it is not appropriate to assume a Poisson distribution for the model
involving overlap of home ranges; if an animal is in an exclusive part of its home range,
for instance, it is likely to remain so for some time, increasing the variance in encounter
rate. In any case the Poisson distribution will underestimate the variance because
trajectories are not straight. Our third approach, based on simulating a more detailed
model of movements allows a more valid estimate of variance for non-straight trajectories
(but here we assume panmixis rather than modelling the overlapping home ranges). We
chose to sample randomly from the observed step lengths and changes of angle (rather
than sampling from fitted distributions). There was no need to incorporate a correlation
between these because it scarcely existed in our data. Our model also matched the
structure of the observations. Observations were made in blocks of typically 10 days, and
since blocks were separated by typically 20 days, our model took a new random starting
configuration each block. Otherwise, the initial location each day was taken as the same as
the night before, but the initial direction was chosen at random. Our model matched the
pattern of observation periods exactly, so that some blocks had more days than others.
Similarly, although the random walk continued for 23 steps each day, on some days there
were periods when data had not been recorded and accordingly any encounters in our
model at the corresponding times were ignored. Using this approach, the 95% confidence
24
limits for the number of encounters were much wider than the Poisson-generated limits
(Table 1).
The simulation also allowed us to investigate one source of observer bias. Two
groups that encounter each other may shortly afterwards re-encounter each other, but
observers might easily not realise that there was a period when the groups were apart. We
could set our simulation to ignore re-encounters starting within some period from the end
of earlier encounters between the same groups. Setting this period to 1 h reduced the
predicted number of 200 m encounters from 29.7 to 19.7; setting it to 4 h reduced it
further to 14.1 (Table 1).
In theory the random walk could be further modified to make it more realistic. One
modification that had almost no effect on the confidence limits was to incorporate the
observed autocorrelation between successive step lengths and turning angles. However,
that version was based on a first-order Markov process (i.e. only the immediately
preceding step influences the next; Root & Kareiva, 1984), not on consistencies in step
lengths and directionality persisting over several hours. In reality the group tended to
spend several days in succession feeding on the same fruiting tree in the morning and
exploring away from it later in the day, whereas on other days it moved more directionally
to locate a new source of fruit (Waser, 1977a). Simulating this would require much
information on fruiting patterns and foraging behaviour. An easier modification to our
simulation would be to simulate just the three neighbouring groups observed, restricting
each to their observed home ranges.
We also tried the approach of counting intercepts with the group’s own trajectory in
other months (so separated by ≥ 14 days, enough for the group to traverse its home range).
This null model’s strength is incorporating any consistent tendency to avoid regions,
perhaps of unsuitable habitat, within a home range. For the most southerly 0.35 of the
home range, we selected steps starting within this region, then systematically paired all
such steps from different months, and found that 0.0239 of these pairings would have
involved a 200 m encounter. The focal group was in that region 0.30 of the time, and so
we assume (as above) that it and its neighbour were there simultaneously 0.302 of the
time. A total of 2261 half-hour steps were observed, so we predict 2261 0.0239 0.302
encounters in the southern area of overlap, plus 2261 0.0358 0.372 in the northern area
of overlap, summing to 16.0 encounters. This figure and the corresponding prediction for
encounters within 500 m are comparable with the predictions from the ideal gas model
when range overlap is incorporated (Table 1). Apparently any tendency to avoid or prefer
particular regions within the home range is at the wrong scale, or otherwise insufficient, to
have much effect on the rate of encounter at these distances.
VIII. DISCUSSION
Our analyses indicate that when individuals do not interact with each other the ideal gas
model makes rather robust predictions about the number of contacts, but less robust
predictions about the variance of this number or the number of different individuals
contacted. In addition, the model will tend to overestimate the number of encounters
scored if multiple contacts within a short interval are overlooked. The model must be
modified when animals are restricted to overlapping home ranges, but encounter rates are
still predictable if appropriate data on home-range geometry and usage have been
collected. When individuals move away from each other on contact, the model
underestimates encounter rate, but not by much unless density is very high.
Reanalysis of the data set used in an early application of the ideal gas model to the
25
investigation of grey-cheeked mangabey intergroup encounters showed that the results
were insensitive to a variety of refinements (more realistic speed distributions, and
allowing correlation among steps in the underlying random walk model of movement). On
the other hand, both a more realistic incorporation of home-range overlap and filtering out
multiple encounters over a short interval lowered the predictions considerably, and would
do so further if these modifications were combined. Also, our simulations indicate that the
confidence limits around predicted encounter rates are wide, making it difficult to
demonstrate statistically significant differences from observations. The confidence limits
would have been even wider had we allowed for some uncertainty in the parameters used,
for instance in density or in speed of other groups.
This fundamental problem of wide confidence limits may mean that counting
encounters is not the best way to test whether animals avoid or are attracted to each other;
for instance, Waser (1976, 1977b) relied not just on the ideal gas model, but also observed
how mangabeys responded to playbacks of neighbours’ calls. For other purposes the ideal
gas model remains useful in providing estimates of encounter rate in a wide variety of
biological interactions. A particular benefit of this analytic approach is that biologists are
immediately alerted to the expected form of the relationship between speed, density or
detection distance and the quantity of interest (not only encounter rate, but encounter
duration, the intervals between encounters, etc.). The form of these relationships may
indeed be what matters most when the ideal gas model is used as a component of larger
models. However, simulations provide a way to add biologically interesting complications
to the ideal gas model.
This may be a ripe time to reconsider the ideal gas model and its variants, because the
diminishing price and size of tracking technologies (e.g. GPS data loggers and satellite
tracking: Kenward et al., 2002) now facilitate near-continuous recording of position
however widely the animals range. These technologies not only provide parameters for the
model (speed distributions, home range overlaps, etc.) but, because multiple animals can
be tracked simultaneously, they also provide a means to monitor encounters themselves.
In this latter context, another recent technological advance is data loggers that record the
presence of other tagged individuals in close proximity (Weihong, White & Clout, 2005).
In reviewing the many diverse applications of the ideal gas model, we were surprised
by several omissions. For example, we failed to find its application among studies of
commercial fishing. The explanation appears to be that the models used to predict fish
capture rates are more sophisticated than the ideal gas model, taking into account the
avoidance behaviour of the fish once they notice the approaching net, and the consequent
greater propensity for fish further from the centre of the net to escape (Barkley, 1964;
Laval, 1974).
More difficult to explain is that we found no mention of the ideal gas model in
epidemiology, despite the acknowledged importance of contact rate in determining the
spread and prevalence of diseases and parasites (e.g. Gompper & Wright, 2005).
Epidemiology goes as far as borrowing the law of mass action from chemistry (that the
rate of contact between two types of individuals is proportional to the product of their
densities; McCallum, Barlow & Hone, 2001; Begon et al., 2002). This result is a
prediction of the ideal gas model; what is extra in the ideal gas model itself, besides the
value of the multiplicative constant, is the dependence of rate of contact on individual
speeds and on the critical distance between individuals that defines a “contact”. As an
example of the possible applicability of the ideal gas model, consider the typical finding in
epidemiology that infection rate shows a non-linear effect of density (Fenton et al., 2002).
In some cases this might be explained by looking at how individual behaviour depends on
density (Dwyer & Elkinton, 1993). For instance, it might be that density affects speed of
26
movement or the degree of overlap of home ranges; application of the ideal gas model
would then allow a revision of the relationship between contact or infection rate and
density. The same sorts of analysis might be used to improve models of information
spread in social animals (cf. Adler & Gordon, 1992).
The story described in the Introduction is of biologists repeatedly and independently
turning to physics to find an equation for collision frequencies between moving particles.
An interesting twist to this tale is that James Clerk Maxwell, the originator of the equation
for collision rate in an ideal gas, was himself inspired by research on populations of
animals (humans). In the social sciences, the statistical approach had reliably described
such phenomena as crime rates and ages of marriage in large populations even though
behaviour of individual humans was obviously unpredictable, which stimulated both
Maxwell and Boltzmann to take the same statistical approach with molecules (Gigerenzer
et al., 1989, p. 62).
IX. CONCLUSIONS
(1) The ideal gas model applies to non-interacting particles moving in randomly oriented
straight lines. It yields a simple equation linking the number of times particles come
within a specified distance of one another to their speed and density. The formula is
straightforward to extend to encounters between two classes of individual differing in their
densities and speeds.
(2) This encounter-rate formula is quite different to that predicting how many
neighbouring particles lie within the specified distance at a single moment of observation,
or summed over a series of such observations. Combining the two different formulae
allows calculation of mean duration of an encounter, which published formulae
overestimate.
(3) The model has been widely applied in biology to analyse rates of encounters between
individuals (e.g. males and females, sperm and eggs, predators and prey, human observers
and the animals that they are counting) and between groups of individuals. It has been
used both as a null model to detect avoidance or association, and as a component of other
models that analyse rates of events dependent on encounters.
(4) There are different versions of the formula depending on whether movement is in two
or three dimensions and on whether speeds are constant or follow Maxwell-Boltzmann
distributions. Versions for other speed distributions usually require numerical integration
to derive the multiplicative constant. This is also the case if directions of movement are
anisotropic.
(5) Local variation in density has a different effect on the predictions depending on
whether it is due to individuals avoiding low-density regions or moving faster within
them. Simple extensions of the model cover the situation when density decreases
progressively as a result of previous encounters (e.g. predators and prey).
(6) When individuals are restricted to home ranges that overlap, a simple correction factor
may often be adequate, but this differs from corrections given in the literature.
(7) If movement is not in straight lines, the formulae still hold for mean number of
encounters, but this is no longer Poisson distributed (invalidating several published
calculations of whether any encounter occurs in a specified time). Many encounters will
now be re-encounters with the same individual, which field observations might well not
distinguish. We know of no analytic formulae for predicting the number of different
individuals encountered, although this will be Poisson distributed and proportional to
density.
27
(8) If individuals back away from one another following contact, the encounter-rate
formulae produce serious underestimates only at high densities, such as might be
experienced by ants in a nest. Other forms of non-independence of movement are
generally not tractable.
(9) Tests of the significance of differences between model predictions and observations
may be based on three approaches to estimating variance: (i) from a series of observations
at different times; (ii) from the Poisson distribution (leading to too many rejections if
paths are not straight); and (iii) from simulations based on a correlated random walk.
(10) A reanalysis of data on encounters between groups of grey-cheeked mangabeys tends
to support the initial conclusion that groups avoid each other, but also suggests three
additional factors that might contribute to the low number of observed encounters relative
to predictions from the ideal gas model: (i) restricted home-range overlap of neighbours
compared with panmixis, (ii) the possibility of overlooking re-encounters within an hour
of an earlier encounter, and (iii) the wide confidence intervals calculated from simulations
of a correlated random walk.
X. ACKNOWLEDGEMENTS
Many thanks to Greg Sandland for help in searching the epidemiological literature, to
Buks van Rensburg for responding to queries about contact rates in dense gases, and to
Jeff Lucas, Walter Jetz, Heike Reise and two referees for their comments on versions of
the manuscript. The authors we criticise have responded in a most positive way; in
particular we thank Kevin Bailey, Greg Farley, Sandy Harcourt, Margaret Kinnaird and
Stam Nicolis for their comments.
XI. REFERENCES
ADLER, F. R. & GORDON, D. M. (1992). Information collection and spread by networks of
patrolling ants. American Naturalist 140, 373–400.
ALLEE, W. C. (1938). The Social Life of Animals. W. W. Norton, New York.
ANDERSON, J. J., GURARIE, E. & ZABEL, R. W. (2005). Mean free-path length theory of
predator-prey interactions: application to juvenile salmon migration. Ecological
Modelling 186, 196–211.
BAILEY, K. M. & BATTY, R. S. (1983). A laboratory study of predation by Aurelia aurita on
larval herring (Clupea harengus): experimental observations compared with model
predictions. Marine Biology 72, 295–301.
BARKLEY, R. A. (1964). The theoretical effectiveness of towed-net samplers as related to
sampler size and to swimming speed of organisms. Journal du Conseil, Conseil
International pour l'Exploration de la Mer 29,146–157.
BARRETT, L. & LOWEN, C. B. (1998). Random walks and the gas model: spacing behaviour
of grey-cheeked mangabeys. Functional Ecology 12, 857–865.
BEGON, M., BENNETT, M., BOWERS, R. G., FRENCH, N. P., HAZEL, S. M. & TURNER, J.
(2002). A clarification of transmission terms in host-microparasite models: numbers,
densities and areas. Epidemiology and Infection 129, 147–153.
BENNETT, E. L. (1984). The banded langur: ecology of a colobine in West Malaysian rainforest. Unpublished Ph.D. dissertation, University of Cambridge.
BERGMAN, C. M., SCHAEFER, J. A. & LUTTICH, S. N. (2000). Caribou movement as a
correlated random walk. Oecologia 123, 364-374.
28
BUCKLAND, S. T., ANDERSON, D. R., BURNHAM, K. P., LAAKE, J. L., BORCHERS, D. L. &
THOMAS L. (2001). Introduction to Distance Sampling: Estimating Abundance of
Biological Populations. Oxford University Press, Oxford.
CHAPMAN, S. & COWLING, T. G. (1952). The Mathematical Theory of Non-uniform Gases.
Cambridge University Press, Cambridge, UK.
CORDS, M. (1987). Mixed-species association of Cercopithecus monkeys in the Kakamega
forest, Kenya. University of California Publications in Zoology 117, 1–109.
COWAN, J. H. JR & HOUDE, E. D. (1992). Size-dependent predation on marine fish larvae
by ctenophores, scyphomedusae, and planktivorous fish. Fisheries Oceanography 1,
113–126.
COX, P. A. (1983). Search theory, random motion, and the convergent evolution of pollen
and spore morphology in aquatic plants. American Naturalist 121, 9–31.
COX, P. A. & SETHIAN, J. A. (1985). Gamete motion, search, and the evolution of
anisogamy, oogamy, and chemotaxis. American Naturalist 125, 74–101.
CROWLEY, P. H., TRAVERS, S. E., LINTON, M. C., COHN, S. L., SIH, A. & SARGENT, R. C.
(1991). Mate density, predation risk, and the seasonal sequence of mate choices: a
dynamic game. American Naturalist 137, 567–596.
DE VITA, J., KELLY, D. & PAYNE, S. (1982). Arthropod encounter rate: a null model based
on random motion. American Naturalist 119, 499–510.
DONCASTER, C. P. (1990). Non-parametric estimates of interaction from radio-tracking
data. Journal of Theoretical Biology 143, 431–443.
DUNBAR, R. I. M. (1988). Primate Social Systems. Croom Helm, London.
DUNBAR, R. I. M. (1995). The mating system of callitrichid primates: I. Conditions for the
coevolution of pair bonding and twinning. Animal Behaviour 50, 1057–1070.
DUNBAR, R. I. M. (2000). Male mating strategies: a modeling approach. In Primate Males:
Causes and Consequences of Variation in Group Composition (ed. P. M. Kappeler),
pp. 259–268. Cambridge University Press, Cambridge.
DUNBAR, R. I. M. (2002). Modelling primate behavioral ecology. International Journal of
Primatology 23, 785–819.
DUSENBERY, D. B. (1992). Sensory Ecology: How Organisms Acquire and Respond to
Information. W. H. Freeman, San Francisco.
DUSENBERY, D. B. (2000). Selection for high gamete encounter rates explains the success
of male and female mating types. Journal of Theoretical Biology 202, 1–10.
DUSENBERY, D. B. (2002). Ecological models explaining the success of distinctive sperm
and eggs (oogamy). Journal of Theoretical Biology 219, 1–7.
DUSENBERY, D. B. (2006). Selection for high gamete encounter rates explains the
evolution of anisogamy using plausible assumptions about size relationships of
swimming speed and duration. Journal of Theoretical Biology 241, 33–38.
DWYER, G. & ELKINTON, J. S. (1993). Using simple models to predict virus epizootics in
gypsy moth populations. Journal of Animal Ecology 62, 1–11.
EVANS, G. T. (1989). The encounter speed of moving predator and prey. Journal of
Plankton Research 11, 415–417.
FARLEY, G. S. (2002). Helical nature of sperm swimming affects the fit of fertilizationkinetics models to empirical data. Biological Bulletin 203, 51–57.
FARLEY, G. S. & LEVITAN, D. R. (2001). The role of jelly coats in sperm-egg encounters,
fertilization success, and selection on egg size in broadcast spawners. American
Naturalist 157, 626–636.
FENTON, A., FAIRBAIRN, J. P., NORMAN, R. & HUDSON, P. J. (2002). Parasite transmission:
reconciling theory and reality. Journal of Animal Ecology 71, 893–905.
FISHER, N. I. (1993). Statistical Analysis of Circular Data. Cambridge University Press,
29
Cambridge, UK.
FUIMAN, L. A. & GAMBLE, J. C. (1989). Influence of experimental manipulations on
predation of herring larvae by juvenile herring in large enclosures. Rapports et
Procès-Verbaux des Réunions du Conseil International pour l’Exploration de la Mer
191, 359–365.
GASTON, A. J. & SMITH, G. E. J. (1984). The interpretation of aerial surveys for seabirds:
some effects of behaviour. Occasional Paper No. 53 Canadian Wildlife Service.
GASTON, A. J., COLLINS, B. L. & DIAMOND, A. W. (1987). The “snapshot” count for
estimating densities of flying seabirds during boat transects: a cautionary comment.
Auk 104, 336–338.
GERRITSEN, J. (1980a). Adaptive responses to encounter problems. In Evolution and
Ecology of Zooplankton Communities (ed. W. C. Kerfoot), pp. 52–62. University
Press, Hanover, NH.
GERRITSEN, J. (1980b). Sex and parthenogenesis in sparse populations. American
Naturalist 115, 718–742.
GERRITSEN, J. & STRICKLER, J. R. (1977). Encounter probabilities and community
structure in zooplankton: a mathematical model. Journal of the Fisheries Research
Board of Canada 34, 73–82.
GIGERENZER, G., SWIJTINK, Z., PORTER, T., DASTON, L., BEATTY, J. & KRÜGER, L. (1989).
The Empire of Chance: How Probability Changed Science and Everyday Life.
Cambridge University Press, Cambridge, UK.
GOMPPER, M. E. & WRIGHT, A. N. (2005). Altered prevalence of raccoon roundworm
(Baylisascaris procyonis) owing to manipulated contact rates of hosts. Journal of
Zoology 266, 215–219.
GORDON, D. M., PAUL, R. E. & THORPE, K. (1993). What is the function of encounter
patterns in ant colonies? Animal Behaviour 45, 1083–1100.
GURSKY, S. (2005). Associations between adult spectral tarsiers. American Journal of
Physical Anthropology 128, 74–83.
HAM, R. M. (1994). Behaviour and ecology of grey-cheeked mangabey (Cercocebus
albigena) in the Lope Reserve, Gabon. Unpublished Ph.D. dissertation, University
of Stirling.
HARCOURT, A. H. & GREENBERG, J. (2001). Do gorilla females join males to avoid
infanticide? A quantitative model. Animal Behaviour 62, 905–915.
HILL, C. M. (1991). A study of territoriality in Cercopithecus diana. Do females take an
active part in territorial defence? Unpublished D. Phil. dissertation, University
College London.
HOLENWEG, A.-K., NOË, R. & SCHABEL, M. (1996). Waser’s gas model applied to
associations between red colobus and Diana monkeys in the Taï National Park, Ivory
Coast. Folia Primatologica 67, 125–136.
JEFFREYS, H. (1961). Theory of Probability (3rd edition). Clarendon Press, Oxford.
JETZ, W., CARBONE, C., FULFORD, J. & BROWN, J. H. (2004). The scaling of animal space
use. Science 306, 266–268.
JOLLY, A., RASAMIMANANA, H. R., KINNAIRD, M. F., O’BRIEN, T. G., CROWLEY, H. M.,
HARCOURT, C. S., GARDNER, S. & DAVIDSON, J. M. (1993). Territoriality in Lemur
catta groups during the birth season at Berenty, Madagascar. In Lemur Social
Systems and Their Ecological Basis (eds. P. M. Kappeler and J. U. Ganzhorn), pp.
85–109. Plenum Press, New York.
KATONA, S. K. (1973). Evidence for sex pheromones in planktonic copepods. Limnology
and Oceanography 18, 574–583.
KAUZMANN, W. (1966). Kinetic Theory of Gases. Thermal Properties of Matter. Volume 1.
30
W. A. Benjamin Inc., New York.
KENWARD, R. E., RUSHTON, S. P., PERRINS, C. M., MACDONALD, D. W. & SOUTH, A. B.
(2002). From marking to modeling: dispersal study techniques for land vertebrates.
In Dispersal Ecology (eds. T. M. Bullock, R. E. Kenward and R. S. Hails), pp. 50–
71. Blackwell, Oxford.
KILTIE, R. A. (1980). Application of search theory to the analysis of prey aggregation as an
antipredation tactic. Journal of Theoretical Biology 87, 201–206.
KINNAIRD, M. F. & O’BRIEN, T. G. (2000). Comparative movement patterns of two semiterrestrial cercopithecine primates: the Tana River crested mangabey and the
Sulawesi crested black macaque. In On the Move: How and Why Animals Travel in
Groups (eds. S. Boinski and P. A. Garber), pp. 327–350. University of Chicago
Press, Chicago.
KOOPMAN, B. O. (1956). The theory of search. I. Kinematic bases. Operations Research 4,
324–346.
LAING, J. (1938). Host-finding by insect parasites. II. The chance of Trichogramma
evanescens finding its hosts. Journal of Experimental Biology 15, 281–302.
LAVAL, P. (1974). Un modèle mathématique de l’évitement d’un filet à plancton, son
application pratique, et sa vérification indirecte en recourant au parasitisme de
l’amphipode hypéride Vibilia armata Bovallius. Journal of Experimental Marine
Biology and Ecology 14, 57–87.
LETCHER, B. H., RICE, J. A., CROWDER, L. B. & ROSE, K. A. (1996). Variability in survival
of larval fish: disentangling components with a generalized individual-based model.
Canadian Journal of Fisheries and Aquatic Science 53, 787–801.
LOTKA, A. J. (1924). Elements of Physical Biology. Williams and Wilkins, Baltimore.
LOWEN, C. & DUNBAR, R. I. M. (1994). Territory size and defendability in primates.
Behavioral Ecology and Sociobiology 35, 347–354.
LUDING, S. (2001). Global equation of state of two-dimensional hard sphere systems.
Physics Review E 63, 042201.
MANLY, B. F. J. (1997). Randomization, Bootstrap and Monte Carlo Methods in Biology
(2nd edition). Chapman and Hall, London.
MAXWELL, J. C. (1860). Illustrations of the dynamical theory of gases. Part I. On the
motions and collisions of perfectly elastic spheres. Philosophical Magazine 19, 19–
32.
MANICA, A. & JOHNSTONE, R. A. (2004). The evolution of paternal care with overlapping
broods. American Naturalist 164, 517–530.
MCCALLUM, H., BARLOW, N. & HONE, J. (2001). How should pathogen transmission be
modelled? Trends in Ecology and Evolution 16, 295–300.
MITANI, J. C., GRETHER, G. F., RODMAN, P. S. & PRIATNA, D. (1991). Associations among
wild orang-utans: sociality, passive aggregations or chance? Animal Behaviour 42,
33–46.
MOSIMANN, J. E. (1958). The evolutionary significance of rare matings in animal
populations. Evolution 12, 246–261.
NICOLIS, S. C., THERAULAZ, G. & DENEUBOURG, J. L. (2005). The effects of aggregates on
interaction rate in ant colonies. Animal Behaviour 69, 535–540.
OKUBO, A. (1980). Diffusion and Ecological Problems: Mathematical Models. SpringerVerlag, Berlin.
OLSON, F. C. W. (1964). The survival value of fish schooling. Journal du Conseil, Conseil
Permanent International pour l’Exploration de la Mer 29, 115–116.
PARADIS, A. R. & PEPIN, P. (2001). Modelling changes in the length-frequency
distributions of fish larvae using field estimates of predator abundance and size
31
distributions. Fisheries Oceanography 10, 217–234.
PARADIS, A. R., PÉPIN, M. & PEPIN, P. (1999). Disentangling the effects of size-dependent
encounter and susceptibility to predation with an individual-based model for fish
larvae. Canadian Journal of Fisheries and Aquatic Science 56, 1562–1575.
PODOLSKY, R. D. (2002). Fertilization ecology of egg coats: physical versus chemical
contributions to fertilization success of free-spawned eggs. Journal of Experimental
Biology 205, 1657–1668.
PRATT, S. C. (2005). Quorum sensing by encounter rates in the ant Temnothorax
albipennis. Behavioral Ecology 16, 488–496.
RANDERSON, J. P. & HURST, L. D. (2001). The uncertain evolution of the sexes. Trends in
Ecology and Evolution 16, 571–579.
ROOT, R. B. & KAREIVA, P. M. (1984). The search for resources by cabbage butterflies
(Pieris rapae): ecological consequences and adaptive significance of Markovian
movements in a patchy environment. Ecology 65, 147–165.
ROSENTHAL, H. & HEMPEL, G. (1970). Experimental studies in feeding and food
requirements of herring larvae (Clupea harengus L.). In Marine Food Chains (ed. J.
H. Steele), pp. 344–364. University of California Press, Berkeley.
ROWCLIFFE, J. M., COWLISHAW, G. & LONG, J. (2003). A model of human hunting impacts
in multi-prey communities. Journal of Applied Ecology 40, 872–889.
SCHÜLKE, O. & KAPPELER, P. M. (2003). So near and yet so far: territorial pairs but low
cohesion between pair partners in a nocturnal lemur, Phaner furcifer. Animal
Behaviour 65, 331–343.
SCUDO, F. M. (1967). The adaptive value of sexual dimorphism: I, anisogamy. Evolution
21, 285–291.
SKELLAM, J. G. (1958). The mathematical foundations underlying the use of line transects
in animal ecology. Biometrics 14, 385–400.
SPEAR, L., NUR, N. & AINLEY, D. G. (1992). Estimating absolute densities of flying
seabirds using analyses of relative movement. Auk 109, 385–389.
STANLEY, J. (1932). A mathematical theory of the growth of populations of the flour
beetle, Tribolium confusum, Duv. Canadian Journal of Research 6, 632–671.
STRUHSAKER, T. T. (1981). Polyspecific associations among tropical rain-forest primates.
Zeitschrift für Tierpsychologie 57, 268–304.
SUGIURA, H., SAITO, C., SATO, S., AGETSUMA, N., TAKAHASHI, H., TANAKA, T., FURUICHI,
T. & TAKAHATA, Y. (2000). Variation in intergroup encounters in two populations of
Japanese macaques. International Journal of Primatology 21, 519–535.
TASKER, M. L., JONES, P. H., DIXON, T. & BLAKE, B. F. (1984). Counting seabirds at sea
from ships: a review of methods employed and a suggestion for a standardized
approach. Auk 101, 567–577.
TRAVIS, J. M. J. & PALMER, S. C. F. (2005). Spatial processes can determine the
relationship between prey encounter rate and prey density. Biology Letters 1, 136138.
VAN RENSBURG, E. J. J. (1993). Virial coefficients for hard discs and hard spheres. Journal
of Physics A: Mathematical and General 26, 4805–4818.
VAN SCHAIK, C. P. (1999). The socioecology of fission-fusion sociality in orangutans.
Primates 40, 69–86.
VAN SCHAIK, C. P. & DUNBAR, R. I. M. (1990). The evolution of monogamy in large
primates: a new hypothesis and some crucial tests. Behaviour 115, 30–62.
VAN SCHAIK, C. P., VAN AMERONGEN, A. & MOUTON, J. W. (1985). The organization of
range use in Sumatran long-tailed macaques. In The Socio-ecology of Sumatran
Long-tailed Macaques (Macaca fascicularis): 1. Costs and Benefits of Group Living
32
(ed. C. P. van Schaik), pp 23–63. Elinkwijk BV, Utrecht.
VOGEL, H., CZIHAK, G., CHANG, P. & WOLF, W. (1982). Fertilization kinetics of sea urchin
eggs. Mathematical Biosciences 58, 189–216.
WASER, P. M. (1975). Spatial associations and social interactions in a “solitary” ungulate:
the bushbuck Tragelaphus scriptus (Pallas). Zeitschrift für Tierpsychologie 37, 24–
36.
WASER, P. M. (1976). Cercocebus albigena: site attachment, avoidance, and intergroup
spacing. American Naturalist 110, 911–935.
WASER, P. (1977a). Feeding, ranging and group size in the mangabey Cercocebus
albigena. In Primate Ecology: Studies of Feeding and Ranging Behaviour in
Lemurs, Monkeys and Apes (ed. T. H. Clutton-Brock), pp. 183–222. Academic Press,
London.
WASER, P. (1977b). Individual recognition, intragroup cohesion and intergroup spacing:
evidence from sound playback to forest monkeys. Behaviour 60, 28–74.
WASER, P. M. (1982). Primate polyspecific associations: do they occur by chance? Animal
Behaviour 30, 1–8.
WASER, P. M. (1984). “Chance” and mixed-species associations. Behavioral Ecology and
Sociobiology 15, 197–202.
WASER, P. M. (1987). Interactions among primate species. In Primate Societies (eds. B. B.
Smuts, D. L. Cheney, R. M. Seyfarth, R. W. Wrangham and T. T. Struhsaker), pp.
210–226. University of Chicago Press, Chicago.
WEIHONG, J., WHITE, P. C. L. & CLOUT, M. N. (2005). Contact rates between possums
revealed by proximity data loggers. Journal of Applied Ecology 42, 595–604.
WHITESIDES, G. H. (1989). Interspecific associations of Diana monkeys, Cercopithecus
diana, in Sierra Leone, West Africa: biological significance or chance? Animal
Behaviour 37, 760–776.
YAPP, W. B. (1956). The theory of line transects. Bird Study 3, 93–104.
33
Table 1. Observed and predicted numbers of encounters for a mangabey group followed
for 2261 half-hour intervals in Kibale Forest, Uganda. We were unable to retrieve the data
from 74 of the original 2474 data points used by Waser (1976) owing to loss or
degradation of the IBM cards. The models are explained in the text. The last column
indicates the basis of the confidence limits: t = t-distribution, P = Poisson distribution, s =
simulation
Encounters < 200 m
mean
95% CL
Observations
4
Encounters < 500 m
mean
95% CL
11
2 ≤ x ≤ 18
t
Predictions of Waser (1976)
Gas model: constant speed
Gas model: Maxwell-Boltzmann speeds
Gas model: observed speed distribution
Ditto, + range overlap
Ditto, + observed usage of overlap
22.5
24.6
27.4
29.7 21 ≤ x ≤ 38
13.5
12.5
50.0
50.1
55.7
60.5 48 ≤ x ≤ 73 P
27.5
25.5
Correlated random walk simulation
Ditto, > 1 h apart
Ditto, > 4 h apart
29.7
19.7
14.1
8 ≤ x ≤ 59
5 ≤ x ≤ 39
4 ≤ x ≤ 27
60.6 27 ≤ x ≤ 103 s
38.9 17 ≤ x ≤ 66 s
27.3 12 ≤ x ≤ 46 s
Intercept with own time-displaced trajectory
16.0
34
21.3
FIGURE CAPTIONS
Fig. 1. (A) Trajectories of three individuals, each travelling at the same speed v, over a
period t, so the length of each trajectory is vt. (B) The same trajectories as in A but viewed
relative to the position of individual 1, which thus now appears to remain in the same spot.
The dashed lines are the translated vectors of individuals 1 and 2 from A used to construct
the relative trajectory of individual 2 (solid line); the length of this relative trajectory is 2vt
sin(θ/2) (since the dashed lines form an isosceles triangle with it). (C) The areas swept out
by the leading edge of a disc of radius D following the trajectories of individuals 2 and 3
relative to 1. This shows that individual 1 comes within D units of individual 2 but never
so close to individual 3. The probability of an encounter with the focal individual depends
on the size of this swept-out area, which for individual 2 is 2D 2vt sin(θ/2). The average
area for a randomly orientated individual (i.e. θ evenly distributed between 0 and π) is thus
π
2 D 2vt sin(θ/2) dθ/π = 2D 4vt/π = 8Dvt/π. The 4vt/π term is the mean relative speed
0
[CORRECTION: THIS SHOULD READ “mean relative trajectory length”].
Fig. 2. Mean relative speed in three dimensions. Dashed lines show the directions of two
individuals at an angle θ to each other; without loss of generality the focal individual is
shown vertically orientated. The trajectory of the other individual relative to the focal
individual (thick solid line) has length 2vt sin(θ/2), as in the two-dimensional case (Fig. 1).
However, randomly orientated individuals now yield some values of θ more commonly
than others: the probability of the angle lying between θ and θ + δθ is the area of a strip of
width δθ following the dotted line on the figure, divided by the surface area of the sphere
= 2π sin θ δθ/(4π) = sin θ δθ/2. Thus the mean length of the relative trajectory is
π
vt sin(θ/2) sin θ dθ = 4vt/3.
0
Fig. 3. (A) A modification of Fig. 1C to describe the situation when the number of
occasions on which an association occurs is counted (instead of the number of new
encounters). The expected number of associations is how many discs cover a random spot,
and is thus the product of the area of each disc (πD2), the number of observations made (5
in this example), and the density of individuals. Speed and direction make no difference:
the sum of the areas occupied by the discs is the same for individual 2 (hatched) as
individual 3 (solid shading). For small values of θ (the angle between trajectories) the
relative trajectories are shorter so that the discs overlap (as they are also more liable to do
if the sampling interval is short compared to the speeds); an association may then persist
from one observation to the next and should then be counted more than once to match
with the model. (B) The number of associations between D – δD and D + δD from a focal
individual is given by the area of the annuli (each has area 4D δD + δD2 ~ 4D δD for
small δD, and is thus proportional to D). (C) The number of encounters within D + δD, but
not closer than D – δD is given by the shaded region. For long periods of observation the
semicircles at each end are a minor component, so the area is roughly 4 δD length of
relative trajectory, and thus is independent of D.
Fig. 4. Mean relative speed of predator and prey as a function of the mean speed of the
predator. Encounter rate is directly proportional to mean relative speed. The mean speed
of the prey is set to 1. The lines differ in whether movement is in two (2-D) or three (3-D)
35
dimensions, and in whether the prey and predator each have constant speed or whether this
varies according to a Maxwell-Boltzmann (M-B) distribution.
Fig. 5. Deviations from the ideal gas model with a correlated random walk. As a baseline,
each simulation consists of 64 straight-line steps, each of 100 m, all individuals move at
the same constant speed, detection distance D = 100 m, density ρ = 1 km–2. Turning angle
follows a von Mises distribution with dispersion parameter κ; we transform κ to the
percentage of times that the direction changes by less than 90º (“% steps forward”): κ = 0
50% forward (uncorrelated random walk), κ = 1 78% forward, κ = 100%
forward (gas model assumption). (A) Distribution of number of encounters: thick line =
gas model, broken line κ = 1, thin line κ = 0. (B) Ratio of variance to mean number of
encounters, as a function of κ; solid line = baseline parameters, dashed line D = 25 m and
ρ = 4 km-2. (C) Number of different individuals encountered as a function of κ; parameters
as in B. (D) Number of different individuals encountered as a function of number of steps;
other parameters as in A. Each point is based on ≥ 105 simulations.
Fig. 6. Deviations from the ideal gas model when individuals rebound. Density is
measured as individuals per unit2 and rebounds occur at a separation of 1 unit between
centres. Speeds are constant, with individuals heading directly away after contact. At high
densities, disks are initially arranged on a triangular lattice on a toroidal surface;
simulation continues until collision rate has reached equilibrium. (A) Number of
encounters relative to ideal gas model. Dashed line shows the analytically derived
correction factor based on virial coefficients up to order 8 (van Rensburg, 1993). Solid line
shows the factor by which collision rate in our simulations exceeded the gas-model
prediction. The inflection is associated with the “phase change” to a crystalline form in
which individuals are unable to slip past one another (see Luding, 2001). Each point is
based on 20000–100000 simulations of 100 disks. (B) Number of different individuals
encountered in a travel distance of 10 units. Solid line is from our simulations, dashed line
is that predicted by the gas model assuming straight-line trajectories. Each point is based
on 10000 simulations of a minimum of 100 disks.
Fig. 7. If each individual can detect another d away, and they form groups of diameter s,
then groups can detect each other when their centres are d + s apart and so this should
replace D in the standard formulae.
Fig. 8. Configurations of overlapping territories in which home range density = h = 1.56.
Home ranges overlapping the stippled home range are shown hatched. For configurations
A, B and C, the correction factor assuming equal usage within a home range = 2(h – 1)/h2
= 0.46. For configuration D, the corresponding correction factor = 4(h – 1)(2h – 1)/h2 =
0.61.
Fig. 9. Encounters between jellyfish moving at speed v at an angle θ to herring moving at
speed u. We assume that the herring has length L but is infinitely thin. A shows this
situation in the plane common to both movement vectors. The trajectory of the herring
relative to the jellyfish is shown by the thick vector and has length y =
t u 2 v 2 uv cos θ [CORRECTION: THIS SHOULD READ t u 2 v 2 2uv cos θ ].
The parallelogram outlined by a dashed line shows those starting positions of an infinitely
small jellyfish that would result in a new encounter within time t. The width of this
parallelogram perpendicular to the relative trajectory is z = (Lv sin θ)/y. However, instead
36
the jellyfish are considered to be spheres with radius rj, and the three circles show
examples of positions that result in one part of the jellyfish just touching the herring. The
dotted outline marks the envelope of positions of the jellyfish centre that result in new
encounters. B shows the cross section of this space cut perpendicular to the relative
trajectory. The area of the cross section is πrj2 + rjz. This must be multiplied by the length
of the relative trajectory of the herring (y) to give the volume of the space: y(πrj2 +
(rjLv sin θ)/y) = yπrj2 + rjLv sin θ. Calculating the expected number of encounters then
involves integrating this over θ (see Fig. 2).
Fig. 10. Calculation of the number of birds entering a rectangular area (2D a) in front of
a ship moving with speed u. Consider the ship’s motion relative to a bird flying with speed
v at an angle of θ to the course of the ship. Relative to the bird the ship moves a distance
t u 2 v 2 uv cos θ [CORRECTION: THIS SHOULD READ t u 2 v 2 2uv cos θ ] in
time t (shown as thick vector in the figure). The movement of the rectangular area along
this relative trajectory covers a strip with sides of this length and with width
(2 D) 2 a 2 sin( + |α|), where = tan–1(2D/a) and α = tan–1((v sin θ)/(u – v cos θ)). The
area of the strip is the product of this width and length, and the expected number of
encounters is this area times bird density (integrating over θ if birds are not all travelling
in the same direction).
37
Fig. 1
38
Fig. 2
39
Fig. 3
40
Fig. 4
Mean relative speed
3
Prey
Predator
2-D constant constant
3-D constant constant
2-D
M-B
constant
2/3-D
M-B
M-B
2
1
0
1
2
Mean speed of predator
41
3
0.5
Var./mean number of encounters
Fig. 5
A
B
3.0
0.4
Proportion
2.5
0.3
2.0
0.2
1.5
0.1
0.0
0
5
1.0
10
50
C
1.6
Number of individuals encountered
Number of individuals encountered
Number of encounters
60
70
80
90
% steps forward
100
D
3.0
1.4
2.0
1.2
1.0
1.0
0.8
0.0
50
60
70
80
% steps forward
90
100
0
42
32
64
Number of steps
96
128
Fig. 6
Encounter rate / ideal-gas prediction
10
A
8
6
4
2
0.0
0.2
0.4
0.6
0.8
1.0
2/3
1.2
0.8
1.0
2/3
1.2
Density
Number of individuals encountered
8
B
6
4
2
0
0.0
0.2
0.4
0.6
Density
43
Fig. 7
44
Fig. 8
45
Fig. 9
46
Fig. 10
47