Evidence of peak-to-peak dynamics in ecology
-
Upload
sergio-rinaldi -
Category
Documents
-
view
214 -
download
2
Transcript of Evidence of peak-to-peak dynamics in ecology
R E P O R TEvidence of peak-to-peak dynamics in ecology
Sergio Rinaldi, Matteo Candaten
and Renato Casagrandi
CIRITA, Politecnico di Milano,
Via Ponzio 34/5, 20133 Milan,
Italy.
Correspondence:
Abstract
Remarkably irregular peaks characterize the dynamics of many plant and animal
populations. As such peaks are often associated with undesirable consequences (e.g. pest
outbreaks, epidemics, forest ®res), the forecast of the forthcoming peak is a problem of
major concern. Here we show, through the analysis of a number of models and of some
of the longest and most celebrated ecological time-series, that the intensity of the
forthcoming peak can often be predicted simply from the previous peaks. When this is
possible, one can also predict the time of occurrence of the forthcoming peak.
Keywords
Nonlinear dynamics, chaos, quasiperiodic behaviour, models, peak-to-peak dynamics,
populations, forecast, strange attractors, fractal dimension, extreme episodes.
Ecology Letters (2001) 4: 610±617
I N T R O D U C T I O N
After transient, any deterministic, dissipative, nonlinear
system settles on an attractor (an equilibrium, a limit cycle, a
torus or a strange attractor) and remains there forever if it is
not perturbed (Strogatz 1994). Insights about the attractor
can be obtained, even in the absence of a mathematical
model, if a single variable y of time t has been recorded for a
long period, while the system was on the attractor (Hastings
et al. 1993). In particular, one can extract from the record all
peaks (local maxima) of the variable y, say yi (i � 0, 1, 2,¼),
and plot them one against the previous one, thus obtaining a
set of points (yi, yi+1), called peak-to-peak plot (PPP). The same
plot is sometimes called next amplitude plot (Olsen & Degn
1985) or next-maximum plot (Scott 1994) or Lorenz plot
(Strogatz 1994). From the same record one can also extract
the times of occurrence of the peaks ti and compute the
return times si+1 � ti+1 ± ti.
If the regime is periodic and there are k peaks per period,
the PPP is composed only of k distinct points. By contrast, if
the regime is quasiperiodic or chaotic, i.e. if the attractor is a
torus or a strange attractor, the points of the PPP are all
distinct and sometimes display ®liform geometries (i.e. the
PPP has fractal dimension close to 1, see Strogatz 1994). More
precisely, when the attractor is a high-dimensional strange
attractor, the PPP is a cloud-like set. Conversely, the points
of the PPP lie on a closed regular curve (the slice of a torus)
when the regime is quasiperiodic and lie roughly on a curve
when the attractor is a low-dimensional strange attractor.
We show in the next section that when the PPP is
®liform, i.e. when the attractor is low-dimensional, the
intensity of the forthcoming peak and its time of occurrence
can be predicted with remarkable accuracy from the
previous peak or from the two previous peaks. The ®rst
possibility has been pointed out by Lorenz himself in his
pioneering paper on chaos (Lorenz 1963) and then
described in more detail by various authors (e.g. Olsen
& Degn 1985, in the context of biology). Peak-to-peak
analysis is a special case of the standard technique for
reconstructing strange attractors (Packard et al. 1980),
popularized in the 1980s among ecologists in a series of
ground-breaking articles (e.g. Schaffer & Kot 1985a, b;
Schaffer 1985; Olsen et al. 1988). Peak-to-peak analysis
requires only very little effort (it can be performed by hand
in a few minutes), especially if it is compared with other
techniques used in ecology (Sugihara & May 1990; Casdagli
1989; Ellner & Turchin 1995; Costantino et al. 1997). As far
as we know, only six ecological models have been studied
until now through peak-to-peak analysis (Funasaki & Kot
1993; Boer et al. 1998; Rinaldi & Solidoro 1998; Blasius &
Stone 2000; Rinaldi et al. 2001a, b).
M A T E R I A L S A N D M E T H O D S
To illustrate why a PPP can be ®liform and to detect the
consequences of this fact in terms of peak predictability, we
consider in this section a paradigmatic example of chaotic
ecosystems, namely the Rosenzweig±MacArthur tritrophic
food chain (Hastings & Powell 1991)
dx1
dt� rx1 1ÿ x1
K
� �ÿ a2x1
b2 � x1
x2 �1�
Ecology Letters, (2001) 4: 610±617
Ó2001 Blackwell Science Ltd/CNRS
dx2
dt� c2
a2x1
b2 � x1
x2 ÿ d2x2 ÿ a3x2
b3 � x2
x3 �2�dx3
dt� c3
a3x2
b3 � x2
x3 ÿ d3x3 �3�
where x1, x2, x3 are the abundances of prey, predator and
top-predator, r and K are intrinsic growth rate and carrying
capacity of the prey and (ai, bi, ci, di) i � 2,3 are maximum
predation rate, half-saturation constant, ef®ciency and death
rate of predator and top-predator, respectively. For suitable
parameter values (see caption of Fig. 1) the model is chaotic
(Hastings & Powell 1991), its strange attractor A looks like a
tea-cup and has fractal dimension slightly greater than 2 (see
Fig. 1A).
Let y(t) � x3(t) and assume that top-predator abundance
is measured for a long period of time. The points in state
space where y(t) is either maximum or minimum are
characterized by dy/dt � 0, and lie on the plane x2 � b3d3/
(c3a3 ± d3) (see Eq. (3) with x3 ¹ 0 and dx3/dt � 0). This
plane, indicated by S in Fig. 1(B), is a particular PoincareÂ
section, namely a manifold transversal to the attractor
(Strogatz 1994). The intersection of A with S provides the
points of the attractor where y is maximum or minimum: of
course, after a maximum (point 0) we have a minimum
(point 1) and then a maximum again (point 2) and so on. In
other words, when the system is on the attractor, Eqs. (1)±(3)
de®ne a transformation, called Poincare map, from any
maximum (point 0) to the next maximum (point 2) on a
subset P of S. Moreover, if the dimension of A is close to 2,
then the dimension of P is close to 1, i.e. P is a ®liform set,
namely a fractal set resembling a thick curve. Thus, in
practice, each contour line y � const. on S intersects P at one
or more distinct points. In the present case (see Fig. 1C),
each contour line (straight line x3 � const.) intersects P only
at one point. This is a fortunate situation: it implies that the
contour line y � yi, where yi is the amplitude of the ith peak,
intersects P only at one point which is then transformed by
the Poincare map into a next maximum point at the
intersection of P with the contour line y � yi+1. Thus, a
unique peak yi+1 is associated to the peak yi, i.e. the ®liform
PPP can be represented by a function
yi�1 � F yi� � �4�as shown in Fig. 1(D). By contrast, when some contour line
y � const. intersects the ®liform set P at h ³ 2 points, the
PPP is still ®liform but cannot be described by Eq. (4),
because the prediction of the next peak yi+1 cannot be
performed exclusively on the basis of the last peak yi.
However, it can be shown (Candaten & Rinaldi 2000) that
independently upon the value of h, each peak yi+1 is uniquely
determined by the two previous peaks, i.e.
yi�1 � F�yi ; yiÿ1� �5�Thus, in conclusion, if the attractor is a torus or a strange
attractor with fractal dimension close to 2, the PPP
is ®liform and successive peaks obey simple recursive
Figure 1 Peak-to-peak dynamics of the
Rosenzweig±MacArthur tritrophic food
chain model (Eqs. 1±3). Parameter values are
r � K � c2 � c3 � 1, a2 � 1.66, b2 � 0.33,
d2 � 0.27, a3 � 0.05, b3 � 0.5, d3 � 0.01.
(A) The teacup strange attractor A in state
space; (B) the Poincare section S where
top-predator abundance is either maximum
or minimum; (C) the subset P where top-
predator abundance is maximum and the
contour lines x3 � const.; (D) the peak-to-
peak plot.
Peak-to-peak dynamics 611
Ó2001 Blackwell Science Ltd/CNRS
equations. When this happens we say that the system has
peak-to-peak dynamics (PPD). Such PPD are called simple in the
case of Eq. (4) and complex when Eq. (5) holds.
Finally, it is worth noticing that when PPD are simple the
peak yi identi®es a unique point on the Poincare section
(point 0 in Fig. 1B) so that Eqs. (1)±(3) can be integrated
from that point until after si+1 units of time, the next peak
yi+1 (point 2 in Fig. 1B) is reached. This means that the
return time si+1 is a function of the previous peak, i.e.
si+1 � T(yi), so that the time of occurrence of the
forthcoming peak can also be predicted from the previous
peak through the formula ti+1 � ti + T(yi). Of course, the
function T(yi) must be substituted by a function T(yi, yi±1) in
the case of complex PPD. Moreover, it can be shown
(Candaten & Rinaldi 2000) that the existence of PPD
implies that the return times satisfy a ®rst-order recursive
equation of the form
si�1 � H �si� �6�or, in the most complex cases, a second-order equation
si�1 � H �si ; siÿ1� �7�The fact that the return times associated with a quasiperi-
odic or low-dimensional chaotic regime obey the laws of an
autonomous system has been noticed for the ®rst time in
physics through very simple experiments on dripping taps
(Shaw 1984). But the same phenomenon is sometimes
present also in biology, for example, in the study on the
in¯uence of light stimuli on the locomotor behaviour of
Halobacterium (see Fig. 1g in Schimz & Hildebrand 1992).
The forecast of the time of occurrence of the next peak is of
great importance in ecology, where in many cases the
amplitude of the peak is hardly measurable, whereas its time
of occurrence is perfectly known (see Rinaldi et al. 2001a for
the forecast of insect pest-outbreaks and ®res in Mediter-
ranean forests).
Results
In order to show that PPD are frequent in ecology, we ®rst
present in Fig. 2 a selected sample of PPPs derived from
various models.
Figure 2(A) refers to the plant-herbivore-carnivore model
recently used (Blasius et al. 1999; Blasius & Stone 2000) to
explain the phase synchronization of hare and lynx popu-
lations in Canada. The model is essentially a Rosenzweing±
MacArthur model (Eqs. 1±3) with very high prey carrying
capacity and top-predator predation rate. Moreover, a small
positive term is added at the right-hand side of Eq. (3) to
accommodate for the alternative food sources available to
the lynx. The time-series obtained through simulation shows
that the return times are almost constant, whereas the peak
amplitudes vary remarkably. Although the changes in
abundance look erratic, they can be perfectly explained with
Eq. (4), as pointed out by the PPP. In other words, the model
suggests that each peak of the lynx population can be
predicted years ahead from the previous peak.
Panel B of Fig. 2 comes from a prey±predator model
(Eqs. (1) and (2) with x3 � 0) with periodically varying food
supply. It has long been known that seasonal perturbations
in prey±predator assemblies can give rise to quasiperiodic
and chaotic behaviours. A general discussion of this topic
can be found in Rinaldi et al. (1993), and a recent application
to the population dynamics of Fennoscandian voles is
available in Turchin & Ellner (2000). In the case examined
in Fig. 2(B) the prey time-series is chaotic, with return time
roughly double than the period of the food supply.
However, as the PPP is ®liform, the randomness of the
peaks can be fairly well explained through Eq. (4).
Figure 2(C) has been obtained from the discrete-time
three-stage (Larvae-Pupae-Adults) insect population model
discussed in Costantino et al. (1997)
Lt�1 � bAt exp ÿcel Lt ÿ ceaAt� � �8�Pt�1 � 1ÿ ll� �Lt �9�
At�1 � 1ÿ la� �At � Pt exp ÿcpaAt
ÿ � �10�As the number of points of the PPP (obtained through
simulation of the model) is practically unlimited, all the
details of the PPP can be detected. Thus, it is possible to
distinguish between chaoticity (open regular curves) and
quasiperiodicity (closed regular curve) without performing
complicated tests or computing Lyapunov exponents
(Costantino et al. 1995, 1997). (Obviously, the same distinc-
tion is not possible in the case of a limited number of
recorded peaks). For the parameter values indicated in the
caption, the regime is quasiperiodic because the points of the
PPP lie on a closed regular curve, so that the PPD are
complex and the next peak can be predicted with a model of
type (5).
Figure 2(D) has been obtained from the very simple
insect model with delayed recruitment rate discussed in
Gurney et al. (1980)
dx�t�dt� ÿax�t� � bx�t ÿ s�exp�ÿx�t ÿ s�� �11�
where x is biomass, a and b are mortality and maximum
recruitment rate and s is the time-delay. Periods of very low
abundance regularly alternate with periods of high abun-
dance characterized by bursts of high-frequency oscillations.
The PPP obtained by using all local peaks is a cloud-like set
(not shown), whereas the PPP obtained by retaining only
the highest peak (superpeak) of each burst (marked with *)
has a ®liform, although complex, geometry. This means that
the dynamics of the superpeaks are described by Eq. (5), i.e.
each superpeak can be predicted from the two previous
612 S. Rinaldi, M. Candaten and R. Casagrandi
Ó2001 Blackwell Science Ltd/CNRS
superpeaks, a rather astonishing result. It is worth noticing,
however, that the structure of the PPP can be fully identi®ed
only if the population is observed for a very long time. For
example, if the available record would be that of Fig. 2(D)
(nine superpeaks), one would obtain eight points in the
space (yi, yi+1) (points marked with · in the right column of
Fig. 2D), which do not reveal all the details of the PPP.
Actually, one would most likely conjecture that these eight
points lie on a closed regular curve and hence conclude that
the regime is quasiperiodic. Of course, other records of the
same length, obtained with the same model, would evidence
other parts of the PPP and therefore suggest chaoticity.
Figure 2 shows only a small sample of ecological models
displaying PPD. In fact, we have also identi®ed PPD in the
following 15 models:
· Ricker's model (Ricker 1954);
· The discrete-time logistic model (May 1976);
· Three age-structured population models (Guckenheimer
et al. 1977; Caswell & Weeks 1986; Higgins et al. 1997);
· A model for host±parasitoid interaction (Lauwerier
& Metz 1986);
· Three rodent-mustelid models (Hanski et al. 1993;
Turchin & Hanski 1997; Turchin & Ellner 2000);
· A model of scramble competition (FerrieÁre & Gatto
1995);
· A plant±herbivore±omnivore model (McCann & Hastings
1997);
· A prey±predator evolutionary model (Abrams & Matsuda
1997);
· A model of tritrophic food chains with allochtonous
resources (McCann et al. 1998);
· An algal competition model (Gragnani et al. 1999);
· A model for snowshoe hare demography (King
& Schaffer 2001).
A ®le on these models and their corresponding PPD is
available on request from Renato Casagrandi.
By contrast, the identi®cation of PPD from ®eld and
laboratory data poses some problems. There are three reasons
Figure 2 Peak-to-peak dynamics in four
ecological models. On the left of each panel
there is a short segment of a very long
simulated time-series used to construct the
PPP shown on the right, where the points ·correspond to all pairs of subsequent peaks
of the short segment. (A) Carnivore popu-
lation of the tritrophic food chain model
used in (Blasius et al. 1999) (see Eqs. (1)±(3)
with an extra-term equal to 0.06 in Eq. (3)).
Parameter values are r � d2 � c2 � c3 � 1,
K � a3 � b3 � 8, a2 � 4, b2 � 2, d3 � 10.
(B) Prey population of a prey±predator
model (Eqs. (1), (2) with x3 � 0) with peri-
odically fed prey. Parameter values are
r � d2 � 2p, a2 � 4p, b2 � 0.3, c2 � 1, and
K � K0(1 + esin(2pt)) with K0 � 1.44,
e � 0.65. (C) Larvae abundance of a three-
stage insect model (see Eqs. 8±10) with
cpa � 0.075. All other parameter values are as
in Costantino et al. (1997), namely b � 6.598,
cel � 0.01209, cea � 0.01155, lL � 0.2055,
la � 0.96. (D) Adult individuals of an insect
model with delayed recruitment (see Eq. 11)
(the peaks marked with * are the superpe-
aks). Parameter values are s � 1, a � 8.21,
b � 821.
Peak-to-peak dynamics 613
Ó2001 Blackwell Science Ltd/CNRS
for this. First, even the longest population time-series are too
short for obtaining reliable estimates of PPD and other
dynamic invariants when a sound supporting model is not
available. Second, false peaks due to measurement and
process noise have a severe impact on the PPP. In fact, a false
peak f between two successive true peaks a and b implies the
loss of point (a, b) in the PPP and the addition of two false
points (a, f ) and ( f, b). In principle, data ®ltering should be
used to eliminate false peaks and ®nd reliable estimates of real
peaks. In practice, exclusion of false peaks through common
sense (as done in Fig. 3) can also be effective. Finally, the data
used to construct a PPP should be collected when the system
is on the attractor. Thus, the initial peaks of any laboratory
time-series should be excluded from the analysis because they
usually correspond to a transient towards the attractor. The
elimination of transient peaks may dramatically reduce the
length of an already too short time-series. For this reason, in
our analysis we have systematically sacri®ced only the ®rst
peak of each laboratory time-series.
Despite these dif®culties, we have constructed the PPPs
of some of the longest available ecological time-series and
the results (described in the following) are encouraging: not
only do the points of the PPPs lie on regular curves, but
such curves are shaped as those predicted from the models.
The results are summarized in Fig. 3, which presents four
well-known population time-series. The PPPs obtained
from these series using all their peaks are not shown,
because they do not display sharp geometries. Instead, the
second column of Fig. 3 displays the PPPs obtained by
eliminating the false peaks (marked with h) and the transient
peaks (marked with s). In each one of these ®gures a dotted
curve (obtained without performing any sort of formal
calibration) has been drawn to help the reader in focusing
on a plausible alignment of the points of the PPP. In other
words, the dotted curves are the result of the most crude
and model-free analysis of the data: plot in a plane all pairs
of subsequent peaks and decide (subjectively!) if the points
of the PPP lie or not on one or more curves.
Figure 3 Four population time-series (®rst
column) and corresponding PPPs (second
column). Peaks marked with s and h are
considered as transient and false peaks,
respectively, and are not used for deriving
the PPP. See text for the interpretation of
the dotted lines. (A) Lynx fur returns (in
thousands) in the MacKenzie River Region
(Elton & Nicholson 1942); (B) Paramecium
aurelia (in hundreds) in a controlled experi-
ment (Veilleux 1979); (C) larvae of ¯our
beetle Tribolium castaneum (in hundreds)
(see Fig. 2 with cpa � 0.25 in Costantino
et al. 1997); (D) Nicholson's long experiment
on blow¯y Lucilia cuprina (in thousands)
(Nicholson 1957). The peaks marked
with * are the superpeaks.
614 S. Rinaldi, M. Candaten and R. Casagrandi
Ó2001 Blackwell Science Ltd/CNRS
Panel A of Fig. 3 shows the pattern of the abundance of
the Canadian lynx in the MacKenzie River Region over
almost one century (Elton & Nicholson 1942). The data are
estimates of fur returns obtained from the archives of the
Hudson's Bay Company. Similar data, available for the
period 1913±40, have not been considered because they
come from different archives and most likely are in a ®xed
but unknown ratio with the data of Fig. 3(A). Schaffer (1984)
was the ®rst to suggest that the randomness of the lynx time-
series could actually be low-dimensional chaos. After that,
various methods have been applied to these data with
somewhat controversial results. By contrast, the eight points
of our PPP align pretty well on a regular curve, thus
suggesting the existence of a low-dimensional strange
attractor. Moreover, the shape of the dotted curve in
Fig. 3(A) recalls those of Figs 1(D) and 2(A), which have
been obtained from food chain models. This is not surprising
because the most convincing interpretations of the 10-year
lynx cycle are indeed based on plant±herbivore±carnivore
models (Gamarra & Sole 2000; King & Schaffer 2001).
Figure 3(B) shows Veilleux's data (Veilleux 1979) on the
evolution of a protozoan population (Paramecium aurelia) in
the presence of its predator population (Didinium nasutum).
During the experiment the protozoa were fed periodically
every 2 days and the return times were roughly double. If
the two peaks marked with h are considered as false peaks
and, thus, eliminated and the ®rst peak is sacri®ced (as
suggested by Veilleux himself) the seven points of the PPP
nicely lie on a curve shaped as that of Fig. 2(B), obtained
from a prey±predator model with periodic food supply.
Panel C of Fig. 3 reports an 80-week time-series of larvae
of the ¯our beetle Tribolium castaneum (®fth time-series of
Fig. 2 in Costantino et al. 1997). In this case the PPP is
composed of 12 points roughly distributed on a closed
curve. This suggests that the regime is quasiperiodic or
weakly chaotic, in good agreement with the estimate of an
almost vanishing Lyapunov exponent given in Costantino
et al. (1997).
Finally, Fig. 3(D) reports the outcome of Nicholson's
long experiment on the sheep blow¯y Lucilia cuprina
(Nicholson 1957). Ups and downs of abundance vary
remarkably in amplitude and multiple local peaks, attributed
to time-delays in the reproductive cycle (Gurney et al. 1980),
often characterize periods of high abundance. If we eliminate
the ®rst (transient) superpeak (marked with s) and we retain
the other superpeaks (marked with *) we obtain eight points
in the space (yi, yi+1), which suggest that the attractor is a low-
dimensional strange attractor. It is worth noticing that these
eight points are not enough to fully evidence the geometry of
the complex PPP of the supporting model (see Fig. 2D). A
similar result holds for Nicholson's short experiment
(Nicholson 1954). All this supports the idea that a simple
recursive relationship regulates the superpeaks and compares
quite favourably with the outcome of some statistical
methods, which suggest that these time-series are almost
chaotic (Ellner & Turchin 1995).
C O N C L U S I O N
Our analysis has revealed that PPD are almost the rule in
standard ecological models (indeed, we are now aware of 25
different ecological models with PPD). Moreover, the crude
analysis of four of the longest ®eld and laboratory
population time-series suggests that PPD are quite plausible,
and the qualitative comparison with some supporting
models reinforces this conclusion. These results are in
agreement with the ®ndings of many authors about the role
of density dependence mechanisms in the dynamics of
ecosystems.
Once PPDs have been identi®ed, the population can be
described by a very simple reduced-order model involving
only the peaks of the variable of concern (see Eqs. (4) and
(5)) or the corresponding return times (see Eqs. (6) and (7)).
These models can be of paramount importance in applica-
tions, because high population peaks are often associated
with undesirable consequences, which can be attenuated if
predicted in good time. In this respect, it is worth noticing
that if one has a reliable mathematical model of the standard
form dx/dt � f(x), or equivalently x(t + 1) � f(x(t)), the
forecast of the forthcoming peak requires the knowledge of
the state x of the system at a given initial time. But this is
never the case in real situations. For example, one may know
the lynx population in a given year, but not the hare
population and/or the density of grass and plants. By
contrast, PPD models are in the form (4)±(7) and can
therefore be used even if the available information is
restricted to the peaks of the variable of interest.
In closing this paper it is fair to make a few comments on
the identi®cation of PPD from real, i.e. noisy, data. The
technical literature on this matter clearly indicates that,
without the support of a model, hundreds of peaks are
needed for estimating the invariants (Eckmann et al. 1986;
Farmer & Sidorowich 1987; Casdagli 1991; Kennel
& Isabelle 1992; Mitschke & Damming 1993; Aguirre
& Billings 1995), a condition which is far from being
satis®ed by ecological time-series. Hence, it is mandatory to
make a combined use of data and models if the target is the
identi®cation of the PPD of a given population. A very
quick, but also very naõÈve way of proceeding has been
indicated in this paper, where the supporting models have
been used only for a qualitative consistency check. Better
estimates of PPD should be obtained by ®xing at realistic
values all parameters of the supporting model, except a
couple of strategic parameters which could then be varied to
®nd the best ®t of the simulated PPP with the data. This
calibration procedure is certainly justi®ed if the aim is to
Peak-to-peak dynamics 615
Ó2001 Blackwell Science Ltd/CNRS
build up an operational forecasting technique based only on
information on past peaks. For improving the estimates of
the PPD one should also consider the possibility of
pre®ltering the data, as there is here a very speci®c feature,
namely the fact that the structure of the PPP is sensitive to
high frequency noise (see discussion on false vs. true peaks).
General and effective pre®ltering rules that might counter-
act this peculiar weakness are hard to discover, but the
challenge is worth trying. By contrast, in studying speci®c
populations it might be possible to ®nd ways to circumvent
speci®c obstacles. This is, for example, what has been done
in Rinaldi & Solidoro (1998), where the impact on PPD
identi®ability of the plankton sampling period and of
measurement noise and weather variability has been
evaluated, together with the ef®ciency of various pre®ltering
algorithms.
A C K N O W L E D G E M E N T S
This study has been ®nancially supported by Consiglio
Nazionale delle Ricerche, Project ST/74 `Mathematical
Methods and Models for the Study of Biological Phenomena'.
The authors are grateful to Alessandra Gragnani and Carlo
Piccardi for their help and advice. Three anonymous
reviewers and Mercedes Pascual are also acknowledged for
their constructive criticisms.
R E F E R E N C E S
Abrams, P.A. & Matsuda, H. (1997). Prey adaptation as cause of
predator-prey cycles. Evolution, 51, 1742±1750.
Aguirre, L.A. & Billings, S.A. (1995). Retrieving dynamical invari-
ants from chaotic data using NARMAX models. Int. J. Bifurcation
Chaos, 5, 449±474.
Blasius, B., Huppert, A. & Stone, L. (1999). Complex dynamics and
phase synchronization in spatially extended ecological systems.
Nature, 399, 354±359.
Blasius, B. & Stone, L. (2000). Chaos and phase synchronization in
ecological systems. Int. J. Bifurcation Chaos, 10, 2361±2380.
Boer, M.P., Kooi, B.W. & Kooijman, S.A.L.M. (1998). Food chain
dynamics in the chemostat. Mathemat. Biosciences, 150, 43±62.
Candaten, M. & Rinaldi, S. (2000). Peak-to-peak dynamics: a critical
survey. Int. J. Bifurcation Chaos, 10, 1805±1819.
Casdagli M. (1991). Chaos and deterministic versus stochastic non-
linear modelling. J. Royal Statist. Soc. Series B, 54, 303±328.
Casdagli, M. (1989). Nonlinear prediction of chaotic time series.
Physica D, 35, 335±356.
Caswell, H. & Weeks, D.E. (1986). Two-sex models: Chaos,
extinction, and other dynamic consequences of sex. Am. Naturalist,
128, 707±735.
Costantino, R.F., Cushing, J.M., Dennis, B. & Desharnais, R.A.
(1995). Experimentally induced transitions in the dynamic
behaviour of insect populations. Nature, 375, 227±230.
Costantino, R.F., Desharnais, R.A., Cushing, J.M. & Dennis, B.
(1997). Chaotic dynamics in an insect population. Science, 275,
389±391.
Eckmann, J.P., Kamphorst, S.O., Ruelle, D. & Ciliberto, S. (1986).
Liapunov exponents from time series. Phys. Rev. A, 34, 4971±
4979.
Ellner, S. & Turchin, P. (1995). Chaos in a noisy world: new
methods and evidence from time-series analysis. Am. Naturalist,
145, 343±375.
Elton, C. & Nicholson, M. (1942). The ten-year cycle in numbers
of the lynx in Canada. J. Anim. Ecol., 11, 215±244.
Farmer, J.D. & Sidorowich, J.J. (1987). Predicting chaotic time
series. Phys. Rev. Lett., 59, 845±848.
FerrieÁre, R. & Gatto, M. (1995). Lyapunov exponents and the
mathematics of invasion in oscillatory and chaotic populations.
Theoret. Population Biol., 48, 126±171.
Funasaki, E. & Kot, M. (1993). Invasion and chaos in a periodically
pulsed mass-action chemostat. Theoret. Population Biol., 44, 203±
224.
Gamarra, G.P. & SoleÂ, R.V. (2000). Bifurcations and chaos in
ecology: Lynx returns revisited. Ecol. Lett., 3, 114±121.
Gragnani, A., Scheffer, M. & Rinaldi, S. (1999). Top-down control
of cyanobacteria: a theoretical analysis. Am. Naturalist, 153,
59±72.
Guckenheimer, J., Oster, G. & Ipaktchi, A. (1977). The dynamics
of density dependent population models. J. Mathemat. Biol., 4,
101±147.
Gurney, W.S.C., Blythe, S.P. & Nisbet, R.M. (1980). Nicholson's
blow¯ies revisited. Nature, 287, 17±21.
Hanski, I., Turchin, P., KorpimaÈki, E. & Henttonen, H. (1993).
Population oscillations of boreal rodents: Regulation by mustelid
predators leads to chaos. Nature, 364, 232±235.
Hastings, A., Hom, C.L., Ellner, S., Turchin, P. & Godfray, H.C.J.
(1993). Chaos in ecology: Is Mother Nature a strange attractor?
Annu. Rev. Ecol. Syst., 24, 1±33.
Hastings, A. & Powell, T. (1991). Chaos in a three-species food
chain. Ecology, 72, 896±903.
Higgins, K., Hastings, A. & Botsford, L.W. (1997). Density
dependence and age structure: nonlinear dynamics and popu-
lation behavior. Am. Naturalist, 149, 247±269.
Kennel, M.B. & Isabelle, S. (1992). Method to distinguish possible
chaos from colored noise and to determine embedding param-
eters. Phys. Rev. A, 46, 3111±3118.
King, A.A. & Schaffer, W.M. (2001). The geometry of a population
cycle: a mechanistic model of snowshoe hare demography.
Ecology, 82, 814±830.
Lauwerier, H.A. & Metz, J.A.J. (1986). Hopf bifurcation in
host-parasitoid models. IMA J. Mathemat. Appl. Med. Biol., 3,
191±210.
Lorenz, E.N. (1963). Deterministic nonperiodic ¯ow. J. Atmospheric
Sci., 20, 130±141.
McCann, K. & Hastings, A. (1997). Re-evaluating the omnivory-
stability relationship in food webs. Proc. Roy. Soc. Lond., Series B,
264, 1249±1254.
McCann, K., Hastings, A. & Huxel, G.R. (1998). Weak trophic
interactions and the balance of native. Nature, 395, 794±798.
Mitschke, F. & Damming, M. (1993). Chaos versus noise in
experimental data. Int. J. Bifurcation Chaos, 3, 693±702.
Nicholson, A.J. (1954). An outline of the dynamics of natural
populations. Aust. J. Zool., 2, 9±65.
Nicholson, A.J. (1957). The self adjustment of populations to
change. Cold Spring Harbour Symp. Quantitative Biol., 22,
153±173.
616 S. Rinaldi, M. Candaten and R. Casagrandi
Ó2001 Blackwell Science Ltd/CNRS
Olsen, L.F. & Degn, H. (1985). Chaos in biological systems. Q. Rev.
Biophysics, 18, 165±225.
Olsen, L.F., Truty, G.L. & Schaffer, W.M. (1988). Oscillations and
Chaos in epidemics ± a nonlinear dynamic study of 6 childhood
diseases in Copenhagen, Denmark. Theoret. Population Biol., 33,
344±370.
Packard, N., Crutch®eld, J., Farmer, J. & Shaw, R. (1980).
Geometry from a time series. Phys. Rev. Lett., 45, 712±716.
Ricker, W.E. (1954). Stock and recruitment. J. Fish Res. Board
Canada, 11, 559±623.
Rinaldi, S., Casagrandi, R. & Gragnani, A. (2001a). Reduced order
models for the prediction of the time of occurrence of extreme
episodes. Chaos, Solitons Fractals, 12, 313±320.
Rinaldi, S., Muratori, S. & Kuznetsov, Y. (1993). Multiple attrac-
tors, catastrophes and chaos in seasonally perturbed predator
prey communities. Bull. Mathemat. Biol., 55, 15±35.
Rinaldi, S., Piccardi, C. & Gragnani, A. (2001b). Pest outbreaks
control: The approach of peak-to-peak dynamics. Natural
Resource Modeling, 14, 177±195.
Rinaldi, S. & Solidoro, C. (1998). Chaos and peak-to-peak dynamics
in a plankton-®sh model. Theoret. Population Biol., 54, 62±77.
Schaffer, W.M. (1984). Stretching and folding in lynx fur returns:
Evidence for a strange attractor in Nature? Am. Naturalist, 124,
798±820.
Schaffer, W.M. (1985). Order and chaos in ecological systems.
Ecology, 66, 93±106.
Schaffer, W.M. & Kot, M. (1985a). Do strange attractors govern
ecological-systems? Bioscience, 35, 342±350.
Schaffer, W.M. & Kot, M. (1985b). Nearly one dimensional
dynamics in an epidemic. J. Theoret. Biol., 112, 403±427.
Schimz, A. & Hildebrand, E. (1992). Nonrandom structures in the
locomotor behavior of Halobacterium: a bifurcation route to
chaos. Proc. Natl. Acad. Sci., 89, 457±460.
Scott, S.K. (1994). Oscillations, Waves and Chaos in Chemical Kinetics.
Oxford University Press, Oxford.
Shaw, R. (1984). The Dripping Faucet as a Model Chaotic System. Aerial
Press, Santa Cruz, CA.
Strogatz, S.H. (1994). Nonlinear Dynamics and Chaos. Studies in Non-
linearity. Addison-Wesley, Reading, MA.
Sugihara, G. & May, R.M. (1990). Nonlinear forecasting as a way of
distinguishing chaos from measurement error in time series.
Nature, 344, 734±741.
Turchin, P. & Ellner, S. (2000). Living on the edge of chaos:
Population dynamics of Fennoscandian voles. Ecology, 81, 3099±
3116.
Turchin, P. & Hanski, I. (1997). An empirically based model for
latitudinal gradient in vole population dynamics. Am. Naturalist,
149, 842±874.
Veilleux, B.G. (1979). An analysis of the predatory interaction
between Paramecium and Didinium. J. Anim. Ecol., 48, 787±803.
B I O S K E T C H
Sergio Rinaldi is interested in theoretical population dynamics.
At the moment he is involved in studies on tritrophic food
chains, plankton dynamics and forest ecosystems.
Editor, M Pascual
Manuscript received 7 May 2001
First decision made 18 May 2001
Second decision made 8 August 2001
Manuscript accepted 30 August 2001
Peak-to-peak dynamics 617
Ó2001 Blackwell Science Ltd/CNRS