L ETTER

Regime shifts in ecological systems can occur with no

warning
Alan Hastings
1
* and Derin B.
Wysham
1,2
1
Department of Environmental
Science and Policy, University of
California, Davis, CA 95616, USA
2
Department of Computational
and Systems Biology, John Innes
Center, Norwich NR4 7UH, UK
*Correspondence: E-mail:
amhastings@ucdavis.edu
Abstract
Predicting regime shifts drastic changes in dynamic behaviour is a key challenge in
ecology and other elds. Here we show that the class of ecological systems that will
exhibit leading indicators of regime shifts is limited, and that there is a set of ecological
models and, therefore, also likely to be a class of natural systems for which there will be
no forewarning of a regime change. We rst describe how nonlinearities in combination
with environmental variability lead to model descriptions that will not have smooth
potentials, concluding that many ecological systems are described by systems without
smooth potentials and thus will not show typical leading indicators of regime shifts. We
then illustrate the impact of these general arguments by numerically examining the
dynamics of several model ecological systems under slowly changing conditions. Our
results offer a cautionary note about the generality of forecasting sudden changes in
ecosystems.
Keywords
Chaos, dynamics, potential, predatorprey, regime shifts, Ricker model.
Ecology Letters (2010) 13: 464472
I NT RODUCT I ON
A key problem in ecology and population biology is both to
understand and to predict the sudden shifts that occur in a
variety of ecological systems (Scheffer et al. 2001; Conners
et al. 2002; Bond et al. 2003; Foley et al. 2003; Scheffer &
Carpenter 2003; Scheffer & van Nes 2004; van Nes &
Scheffer 2005; Brock & Carpenter 2006; Carpenter & Brock
2006; Guttal & Jayaprakash 2007, 2008; Carpenter et al.
2008; Biggs et al. 2009; Contamin & Ellison 2009; Chisholm
& Filotas 2009; Takimoto 2009). Current research has
focused on the possibility of determining leading indicators
that would precede these regime shifts in both ecological
and other more general settings (Scheffer et al. 2009). There
is a wide variety of natural systems that can exhibit sudden
changes in their state (Folke et al. 2004), with dramatic shifts
possible in both the levels of different variables, such as
vegetation, or shifts in the kinds of observed dynamics.
Often the sudden changes lead from a more desirable
ecosystem state to a less desirable one, such as eutrophi-
cation in lakes (Carpenter 2005; Amemiya et al. 2007) or
desertication (Guttal & Jayaprakash 2007). In most cases, it
would be benecial to predict when a change is going to
occur before it does so, in order to allow for preventative
management solutions or for ameliorative steps to lessen
impacts.
Studies (Brock & Carpenter 2006; Carpenter & Brock
2006; Carpenter et al. 2008) that have tried to determine
likely leading indicators of regime shifts have typically begun
with a description of an ecosystem using a dynamic model
that includes both slowly changing underlying conditions
(a slowly changing parameter) and some form of stochas-
ticity. It has been argued that there is an increase in variance
(Carpenter & Brock 2006) or skew (Guttal & Jayaprakash
2008, 2009) or slowing down in dynamics (Chisholm &
Filotas 2009) prior to a regime shift. The change in variance
would be an increase in the variance around the mean
population size or some other measure, a change in skew
would be a change in the third statistical moment, and a
slowing in dynamics would be reected by a shift in the
power spectrum toward lower frequencies. In some
instances the presence of leading indicators has been
supported by the behaviour of entire distributions of
solutions. We suggest that real systems only present one
case history for study, so any successful leading indicator
that would be used in practice must be able to accurately
predict a regime shift from the results using only one initial
condition in a single model simulation.
Ecology Letters, (2010) 13: 464472 doi: 10.1111/j.1461-0248.2010.01439.x
2010 Blackwell Publishing Ltd/CNRS
We take a different approach here and argue that for
classes of ecological systems which exhibit several important
features, including complex dynamics and multiple possible
outcomes, the search for leading indicators of regime shifts
will generally fail. Our argument rst begins with the
underlying mathematical reasons for the lack of leading
indicators of regime shifts for this class of model systems.
Clearly, our general approach does not rule out the presence
of leading indicators of regime shifts in those cases where
natural systems are well described by models that have
smooth potentials, and our work therefore suggests the
importance of a better understanding of which models of
ecological systems can provide a reasonable description of
natural dynamics.
We then conrm and illustrate the predicted lack of
leading indicators in a series of examples. In these models,
we make two kinds of small changes in the parameters that
produce regime shifts: we either change underlying para-
meters by a small amount or we vary the size of the
(stochastic) perturbations. Although we clearly only present
simulation results for a selected set of models, we argue that
these models are a fairly general representation of many
ecological systems. Since essentially any non-spatial model
with strong enough overcompensatory density dependence
will show the period doubling cascade to chaos and any
spatially coupled model will inherit this trait, the spatially
explicit, overcompensatory (i.e. period doubling) models we
choose here are in fact adequate, or at least plausible,
characterizations of a large set of real ecological settings. In
none of the cases we simulate are the proposed leading
indicators of regime shifts found. Obviously preceding any
regime shift there has to be some form of change in
dynamics, but here these do not appear to be those that
have been used to predict a regime shift.
GE NE RAL ARGUME NT
Suggestions that there are leading indicators of regime shifts
have been largely based on the assumed properties of a
potential (Graham & Tel 1984), which is essentially the
mathematical abstraction of the physical notion of potential
energy and acts as a guide for the type of expected
dynamics. It is typically represented by ball and landscape
diagrams such as found in Hastings (1997, p. 123). If the
potential exists, depends smoothly on the underlying state of
the ecosystem, and changes smoothly as underlying condi-
tions change, then the prior arguments for the existence of
leading indicators of regime shifts (summarized in Scheffer
et al. 2009) are valid and in fact have been conrmed
empirically. Simple ecological models based on Lotka-
Volterra equations can satisfy the conditions for the
existence of a smooth potential, and in these cases leading
indicators are found.
However, models for ecological systems can easily exhibit
complex dynamics, and systems like this do not, in general,
have smooth potentials (Graham & Tel 1984, 1986; Graham
et al. 1991). In many cases the description of a system
undergoing a regime shift in the terminology used by
ecologists is the one which undergoes a crisis (Grebogi et al.
1983), which is dened to be a sudden change in (long-term)
system behaviour as a parameter is varied. For most types of
crises to occur, a system must be capable of producing
complex dynamics and thus will not have a smooth
potential. For example, if a chaotic attractor exists, then
the potential has a non-smooth fractal structure (down to
the level of the noise) similar to the fractal structure of the
attractor (Graham et al. 1991). These are precisely the types
of systems for which the presence of leading indicators have
not been studied, as indicated in (Scheffer et al. 2009).
The models that display chaotic or complex dynamics are
only a subset of the systems for which the potential will not
be smooth. If multiple attractors of any type are present,
which is exactly the situation often modelled when looking
to understand regime shifts, then the potential is likely to be
non-smooth precisely in the transition area between two
attractors. The physical meaning of the points where the
potential has discontinuous derivatives is that the stochastic
effects, which must be included in any realistic ecosystem
model, causes there to be several most probable paths the
system may take out of the bottom of any local potential
well. The consequence of only piecewise smooth potentials
for a system is that we would not in general expect that
there would be obvious precursors to a regime shift (or
crisis) as parameters are varied.
The reason for the discrepancy between our results and
those that have been previously reported is subtle. If one
writes down a given model, derives a potential, concludes it
is smooth, and then adds noise during simulation, then this
invariably points to the probable existence of leading
indicators. Whereas, if one writes down a model that
includes a stochastic term, and then derives the (non-
equilibrium) potential using the methodology as presented
in (Graham & Tel 1984), then one will see that this potential
is in fact generically non-smooth and that the proposed
leading indicators will generally no longer apply.
SPECI F I C MODEL S AND MET HODS
We numerically simulate several models that are likely to
lack a smooth potential in order to illustrate the character-
istics of systems prior to, or during, regime shifts. Our
general argument shows there are broad classes of ecological
systems that we would expect to lack a smooth potential and
therefore we would expect to fail to have leading indicators
of regime shifts, namely those that have either multiple
attractors or have at least the possibility of chaotic
Letter Regime shifts with no warning 465
2010 Blackwell Publishing Ltd/CNRS
attractors. The failure to nd strong evidence for chaos in
natural systems does not invalidate the importance of our
argument, as chaotic systems that have arbitrarily small,
positive Liapunov exponents would be almost impossible to
distinguish from non-chaotic ones, yet they would still
typically not have leading indicators. Furthermore, for two
of the three systems we investigate here the Liapunov
exponents are all non-positive, and they take exceptionally
small, positive values in the remaining model.
We use three models that possess the capacity for
complex dynamics in order to illustrate the lack of leading
indicators of regime shifts. The rst model is a single-
species system of eight patches with local population
dynamics given by the Ricker equation and with an
alternation between local population dynamics and passive
diffusion. We make incremental changes to an underlying
parameter, and we indicate the global behavioural changes
that result. The second model is an extension of the rst,
although here space is continuous, leading to a simple
integro-difference model. The third is a three species food
chain, considered both at a single location, and in three or
six locations coupled by passive diffusion. For both these
last two models, we look at regime changes that are a
function of increasing perturbation strength. Our approach
in every case is to simulate the system, and compute the
variance and skew in one or more population levels through
time. We also use the power spectrum to investigate the rst
model. The power spectrum for a signal f (t ) is computed by
first approximating f as a truncated Fourier series:
f t
X
N
kN
c
k
e
2pikt
The size of the Fourier coefcients c
k
then gives the amount
of power that the signal has at each frequency from k = 0,
..., N. The theory of leading indicators (Scheffer et al. 2009)
has thus far concluded that immediately prior to a shift, the
size of the c
k
for low values of k should increase.
In the rst model for the local dynamics, we use a Ricker
map, and consider the case where local population dynamics
alternates with diffusive movement among eight coupled
spatial locations,
~
N t N
i
t e
a1N
i
t z
t

N
i
t 1
X
8
j 1
a
ij
~
N
j
t
1
Here N
i
(t ) is the population size at location i and time t,
~
N t
is the population size after reproduction, but before dis-
persal, z
t
is a uniformly distributed random variable in
()0.001,0.001), and a is a positive constant. The parameters
a
ji
describe the passive dispersal among the patches and are
uniformly distributed in (0.02,0.03) when z
t
> 0, and dis-
tributed in (0.01,0.02) when z
t
< 0.
We will present results below for this rst model in
Fig. 1. The model, as expected, can produce complex
dynamics, and in fact chaos. The parameter a = 2.216 for
Fig. 1e,f, and a = 2.228 for and the remaining panels of
Fig. 1. This small increase in the value of a increases the
dominant Liapunov exponent from 3 10
)2
to
3 10
)1
and results in the destruction of a three-part
chaotic attracting region, one component of which is clearly
visible in the rst 200 generations of Fig. 1a,b.
The second model is an extension of the rst, with an
alternation between local population dynamics given by the
Ricker model with movement in continuous space,
described by a Gaussian dispersal kernel f (x, y). N(x,t ) is
now a density function on x for the population size at time t,
z
t
is now randomly distributed in (-r,r). We scale the
suitable habitat to be of length one, with all individuals
falling outside the suitable habitat dying. The results will be
presented below in Fig. 2.
The third model is a three species food chain in
continuous time with Type II functional responses at a
single spatial location (McCann & Yodzis 1994), or at three
or six locations coupled by movement. The model is made
stochastic by adding a perturbation to the carrying capacity
each time the rst level consumer passes through a Poincare
section. For a perturbation strength r of approximately
r = 0.0045, the top level predator will always eventually
become extinct (see Fig. 3).
As a way to compare with earlier results and demonstrate
the lack of leading indicators we compute possible leading
indicators for our models. For our rst two models, we
compute the variance and skew (Figs 1cf and 3b,d,e), and
the power spectrum in Fig. 1g,h by passing sliding windows
of varying widths over the time series, computing either the
statistical moments or the Fourier representation from the
data within that window. The species within the food web
model follow nearly cyclic patterns. For this reason, we used
simulations in which the top predator survived for a
minimum number of cycles in order to compute the leading
indicators in Fig. 4, and the data in Fig. 5 was computed
using a sliding window of width ten cycles.
S I MUL AT I ON RES UL T S
The systems we consider here always fail to show leading
indicators of a regime shift. Simulations of the eight-patch
Ricker system in Fig. 1a,b show a period of nearly laminar
motion for 200 generations before the population densities
unexpectedly shift to a much more erratic pattern, and
there is no accompanying change in either the variance or
the skew before the abrupt change (Fig. 1c,d). In fact, the
466 A. Hastings and D. B. Wysham Letter
2010 Blackwell Publishing Ltd/CNRS
Figure 1 Populations, both at one location and summed over the entire lattice (eight locations), the computed values of the variance and
skew for parameters that give both bursting and non-bursting dynamics, and the power spectrum for the coupled Ricker model. (a) Five
simulations showing the single population N
1
(t ) for 200 generations before a burst away from the attracting region and for 48 generations
thereafter. Every third generation is plotted. (b) The total population across the lattice for the same five simulations as in (a). (c) The variance
(upper curve in each case) and skew (lower curve) of N
1
(t ) for the first 40 windows of each of the five simulations in (a). Computations were
carried out using a sliding window of width 10 generations. (d ) As (c), but the variance and skew are shown for the last 50 windows before
bursting ensues. In (e, f), For comparison with (c, d ) the variance (upper curve, each panel) and skew (lower curves) computed over sliding
windows of width 10 generations for each of five simulations for which there is no bursting. In (g, h), the power spectrum for one of the time
series in (a). The height of the surface for each k gives the amount of that frequency that is found in the windowed time series (height colour
coded by the colour bar). We see that there is no shift in the spectrum to blue frequencies (smaller k) as the system approaches regime change.
The magnitudes of the Fourier coefficients c
k
for small k are approximately the same for the first 40 windows of length 32 generations (left
panel) as they are for the last 40 generations before the shift (right panel). If anything, the power corresponding to longer wavelengths
(smaller k) has decreased over the course of the time series.
Letter Regime shifts with no warning 467
2010 Blackwell Publishing Ltd/CNRS
magnitudes of the variance and skew, as well as their
temporal pattern (due to the models underlying near-
periodicity), are nearly identical in the generations that
immediately precede the shift to the magnitudes and
patterning found some 200 generations prior (Fig. 1c,d).
No parameters were changed during the course of the
simulations shown in Fig. 1ad, and at future times the
densities settle back into the simpler dynamic until another
abrupt burst ensues. Further, if the growth rate is reduced
by only the smallest fraction (less than 0.4%), then the
variance and skew change only imperceptibly, as seen if
one compares Fig. 1c,d to Fig. 1e,f, and yet the popula-
tions now behave predictably for all time, i.e. the slight
reduction in the growth rate causes any time series to
forever appear nearly identical to the rst 200 generations
of Fig. 1a,b. We have also included computations of the
power spectrum, both for the initial 40 generations of the
time series and for the 40 generations immediately prior to
the burst, showing that there is no shift to lower
frequencies as the regime shift is approached (Fig. 1e,f).
This rst example is particularly important because the
dynamics represent intermittency, a common behaviour in
coupled systems, and yet all the leading indicators fail to
warn of the shift. We note that intermittency is a dynamical
behaviour that in many ways corresponds to the somewhat
irregular behaviour of insect outbreaks in space and time
(kland & Bjrnstad 2006).
Similarly, the values of the variance and skew give no
prior indication of immanent shifts in the continuous-space
model (Fig. 2). Here, the parameters are chosen to be close
to values for which the system has sustained complex
dynamics, but the asymptotic behaviour is simple at low
perturbation strengths (Fig. 2a). The system is highly
effected by the presence of the nearby (in parameter space)
Figure 2 Simulations showing regime shifts
in the total population for the Hastings
Higgins model. Shifts are marked with vertical
blue lines. (a) Aregime shift in the presence of
small external perturbation(r = 0.01) occurs,
and wildly oscillatory behaviour is replaced by
nearly periodic motion. (b) The standard
deviation (square root of the variance) plotted
in black, green, and skewshown in red, purple
for windows of widths 50 and 10, respectively.
(c) Multiple regime shifts occur inthe presence
of large noise (r = 0.1), as the perturbation
strength is strong enough to cause attractor
switching. (d) The variance andskewshownin
the same format as in (b), but around the rst
large shift in (c). (e) The variance and skew
around the second shift in (c).
468 A. Hastings and D. B. Wysham Letter
2010 Blackwell Publishing Ltd/CNRS
complex motion, and mimics those dynamics for long
ecological times. The type of oscillations preceding the
eventual relaxation to a more predictable pattern produces
constant uctuations of varying magnitudes in the variance
and skew (Fig. 2b,d), and so there is little evidence that
these measurements can give a true forecast of a shift. We
do see that the variance reaches a relative maximum right as
the shift occurs; however, larger maxima had been reached
previously in the time series when no shift took place. Also,
when the perturbations are larger, stochastic effects will
push the system back into the regime of complex patterning
and large oscillations (Fig. 2c), and the existence or nal
selection of any asymptotic state depends upon the noise
strength. As shown in Fig. 2e, the time frame preceding the
transitions from one state to another does not show any
leading indicators either, and it is only after the regime
change has occurred that one notices any change in the
variance or skew.
The data in Figs 4 and 5 is conrmation that leading
indicators do not necessarily exist even for the most
dramatic class of regime shift. The three-species food chain
modelled here is spatially explicit, with one, three or six
patches, and the parameters are chosen to satisfy constraints
reecting biological realism (McCann & Yodzis 1994), so
that with no environmental uctuations the top predator
exists indenitely in all patches. As the strength of the
perturbations is slowly increased, there is a critical value
after which the top predator will always eventually become
extinct (Fig. 3). This is a case where the environmental
forcing is undergoing an increase in variance, and yet we see
in Fig. 4 that there are only generally small, random
uctuations in the values of the populations variance and
skew before the critical perturbation strength is surpassed,
and while occasional, nearly linear trends in the data do exist
there is no abrupt increase in leading indicators as the
critical perturbation strength is approached or surpassed.
There is also no clear trend in the data when we make
different choices of the migration rate. The conclusion is
that the proposed leading indicators once again fail to
predict a regime shift.
DI SCUS S I ON
The potential for regime shifts has been well recognized
both empirically and in models. However, whether and how
an impending shift can be predicted in advance has been the
subject of intense recent study. A number of recent papers
have proposed different leading indicators (Takimoto 2009),
primarily based on careful studies of particular models. Yet,
the search for universal indicators of impending regime
z
z z
x x
x x
y y
y y
0.8
1
0.2
1
0.6
0.2
0.6
1
0.2 0.2
0.2
0.6
0.5 0.5 0.5
0.2
0.8 0.8
0 0 0
0 0 0 0 0 0
z z
1 1 1
1 1 1
1 1 1
Patch 1 Patch 2 Patch 3
y
x
z
x
y
Figure 3 Phase portraits of species densities in each of three patches with r = 0.0025 and no predator collapse (top row), and with
r = 0.0075 and assured eventual predator collapse (bottom row).
Letter Regime shifts with no warning 469
2010 Blackwell Publishing Ltd/CNRS
S
k
e
w
Species y Species z
0 0.01 0.01 0.01
0.01 0 0.01
0.085
0.075
0.065
0.055
0.075
0.065
0.055
0.045
0.24
0.2
0.16
0.12
0
0 0.005 0.005
3.3
3.1
2.9
2.7
2.5 5.4
5.2
5
4.8
4.6
0.005 0.005 0.005 0
0.01 0.005 0
0
0.5
1
1.5
2
2.5

Species x
V
a
r
i
a
n
c
e
Figure 4 Variance and skew as a function of perturbation strength for each of the three species in the food web. Each data point represents
the average value over ve simulations for which the top predator survived for at least 25 annual cycles. Colours and line types correspond to
different choices of the number of patches and coupling strengths: a single patch (black, solid); three patches (orange, light green, light blue
and grey, solid lines) with coupling taken to be 10
-4
, 5 10
)4
, 10
)3
and 5 10
)3
, respectively and six patches (red, dark green, dark blue and
black, dashed lines) with the same respective values of the coupling.
0.035
0.045
0.015
0.055 0.6
10 20 30 40 50 10 50 40 30 20
0.6
0.4
0.2
0.8
y
z
x
0.6
0.4
0.2
V
a
r
i
a
n
c
e
S
k
e
w
1
0.2
0.6
0.2
1
0.8
0.6
0.025
Figure 5 The variance (left panel) and skew (middle) for single simulations with varying dispersal and perturbation strength. Shown are the
data for three-patch systems with = 0.001, r = 0.005,0.0075 (species x: light blue; y: light green; z: orange; solid, dashed, respectively), and
= 0.0001, r=0.005,0.0075 (species x: dark blue; y: dark green; z: red; solid, dashed, respectively). The data for which the trajectories have
left the attractor but for which the top predator has not yet begun a monotonic decrease are shown in black, and the right panel is a phase
space representation showing the same portion of the trajectory in black for = 0.0001, r = 0.005.
470 A. Hastings and D. B. Wysham Letter
2010 Blackwell Publishing Ltd/CNRS
shifts has proved difcult. Therefore we have used a
different approach, one that utilizes general dynamical
systems arguments and starts from the viewpoint that typical
ecological models are likely to embody nonlinear complex-
ities that arise from density dependence and strong
interspecic interactions. This approach implies that any
underlying potential for an ecosystem cannot a priori be
assumed to be smooth.
Non-equilibrium potential theory implies that ecological
systems which can exhibit complex dynamics or multiple
outcomes will typically not have leading indicators of regime
shifts. Since ecosystems with complex dynamics are certainly
at least some of the ones likely to exhibit regime shifts, drastic
changes can appear in nature without warning. Althoughthere
have been surveys that have suggested that chaos in particular
is uncommon in ecological systems, this work has been based
on the failure to nd positive Lyapunov exponents (Hastings
et al. 1993). However, since many ecological models, such as
the food chain model (Hastings & Powell 1991) we simulate
here have positive Lyapunov exponents on the order of 10
)2
it
will be very difcult to demonstrate positive Lyapunov
exponents in natural systems. Thus, we suggest that the classes
of natural systems that would in fact fail to show leading
indicators is large, but recognize that this is to some extent an
open question.
Since it can be difcult to understand what kinds of
ecological models will typically satisfy our general condi-
tions, we have also undertaken simulations of specic
models drawn from an important and broad class to
conrm that they lack leading indicators. Using a range of
models that have strong nonlinearities, we demonstrate the
absence of leading indicators before regime shifts.
How can our results be explained in the light of work that
has shown the presence of leading indicators? Our
arguments do not contradict prior evidence (Scheffer et al.
2009) demonstrating that leading indicators are likely to be
observed in a particular class of systems, namely those with
smooth potentials. Unfortunately, it is not likely to be easy
to identify which systems in natural or managed ecosystems
are well described by models with smooth potentials. Our
work therefore does not unequivocally rule out the
appearance of leading indicators of regime shifts in a variety
of systems.
In terms of management decisions, we would argue that
rather than act on the basis of leading indicators (Biggs et al.
2009), it may often be necessary to act extremely quickly in
response to observations of shifts. Our ndings also
indicate that predicting shifts in many natural systems may
require efforts similar to the rst exit time approach that has
proved successful in other elds, but, most importantly,
much further work will be needed to understand which
classes of models can best represent natural system
dynamics.
ACKNOWL EDGEMENT S
This work is supported by NSF grants EF 0434266 and EF
0742674. We thank the referees for their helpful comments.
REF E RENCES
Amemiya, T., Enomoto, T., Rossberg, A., Yamamoto, T., Inamori,
Y. & Itoh, K. (2007). Stability and dynamical behaviour in a lake-
model and implications for regime shifts in real lakes. Ecol.
Model., 206, 5462.
Biggs, R., Carpenter, S.R. & Brock, W.A. (2009). Turning back
from the brink: detecting an impending regime shift in time to
avert it. Proc. Natl. Acad. Sci. USA, 106, 826831.
Bond, N.A., Overland, J.E., Spillane, M. & Stabeno, P. (2003).
Recent shifts in the state of the North Pacic. Geophys. Res. Lett.,
30, 2183.
Brock, W.A. & Carpenter, S.R. (2006). Variance as a leading
indicator of regime shift in ecosystem services. Ecol. Soc., 11, 9.
Carpenter, S.R. (2005). Eutrophication of aquatic ecosystems:
bistability and soil phosphorus. Proc. Natl. Acad. Sci. USA, 102,
1000210005.
Carpenter, S.R. & Brock, W.A. (2006). Rising variance: a leading
indicator of ecological transition. Ecol. Lett., 9, 308315.
Carpenter, S.R., Brock, W.A., Cole, J.J., Kitchell, J.F. & Pace, M.L.
(2008). Leading indicators of trophic cascades. Ecol. Lett., 11,
128138.
Chisholm, R.A. & Filotas, E. (2009). Critical slowing down as an
indicator of transitions in two-species models. J. Theor. Biol., 257,
142149.
Conners, M.E., Hollowed, A.B. & Brown, E. (2002). Retrospective
analysis of Bering Sea bottom trawl surveys: regime shift and
ecosystem reorganization. Prog. Oceanogr., 55, 209222.
Contamin, R. & Ellison, A.M. (2009). Indicators of regime shifts in
ecological systems: what do we need to know and when do we
need to know it? Ecol. Appl., 19, 799816.
Foley, J.A., Coe, M.T., Scheffer, M. & Wang, G.L. (2003). Regime
shifts in the Sahara and Sahel: interactions between ecological
and climatic systems in northern Africa. Ecosystems, 6, 524539.
Folke, C., Carpenter, S., Walker, B., Scheffer, M., Elmqvist, T.,
Gunderson, L., et al. (2004). Regime shifts, resilience, and bio-
diversity in ecosystem management. Annu. Rev. Ecol. Evol. Syst.,
35, 557581.
Graham, R., Hamm, A. & Tel, T. (1991). Nonequilibrium poten-
tials for dynamic systems with fractal attractors or repellers. Phys.
Rev. Lett., 66, 30893092.
Graham, R. & Tel, T. (1984). Existence of a potential for dissi-
pative dynamical systems. Phys. Rev. Lett., 52, 912.
Graham, R. & Tel, T. (1986). Nonequilibrium potentials for
coexisting attractors. Phys. Rev. A, 33, 13221337.
Grebogi, C., Ott, E. & Yorke, J. (1983). Crises, sudden changes in
chaotic attractors, and transient chaos. Phys. D: Nonlinear Phe-
nom., 7, 181200.
Guttal, V. & Jayaprakash, C. (2007). Self-organization and pro-
ductivity in semi-arid ecosystems: implications of seasonality in
rainfall. J. Theor. Biol., 248, 490500.
Guttal, V. & Jayaprakash, C. (2008). Changing skewness: an early
warning signal of regime shifts in ecosystems. Ecol. Lett., 11,
450460.
Letter Regime shifts with no warning 471
2010 Blackwell Publishing Ltd/CNRS
Guttal, V. & Jayaprakash, C. (2009). Spatial variance and spatial
skewness: leading indicators of regime shifts in spatial ecological
systems. Theor. Ecol., 2, 312.
Hastings, A. (1997). Population Biology. Springer, New York.
Hastings, A., Hom, C., Ellner, S., Turchin, P. & Godfray, H.C.J.
(1993). Chaos in ecology: is mother nature a strange attractor?
Annu. Rev. Ecol. Syst., 24, 133.
Hastings, A. & Powell, T. (1991). Chaos in a three species food
chain. Ecology, 72, 896903.
McCann, K. & Yodzis, P. (1994). Nonlinear dynamics and popu-
lation disappearances. Am. Nat., 144, 873879.
van Nes, E.H. & Scheffer, M. (2005). Implications of spatial het-
erogeneity for catastrophic regime shifts in ecosystems. Ecology,
86, 17971807.
kland, B. & Bjrnstad, O. (2006). A resource-depletion model of
forest insect outbreaks. Ecology, 87, 283290.
Scheffer, M., Bascompte, J., Brock, W.A., Brovkin, V., Carpenter,
S.R., Dakos, V., et al. (2009). Early-warning signals for critical
transitions. Nature, 461, 5359.
Scheffer, M. & Carpenter, S.R. (2003). Catastrophic regime shifts in
ecosystems: linking theory to observation. Trends Ecol. Evol., 18,
648656.
Scheffer, M., Straile, D., van Nes, E.H. & Hosper, H. (2001).
Climatic warming causes regime shifts in lake food webs. Limnol.
Oceanogr., 46, 17801783.
Scheffer, M. & van Nes, E.H. (2004). Mechanisms for marine
regime shifts: can we use lakes as microcosms for oceans? Prog.
Oceanogr., 60, 303319.
Takimoto, G. (2009). Early warning signals of demographic regime
shifts in invading populations. Popul. Ecol., 51, 419426.
Editor, John Fryxell
Manuscript received 21 December 2009
Manuscript accepted 22 December 2009
472 A. Hastings and D. B. Wysham Letter
2010 Blackwell Publishing Ltd/CNRS