You are on page 1of 23

ES46CH07-Scheffer ARI 11 September 2015 10:58

Review in Advance first posted online


V I E W
E on September 21, 2015. (Changes may
R

still occur before final publication

S
online and in print.)

C E
I N

A
D V A

Generic Indicators of
Ecological Resilience:
Inferring the Chance
of a Critical Transition
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

Marten Scheffer,1 Stephen R. Carpenter,2


Vasilis Dakos,3 and Egbert van Nes1
1
Department of Environmental Sciences, Wageningen University, Wageningen,
The Netherlands; email: Marten.scheffer@wur.nl
2
Center for Limnology, University of Wisconsin–Madison, Madison, Wisconsin;
email: steve.carpenter@wisc.edu
3
Estación Biológica de Doñana, Consejo Superior de Investigaciones Cientificas, Seville,
Spain; email: vasilis.dakos@gmail.com

Annu. Rev. Ecol. Evol. Syst. 2015. 46:145–67 Keywords


The Annual Review of Ecology, Evolution, and resilience, tipping point, critical transitions, early warning signals, big data,
Systematics is online at ecolsys.annualreviews.org
alternative stable states
This article’s doi:
10.1146/annurev-ecolsys-112414-054242 Abstract
Copyright  c 2015 by Annual Reviews. Ecological resilience is the ability of a system to persist in the face of per-
All rights reserved
turbations. Although resilience has been a highly influential concept, its
interpretation has remained largely qualitative. Here we describe an emerg-
ing family of methods for quantifying resilience on the basis of observations.
A first set of methods is based on the phenomenon of critical slowing down,
which implies that recovery upon small perturbations becomes slower as
a system approaches a tipping point. Such slowing down can be measured
experimentally but may also be indirectly inferred from changes in natu-
ral fluctuations and spatial patterns. A second group of methods aims to
characterize the resilience of alternative states in probabilistic terms based
on large numbers of observations as in long time series or satellite images.
These generic approaches to measuring resilience complement the system-
specific knowledge needed to infer the effects of environmental change on
the resilience of complex systems.

145

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

INTRODUCTION
The term resilience is widely used to characterize the capacity to deal with perturbations. For
instance, in material science resilience is the energy that can be absorbed elastically without
creating a permanent distortion, and in psychology the term refers to the capacity of individuals
to cope with stress and adverse events. In ecology, the term resilience has been used in different
ways (Grimm & Wissel 1997). One common interpretation has been the rate of recovery upon
perturbation (DeAngelis 1980). Clearly such a recovery rate is a highly relevant feature from a
management point of view. For instance, overexploited rocky intertidal shellfish communities may
fully recover in a decade upon protection (Gelcich et al. 2010), whereas old-growth temperate
rainforests require at least the lifetime of multicentennial trees to fully recover.
Although recovery rate per se is important, an alternative and highly influential view of re-
silience was presented in this journal by C.S. Holling more than four decades ago (Holling 1973).
Holling stressed that as ecological systems could have multiple domains of attraction (Lewontin
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org

1969), the term resilience could be used to characterize the size of alternative attraction basins.
Access provided by University of Otago on 10/14/15. For personal use only.

He argued that management should focus on such resilience rather than stability and made the
point that if a domain of attraction becomes smaller, the “resilience is lost or reduced so that a
chance and rare event that previously could be absorbed can trigger a sudden dramatic change
and loss of structural integrity of the system” (Holling 1973, p. 21).
Here we focus on this interpretation of resilience and consider how we may assess such resilience
in practice. Although the original work on ecological resilience has been tremendously influential,
the literature has long remained largely conceptual. This changed with the notion that catastrophic
shifts between alternative states were more common than previously thought (Scheffer et al. 2001),
followed by the discovery of potential empirical indicators of resilience (Scheffer et al. 2009,
2012a). Because the loss of ecological resilience implies an increasing risk of a critical transition
to a contrasting state, the term early warning signals has been widely used to label the work on
indicators of resilience. As we show, recent work allows us to unify some of the earlier concepts. For
instance, recovery rate (Pimm 1984), sometimes termed engineering resilience (Holling 1996),
is actually a proxy for the size of the basin of attraction [termed ecological resilience by Holling
(1996)] in some cases. After a short primer on the theory of resilience and tipping points, we
first review different fundamental classes of resilience indicators. Subsequently, we take a practical
perspective by asking how different kinds of data can be used. Lastly, we reflect on the prospects
of this emerging field of research.

THE THEORY OF RESILIENCE AND TIPPING POINTS


IN A NUTSHELL
An extensive introduction to the theory of dynamical systems focusing on the tipping points and
related phenomena can be found in the book, Critical Transitions in Nature and Society (Scheffer
2009). What follows is a brief version highlighting only the key concepts that are referred to in
the remainder of this review.

A Simple View of Ecological Resilience


Although the theory of resilience as the size of a basin of attraction is basically mathematical
(see Supplemental Materials; follow the Supplemental Material link from the Annual Reviews
home page at http://www.annualreviews.org), its essence can easily be explained without equa-
tions. We start from the notion that the state of a complex system is typically some kind of

146 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

a d

Perturbation

b
System state
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

c
F2
F2

s
on
iti
nd
F1

Co
F1

Conditions System state

Figure 1
A classic generic model of tipping points and resilience. Although some systems (a) respond smoothly to
changing conditions, it is also common to see (b) a relatively sharp response at a critical condition. If the
curve depicting the equilibrium of a system as a function of external conditions is folded as in panel c,
alternative equilibria arise. The arrows in the graphs indicate the direction in which the system moves if it is
not on the equilibrium curve. They show that all curves represent stable equilibria, except for the unstable
equilibrium (dashed line) in panel c. If the system is pushed away a little bit from this part of the curve it will
move farther away instead of returning. Hence, equilibria on this part of the curve are unstable and represent
(d ) the border between the basins of attraction of the two alternative stable states on the upper and lower
branches. Points F1 and F2 in panels c and d are saddle-node bifurcation points, often referred to as tipping
points. Implications for the resilience of a system can be intuitively understood from stability landscapes at
different conditions (as in panel d ). One aspect of resilience is the width of the basin of attraction around an
equilibrium, which corresponds to the maximum displacement that can be taken without causing a shift to an
alternative stable state. Other aspects of resilience are the depth of the basin, which represents the energy
needed to move the system out of the basin, and the slope around the equilibrium, which represents the
return rate upon small perturbations (also known as engineering resilience). This model illustrates how all
three aspects of resilience tend to decline in concert as a system moves toward a tipping point.

equilibrium state. In response to changes in the environment, the system state may change
smoothly (Figure 1a). Commonly, however, the system is rather insensitive over certain ranges of
the external conditions yet responds relatively strongly around a threshold condition (Figure 1b).
For instance, mortality of a species usually increases sharply around some critical concentration of
a toxicant (Maltby et al. 2005). In such a situation, a strong response happens when a threshold is
passed. Such thresholds are important to understand; however, an even more interesting situation
occurs if the curve that describes the response of the equilibrium to environmental conditions is

www.annualreviews.org • Generic Indicators of Ecological Resilience 147

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

folded (Figure 1c). This folded curve implies that, for a certain range of environmental conditions,
the system has two alternative stable states separated by an unstable equilibrium that marks the
border between the basins of attraction of the alternative stable states.
When the system is in a state on the upper branch of the folded curve, it cannot pass to the
lower branch smoothly. Instead, when conditions change enough to pass the threshold (F2 ), a
catastrophic transition to the lower branch occurs (Figure 1c). The term tipping points is used
for thresholds such as F1 and F2 , where a tiny change in conditions can cause a major shift. Such
a tipping point falls under the broader mathematical class of bifurcation points—critical points in
which the attractors (and/or repellors) of a system change (Strogatz 1994). In the case of F2 , we
have a fold bifurcation, also known as a saddle-node bifurcation, in which a stable point (the upper
solid section of the curve, a node) collides with an unstable point (the dashed section, a saddle).
This is a so-called catastrophic bifurcation point in which an attractor vanishes, leaving only the
remaining alternative attractor (lower solid curve section) as a stable point for the system.
Although saddle-node bifurcation points, such as F1 and F2 illustrated here, are among the
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

best-known bifurcation points, there are many other mathematical bifurcations. One example is
the point at which a logistically growing population smoothly goes extinct as conditions become
gradually less favorable. This corresponds to a transcritical bifurcation in which the stable state
touches the trivial zero-population equilibrium and becomes negative. Although this is qualita-
tively important, it is not a catastrophic bifurcation because there is no big jump to an alternative
attractor.
To see intuitively how resilience is affected by change in conditions, we may construct stability
landscapes [also known as potential landscapes (see Supplemental Materials)] for different values
of the conditioning factor (Figure 1c). The slope in these landscapes represents the local rate of
change (Strogatz 1994). For conditions in which two alternative stable states exist, these states
appear as valleys separated by a hilltop. This hilltop is also an equilibrium (the slope of the landscape
is zero). However, this equilibrium is unstable—it is a repellor. Even the slightest change away
from it leads to a self-propagating runaway process that moves the system toward an attractor.
(Note that, although intuitively one could also call this repellor a tipping point, we reserve this
term for bifurcation points such as F1 and F2 in which an attractor ceases to exist upon collision
with the saddle point that marked the boundary of its basin of attraction.)
If a system in equilibrium is perturbed but there is only one basin of attraction, it settles back to
the same state. However, if there are alternative equilibria, a sufficiently severe perturbation may
bring the system into the basin of attraction of another equilibrium. Obviously, the likelihood
of this happening depends not only on the perturbation but also on the size of the attraction
basin. The term resilience can refer to the width of the valley or the basin of attraction around
an equilibrium; this area corresponds to the maximum displacement that can be taken without
causing a shift to an alternative stable state. Two other aspects of resilience are the depth of the
basin, which represents the energy needed to move the system out of the basin, and the slope
around the equilibrium, which represents the return rate upon small perturbations (also known as
engineering resilience).

From Simple Concepts to Complex Reality


Although this way of representing resilience may seem straightforward, it represents only the tip
of the iceberg when applied to the complexity of most real situations. The concept of a stable state
is obviously too simple to represent any real situation. Models as well as experiments and observed
time series suggest that many ecological systems may have intrinsic cycles or even chaotic attractors
instead of stable states. Although systems with such complex attractors may also have alternative

148 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

basins of attraction, such situations cannot be captured by stability landscapes, in which the slope
correctly represents the rate of change. Also, the intuitive interpretation of metaphors such as
stability landscapes is a bit tricky. For instance, they suggest that the ball gains momentum as it
rolls down the hill and therefore may, for instance, overshoot a small valley. This is wrong. There
is no momentum. A more fundamental, technical limitation of the classic stability landscapes is
that they can be constructed only for systems with one state variable [except for very particular
symmetrical cases (Brock & Carpenter 2006)]. Another issue is the suggested constancy of the
landscape. Most work stresses that stability landscapes may change gradually as the result of a
slow change in the conditions. However, the divide between slow change in external control
variables and fast stochastic perturbations that occasionally push the state of the system away
from equilibrium is really somewhat artificial. In practice, conditions are subject not only to slow
trends but also to stochastic fluctuations that also move the system away from equilibrium. This
implies that one might imagine stability landscapes as vibrating and wobbling. Continuing this
stochastic point of view, the system typically does not rest at the bottom of a valley, and if at
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

some point in time its actual state is close to the basin boundary, chances for a shift into the
other valley are diminished even if the basin of attraction is large. Also, the very idea that there
are external control variables (environmental conditions) is usually a simplification. What appears
to be an environmental condition during one timescale may often also be seen as part of the
same dynamical system during another timescale. Thus, as in all concepts and theories, the idea
of tipping points and attraction basins represents merely stylized aspects of reality. Nonetheless,
these highly stylized representations may still help us understand key properties that dominate
the behavior of highly complex systems such as the tropical rainforest.
In the following text we use the view of stability basins and tipping points in one-dimensional
systems (i.e., systems with one-state variable plus the conditioning factor) as a basic framework for
discussing indicators of resilience. From there we discuss extensions to more complex situations
such as spatially extensive systems and networks of species. We repeat this approach for three
classes of situations that differ fundamentally in their discernable resilience indicators: (a) systems
in which the stochastic regime of natural perturbations (i.e., noise) causes only small fluctuations
so that the system remains always close to an equilibrium, (b) systems in which the environmental
noise causes wilder fluctuations so that the state is mostly far from any equilibrium, and (c) systems
that even in the absence of any environmental perturbations display internally generated dynamics
or self-organized spatial patterns. An overview of the studies addressing these different classes of
resilience indicators is given in Table 1. Obviously, systems in practice often fall between those
simplified categories, but the distinction is a good starting point to discuss different fundamental
classes of behavior.

INFERRING RESILIENCE FROM CRITICAL SLOWING DOWN:


CLOSE TO EQUILIBRIUM
Critical transitions to a contrasting state are most striking if the usual fluctuations in a system
are rather small. Much of the recent work on early warning signals for critical transitions focuses
on this situation and makes use of fundamental change—known as critical slowing down—that
happens around equilibrium points as they are about to become unstable.

Time Series
The term critical slowing down (Strogatz 1994) refers to a remarkably generic phenomenon:
systems approaching a particular class of critical points recover more and more slowly from small

www.annualreviews.org • Generic Indicators of Ecological Resilience 149

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Table 1 A summary of the variety of studies that have developed and applied resilience indicators and methods using
models or data
Indicator/Method Models Data
Critical slowing down
Recovery rate/recovery time Bailey 2011, Carpenter et al. 2008, Chisholm & Carpenter et al. 2011, Drake & Griffen
Filotas 2009, Dakos et al. 2011, Gandhi et al. 2010, Veraart et al. 2012, Wissel 1984
1998, Ives 1995, van Nes & Scheffer 2007
Autocorrelation at lag 1 Biggs et al. 2009, Dakos & Bascompte 2014, Carpenter et al. 2011, Dai et al. 2012,
Dakos et al. 2011, O’Regan & Drake 2013 Drake & Griffen 2010, Krkosek & Drake
2014, Lindegren et al. 2012, Veraart et al.
2012, Wang et al. 2012, Wouters et al. 2015
Variance/coefficient of variation Biggs et al. 2009, Brock & Carpenter 2010, Beaugrand et al. 2008, Bestelmeyer et al.
Carpenter & Brock 2006, Carpenter et al. 2008, 2013, Carpenter et al. 2011, Dai et al. 2012,
Contamin & Ellison 2009, Dakos & Bascompte Drake & Griffen 2010, Hsieh et al. 2006,
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org

2014, Dakos et al. 2011, Fung et al. 2013, Lindegren et al. 2012, Veraart et al. 2012,
Access provided by University of Otago on 10/14/15. For personal use only.

Hastings & Wysham 2010, O’Regan & Drake Wang et al. 2012, Wouters et al. 2015
2013, Takimoto 2009
Conditional heteroskedasticity Seekell et al. 2011 Seekell et al. 2012
Spectral reddening Carpenter et al. 2008, Contamin & Ellison 2009, Carpenter et al. 2011
O’Regan & Drake 2013
Eigenvalue of variance- Carpenter et al. 2008, Dakos & Bascompte 2014, —
covariance matrix Suweis & D’Odorico 2014
Discrete Fourier transform Carpenter & Brock 2010 Carpenter et al. 2011, Cline et al. 2014
Spatial correlation Dakos et al. 2011, Dakos et al. 2010, Donangelo Dai et al. 2013, Drake & Griffen 2010,
et al. 2010 Lindegren et al. 2012
Spatial variance Dakos et al. 2011, Donangelo et al. 2010, Cline et al. 2014, Dai et al. 2013, Hewitt &
Guttal & Jayaprakash 2009, Oborny et al. 2005 Thrush 2010, Lindegren et al. 2012
Likelihood ratio Boettiger & Hastings 2012 Boettiger & Hastings 2012
Drift-diffusion-jump models Brock & Carpenter 2012, Carpenter et al. 2011 —
Eigenvalue from general models Lade & Gross 2012 —
Time-varying autoregressive Ives & Dakos 2012 —
models
Interaction network–based Tirabassi et al. 2014 —
indicators
Flickering
Skewness Biggs et al. 2009, Carpenter et al. 2008, Dakos Carpenter et al. 2011, Dai et al. 2012,
et al. 2011, Dakos et al. 2013, Fung et al. 2013, Drake & Griffen 2010, Wang et al. 2012
Guttal & Jayaprakash 2008, Hastings & Wysham
2010, Takimoto 2009
Kurtosis Biggs et al. 2009 —
Spatial skewness Dakos et al. 2011, Guttal & Jayaprakash 2009 —
Brock Dechert Scheinkman test Dakos et al. 2012a Batt et al. 2013a, Elser et al. 2014
Quickest detection method Carpenter et al. 2014 Batt et al. 2013b
Threshold autoregressive Ives & Dakos 2012 —
models
Potential analysis Livina et al. 2010 Hirota et al. 2011, Scheffer et al. 2012b
Probability distributions Livina et al. 2010 Hirota et al. 2011, Scheffer et al. 2012b
(Continued )

150 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Table 1 (Continued )
Indicator/Method Models Data
Mean exit time Guttal & Jayaprakash 2008 —
Repair time/recovery length Dai et al. 2013 Dai et al. 2013
Speed of traveling waves Kuehn 2013, van de Leemput et al. 2015 —
Autonomous fluctuations and patterns
Resonance Benincá et al. 2011 —
Variance Batt et al. 2013a, Kéfi et al. 2013 —
Autocorrelation Kéfi et al. 2013 —
Turing patterns Rietkerk et al. 2004, von Hardenberg et al. 2001 —
Patch size distributions Dakos et al. 2011, Kéfi et al. 2007 Kéfi et al. 2007

The list is not exhaustive but reveals gaps and opportunities in the testing of specific methods. More background on the specific methods and indicators, as
well as software tools, can be found at http://www.early-warning-signals.org/.
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

perturbations (Figure 2). More precisely formulated, the return rate to equilibrium goes to zero
in a range of bifurcation points including the saddle-node bifurcation point known as the classic
tipping point (Wissel 1984) (Figure 1). In mathematical models, this return rate is characterized
by the dominant eigenvalue of the Jacobian matrix (see Supplemental Materials). Importantly, in
many situations the slowing down sets in far from the tipping point at which the basin of attraction
ceases to exist (van Nes & Scheffer 2007), so that it may really be used to indicate change in the
size of the basin of attraction (Scheffer et al. 2009). The slowness of a system can be detected
from the rate of recovery from (experimental) perturbations (van Nes & Scheffer 2007, Veraart
et al. 2012); however, it can also be inferred indirectly from the small fluctuations in the state
of a system induced by the natural stochastic regime of environmental perturbations (Ives 1995)
(Figure 2). The character of the fluctuations of the system’s state changes as a tipping point comes
close because the slowness of the system’s intrinsic rates causes it to differ less from its previous
state at any moment. Thus, memory, as it were, increases—as reflected in higher correlation to
the state 1 time step before (the lag-autocorrelation, Figure 2h versus Figure 2g) (Dakos et al.
2008, Ives 1995). The behavior of the system becomes more like a random walk, and in addition to
the rise in lag-autocorrelation, this behavior leads to an increase in variance (Carpenter & Brock
2006). In the limit case, at the bifurcation point the temporal autocorrelation tends to unity and
the variance to infinity (Carpenter & Brock 2006, Dakos et al. 2008, Ives 1995). As we discuss in
more detail below, the stability basin also tends to become less symmetric as an unstable point (the
hilltop in Figures 1 and 2) approaches the attractor (the bottom of the valley). This leads to an
asymmetric increase in variance and causes a skewed distribution of states (Guttal & Jayaprakash
2008).

Spatial Information
As a next step, imagine a spatial extension in which the dynamical system with alternative attractors
(e.g., Figures 1 and 2) is repeated in many units, which are then connected through diffusion of
the variable. So we still look at the same variable but in many connected places. For instance,
one could think of a spatial network of coral reefs or of a regular lattice representing a spatially
extensive population. As resilience of the state in a grid cell (a single unit) becomes small, critical
slowing down of the internal dynamics increases the relative effect of the diffusive connection to
linked units (just as the relative effect of past states, the memory, becomes stronger). This can be
shown mathematically (Dakos et al. 2010), but intuitively one may understand it from the weaker

www.annualreviews.org • Generic Indicators of Ecological Resilience 151

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

a High resilience b Low resilience

c d e f
System state

System state
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

Time Time Time Time

g h
State (t + 1)

State (t + 1)
State (t) State (t)

Figure 2
Critical slowing down as an indicator of low resilience. Recovery rates upon perturbations (c and e) are
slower when the basin of attraction is shallow (panels b, e, f, h) than when the basin of attraction is deeper
(panels a, c, d, g). The effect of this slowness is reflected in natural (externally driven) fluctuations in the state
of the system (d and f ) and can be detected as increased temporal autocorrelation and variance ( g and h).
Abbreviation: t, time. Figure adapted from Scheffer et al. (2012a) with permission.

tendency of the local system to move to its intrinsic local equilibrium. A corollary is that a small
local perturbation causes a larger part of space to be affected, and recovery from this temporarily
spreading impact will be slow (Dai et al. 2013). The increased relative effect of neighboring units
also leads to an increase in spatial autocorrelation (Dakos et al. 2010) as well as variance and
skewness (Guttal & Jayaprakash 2009). Thus, although there are many subtleties (Dakos et al.
2011), critical slowing down tends to be reflected in space much in the same way as it is in time.

Multivariate Information
Another obvious extension of the simplest one-variable-one-place case is to consider a network of
interacting variables, such as a network of many interacting species. As conditions change, such
networks may approach tipping points for critical transitions (Lever et al. 2014, van Nes & Scheffer
2004), and the search for indicators of loss of resilience before such tipping points are reached
is still ongoing. The findings so far point at the same indicators of slowing down that we find
in spatial systems (Dakos & Bascompte 2014, Suweis & D’Odorico 2014). However, an exciting
extra dimension in this field is that slowing down happens in quite specific ways that may indicate
which species in the network are most likely to be involved in an upcoming critical transition.

152 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

More specifically, the pattern of changing temporal autocorrelations (Dakos & Bascompte 2014),
and also cross-correlations between species, may suggest which species could become extinct and
which ones could become the new dominant species (van de Leemput et al. 2014).

INFERRING RESILIENCE FROM FLICKERING:


FAR FROM EQUILIBRIUM
As we have just seen, critical slowing down is generic in the sense that it is not dependent on the
exact underlying model. However, as the concept of critical slowing down is based on linearization
of the model close to the equilibrium state, we can infer such slowing down only from dynamics
close to equilibrium. In practice, ecological systems are mostly in transient states due to high
environmental stochasticity and/or intrinsic processes (DeAngelis & Waterhouse 1987, Hastings
2004). In this section, we focus on indicators of resilience in systems with alternative states in
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org

which environmental stochasticity causes strong fluctuations. The topic of intrinsically driven
Access provided by University of Otago on 10/14/15. For personal use only.

oscillations and chaos is discussed in the next section.

Time Series
As systems fluctuate more wildly, critical slowing down is less likely to be detectable, because
the system is far from the equilibrium around which the slowing down happens. However, such
systems sample a larger part of the state space, implying that one can infer much more about
the dynamical properties beyond the direct vicinity of the equilibrium. For instance, when the
perturbation regime is strong enough to bring the system close to the boundary between two
basins of attraction, the probability distribution becomes skewed toward the alternative stable
state (Guttal & Jayaprakash 2008). This is because the dynamics are also slow around the unstable
equilibrium at the basin boundary, causing the system to dwell there longer than in other transient
states.
If the stochasticity is high enough to push the system over the boundary of the basin of at-
traction, the system may repeatedly visit different basins of attraction, a process termed flickering
(Scheffer et al. 2009). In such cases not only the resilience of the current state is important but
also the resilience relative to that of the other attractor. This relative resilience together with the
strength of stochasticity determines which fraction of the time is spent in either of the basins of
attraction and how often regime shifts will take place (Figure 3). There are two ways of charac-
terizing this relative resilience, either by analyzing the probability distribution of states (Livina
et al. 2010) or by estimating the mean exit time (Grasman & Van Herwaarden 2010, Hanson &
Tuckwell 1978).
Potential analysis (Livina et al. 2010) is a way to reconstruct the stability landscapes based on
data. If one has a long enough time series for given conditions, and if both basins of attraction are
visited, the probability distribution is bimodal with the position of the modes corresponding to
the equilibria as the system spends relatively more time in the vicinity of the strongest attractors.
Thus, the probability distribution of states carries information about the actual stability landscape.
Indeed, one may compute (an approximation of ) the stability landscape from the probability
distribution of states through a simple log transformation (Livina et al. 2010) (see Supplemental
Materials). This can be shown analytically for simple models applying the Fokker–Planck equation
(Carpenter & Brock 2006, Gardiner 2004, Grasman & Van Herwaarden 2010). From the resulting
stability landscapes one can objectively determine how many alternative attractors exist. More
importantly, such landscapes also allow us to infer the relative resilience of each of the attractors

www.annualreviews.org • Generic Indicators of Ecological Resilience 153

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Control parameter (c)


1 1.5 2 2.5 3
15
a

10

System state

5
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org

0
Access provided by University of Otago on 10/14/15. For personal use only.

0 2,500 5,000 7,500 10,000


Time (a.u.)

0.3
Probability density

b c
0.2

0.1

0
3
d e
Potential

0
0 5 10 15 0 5 10 15
System state System state

Figure 3
Flickering between alternative states in highly stochastic systems. As conditions change over time, the system
may jump back and forth between alternative basins of attraction rather than going through a single
catastrophic transition (panel a). The probability distribution of states, shown in panels b and c, can be used
to infer the shape of the stability landscapes, shown in panels d and e. Figure adapted from Scheffer et al.
(2012a) with permission.

and determine how much time the system should be expected to spend in either of the attractors,
given the external conditions.
The second way to characterize the relative resilience of a flickering system is by measuring
the mean exit time (also termed mean first passage time or recurrence time). This is defined as the
average time it takes before the system leaves the basin of attraction (Gardiner 2004, Grasman &
Van Herwaarden 2010). The mean exit time was already proposed by Hanson & Tuckwell (1978)
as a measure of resilience, but even though it has been applied in climate science (Berglund &
Gentz 2002, Kleinen et al. 2003), its use in ecology thus far has been limited mainly to population
viability studies (Cooper 1984, Leigh 1981, Sæther et al. 1998). In simple models the mean exit time

154 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

from each basin of attraction can be determined by applying the backward time Fokker–Planck
equation (Gardiner 2004). For instance, mean exit time to overexploitation was shown to decrease
in a vegetation model with alternative states (Guttal & Jayaprakash 2008). In real data a reliable
determination of the mean exit time is difficult as the exit from a basin of attraction is a rare event.
The probabilistic approach to estimating stability landscapes, resilience of alternative states,
and mean exit times becomes especially powerful if data are available on massive amounts of
systems. We elaborate further on this possibility in the section entitled A Practical Angle.

Spatial Boundaries and Recovery from Local Perturbations


In the previous section we considered a single system for simplicity. However, if we get back to the
issue of spatial dynamics, it becomes interesting to ask how contagious local flips to a different state
are. When should we expect a domino effect to cause cascading change through a landscape? Are
there indicators for the resilience to such shifts? Before that discussion, we must briefly elaborate on
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

the nature of spatial tipping points.


As discussed in the section entitled Inferring Resilience from Critical Slowing Down, one
model formalization is to make a spatial grid in which diffusion of biomass between neighboring
cells occurs. However, if one formulates models in a spatially continuous way, it becomes clear that
a spatial boundary between two alternative stable states will usually be unstable (van de Leemput
et al. 2015). Specifically, the boundary always moves toward the state with the lowest resilience
(the lowest potential—i.e., the smallest basin of attraction in the stability landscape viewed at the
local scale). Such a moving boundary is known as a traveling wave (Murray 2002). If the ecosystem
is perturbed locally—setting a patch to the alternative stable state—the resulting spot tends to
either grow (if the new state is more resilient) or decay (if the new state is less resilient) (Keitt et al.
2001). The size of the perturbed patch remains constant only if both states are equally resilient
(unless there is some heterogeneity in the spatial exchange rates or in the conditions) (van de
Leemput et al. 2015).
The critical point at which both of the alternate states are equally resilient is termed the Maxwell
point (Bel et al. 2012, Pomeau 1986). Clearly the Maxwell point is crucial in terms of the behavior
of spatially extensive systems, as it marks the conditions in which an entire landscape may be swept
from one state to another state by a traveling wave. It is therefore important that the distance to
this special point may also be sensed through recovery time upon local perturbations. The speed
of recovery of a local perturbation goes to zero as the system gets closer to the Maxwell point
(Keitt et al. 2001); thus, one should expect that local experimental perturbations can be used to
estimate resilience (van de Leemput et al. 2015). Also, the transient spread of a perturbation in
space carries information about such resilience (Dai et al. 2013).

INFERRING RESILIENCE IN SYSTEMS WITH AUTONOMOUS


FLUCTUATIONS AND PATTERNS: NO EQUILIBRIUM
So far, we have considered indicators of resilience by envisioning situations in which perturbations
affect a dynamical system that in a constant environment would have stable equilibria. However,
there are also systems that even under constant environmental conditions keep fluctuating in a
periodic (Fussmann et al. 2000) or chaotic way (Benincà et al. 2008, Bjornstad & Grenfell 2001).
Similarly, some systems spontaneously form spatial patterns, even if the initial environment is
homogeneous (Rietkerk et al. 2004, von Hardenberg et al. 2001). The onset of such dynamical or
spatially patterned behavior in models happens when critical parameter values are passed and the
proximity of such bifurcation points is announced by critical slowing down. At the same time, there

www.annualreviews.org • Generic Indicators of Ecological Resilience 155

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

may be critical parameter values for shifts between alternative complex attractors (e.g., alternative
strange attractors or spatial patterns). Such tipping points correspond to nonlocal bifurcations,
also referred to as basin-boundary collisions (Vandermeer & Yodzis 1999). The distance to such
bifurcation points is typically less easily inferred from the behavior of a system. In a sense, the
distinction between externally induced and intrinsically generated dynamics and spatial patterns
is artificial, as environmental variation blends with intrinsic dynamics to produce a compound,
complex pattern (Bjornstad & Grenfell 2001). Nonetheless, it is helpful to separate these sources
of system fluctuations in space and time.

Cycles and Chaos


As an example of how we can infer the distance to a critical point at which a system becomes
unstable, consider the Hopf bifurcation (Strogatz 1994). Here, a stable equilibrium gives way to
an unstable one, which is surrounded by a cyclic attractor—as known, for instance, from predator–
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

prey dynamics (Rosenzweig 1971). Dynamics around such local bifurcations (involving a stable
point) are relatively easy to understand mathematically (see Supplemental Materials). In the
vicinity of the bifurcation, the system returns to the stable state upon a perturbation through
damped oscillations (Rinaldi & Muratori 1992). Thus, the stable state is a damped oscillator, much
like a bell or wine glass. As the stable system comes closer to the Hopf bifurcation, we see critical
slowing down in the sense that the dampening of the cycles is slower. This form of critical slowing
down was shown experimentally in an isolated squid axon in the 1970s (Matsumoto & Kunisawa
1978). Now if, instead of applying a single perturbation, we expose such damped oscillators to
a stochastic regime of perturbations (i.e., environmental noise), they tend to resonate (Benincá
et al. 2011). That is, the system starts fluctuating in a quite regular way with a dominant frequency
equal to that of the oscillations in the cyclic attractor that appears at the Hopf bifurcation. Such
resonance is similar to what happens when one blows across a bottleneck, causing a soft noise that
resonates in the bottle and produces its characteristic tone, in which one particular frequency is
overwhelmingly strong. Thus, just as in a classic tipping point (Figures 1 and 2), critical slowing
down before the Hopf bifurcation is reflected in elevated temporal autocorrelation and variance
(Batt et al. 2013a, Kéfi et al. 2013). However, resonance takes the specific form of producing
fluctuations with a specific frequency. Other bifurcations in oscillating systems, such as phase
locking between different oscillators, are also preceded by forms of critical slowing down (Leung
2000) that could serve as indicators of resilience.
Probably most difficult when it comes to empirical indicators of resilience are situations in
which intrinsic cycles or other dynamics bring the system to the border of the basin of attraction,
at which point they tip toward an alternative attractor (Hastings & Wysham 2010, Suzuki &
Yoshida 2015, Vandermeer 2011). This is not to say that indicators are not possible. For instance,
when approaching such a basin boundary collision (Vandermeer & Yodzis 1999), cycles become
stretched in time as they slow down at each passage close to the exit point (Rinaldi & Scheffer
2000, Suzuki & Yoshida 2015). However, resilience in the sense of the likelihood of staying in the
same attractor obviously becomes difficult to infer from data, especially if the intrinsic dynamics
blend with the effects of environmental stochasticity.

Self-Organized Spatial Patterns


Although intrinsic oscillations emerge when the stable state becomes unstable, there are also
dynamical systems in which a stable state becomes unstable if spatial diffusion is allowed. Such
instability gives rise to regular spatial patterns. Alan Turing showed this fundamental possibility

156 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

0.4
Gaps
Spatial average vegetation biomass

0.3 Labyrinth

Turing
instability

0.2 Dots
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

0.1

0
Hyper-
arid Arid Semiarid Dry–subhumid
0 0.6
Precipitation
Figure 4
Self-organized patterns of green biomass of perennial grasses in the Negev desert (photographs, top)
compared with self-organized Turing patterns generated by a simple model of spatial dynamics of dryland
vegetation (inset in graph). The curves reveal how the total spatially averaged biomass changes rather
gradually, with small jumps between alternative spatial patterns, until a collapse to a barren state occurs at dry
conditions. This state is an alternative stable state, as revealed by the hysteresis in recovery as precipitation
increases. The patterns reflect how far the system is from catastrophic collapse. The dashed line represents
an unstable homogeneous state. Figure adapted from von Hardenberg et al. (2001) with permission.

by addressing how patterns could arise spontaneously during the formation of an embryo from
an initially symmetrical clump of cells (Turing 1952). A range of regular patterns in ecosystems,
including desert vegetation (von Hardenberg et al. 2001), sea grass fields (van der Heide et al.
2010), and mussel beds (van de Koppel et al. 2008), are thought to be caused by such Turing
instability (Rietkerk & van de Koppel 2008). Usually, if the environment is changed from favorable
to harsh (Figure 4), patterns change in a characteristic sequence: (a) First the homogeneous
distribution (Turing bifurcation) is broken up and regular patterns emerge; (b) then these patterns
shift in character from holes to labyrinths or stripes, eventually leaving only dots; (c) and complete
extinction follows, leaving a bare space. Loss of resilience of the stable, unpatterned state in such
systems is reflected by slowing down of the recovery rate upon perturbations, and the same is true
for loss of resilience of the dotted state before collapse (Bailey 2011, Dakos et al. 2011). However,
the characteristic patterns themselves may also be seen as indicating the distance to the critical
points at which the patterns are lost and the system becomes homogeneously covered or entirely
empty (Rietkerk et al. 2004).
Many ecosystems show more or less irregular clumpy patterns rather than regular Turing-
like patterns. These clumps can also arise from self-organization, and for particular cases, the

www.annualreviews.org • Generic Indicators of Ecological Resilience 157

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

size distribution of clumps may be informative as an indicator of resilience (Kéfi et al. 2007).
Normally the patch sizes approximately follow a power law, in the sense that plotting the patch
size against the frequency at which that size occurs follows a straight line on a double-log scale.
This reflects that there are systematically fewer large patches than small patches. Under adverse
conditions in which vegetation may be close to complete collapse, the power law distribution
becomes truncated as the large patches are lost and only smaller ones are left. In such systems, the
patch size distribution can be seen as a measure of resilience.

A PRACTICAL ANGLE: POSSIBILITIES AND LIMITATIONS


In the previous sections we have looked at the fundamentally different kinds of resilience indicators
one may theoretically expect. Now we turn to a more practical perspective, asking how we may
in practice estimate resilience given different kinds of data (Figure 5). First, we look at the
possibilities of using the previously discussed indicators to qualitatively determine whether some
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

instances of a system (e.g., which reef, population, or lake) are likely to be more resilient than others.
Subsequently, we discuss the potential for and limitations of monitoring change of resilience in a
particular system (e.g., a fish stock or reef ) over time. Lastly, we review the possibilities of using
massive data sets to obtain probabilistic estimates of resilience in relation to potential drivers.

Ranking Resilience
One of the ways to use resilience indicators in practice is to rank systems of a similar kind.
For instance, we may want to rank coral reefs along the Great Barrier Reef according to their
resilience (DeVantier et al. 1998) or compare the resilience of populations of a species within
a given geographic area. To do this, we need to estimate resilience indicators for each instance
of such systems (Figure 5a). Many approaches for detecting critical slowing down from either
time series or spatial data have been developed. Generally, indicators can be distinguished in
metric-based and model-based approaches (Dakos et al. 2012a), and most can be applied to both
time series and spatial data (Dakos et al. 2012a, Kéfi et al. 2014). The distinction is quite simple:
Metric-based indicators quantify changes in the statistical properties of the observed data, whereas
model-based indicators first fit the data to a generic model and then use the model parameters or
model output to quantify resilience.
Metric-based indicators are standard summary statistics—most commonly variance and au-
tocorrelation (Carpenter & Brock 2006, Dakos et al. 2008, Guttal & Jayaprakash 2008, Seekell
et al. 2011), indices of correlation structure (Held & Kleinen 2004, Livina & Lenton 2007), or
quantifications of spatial patterns (Kéfi et al. 2007, Rietkerk et al. 2004, von Hardenberg et al.

−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Figure 5
Examples from three classes of empirical studies of resilience indicators. (a) Class I: (i ) The power law distribution of vegetation patch
sizes (ii ) becomes truncated due to the loss of large patches as the arid ecosystem becomes less resilient in response to higher grazing
pressure (Kéfi et al. 2007). (b) Class II: Loss of resilience as microbial populations of (i, ii ) cyanobacteria and (iii, iv, v) yeast approach a
tipping point. In the cyanobacterial study (Veraart et al. 2012), increasing radiation gradually drives the population to collapse, and
(ii ) recovery rates upon small perturbations reflect the loss of resilience. In the yeast experiments (Dai et al. 2012), increasing dilution
rates drive the loss of resilience as reflected in slowing down and indicated by rising temporal autocorrelation and standard deviation of
fluctuations. (c) Class III: (i ) The probability distribution of satellite-inferred global tropical tree cover varies with the mean annual
precipitation rates. (ii,) At 1,200 mm the probability of finding tree cover >60% is low. (ii,) At 1,800 mm a forested state with tree
cover around 80% becomes more likely. (ii,) At 2,400 mm the fraction of points in the forested state is much higher, reflecting a high
probabilistic resilience (Hirota et al. 2011).

158 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

a Ranking resilience
i Low grazing pressure ii High grazing pressure

10 2 10 2

log(number of
the patches) 10 1 10 1
Power law Truncated
distribution power law
10 0 10 0
10 1 10 2 10 3 10 1 10 2 10 3
log(size of the patches) log(size of the patches)

b Monitoring changes in resilience


Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

p1 p2 p3 p4
i 400 p5 iii 10 6
p6 Stable fixed point
Collapse

Population density
Extinction (m–1)

380
10 5
360 (cells/ml)
10 4
340
Unstable fixed point
10 3
320
Extinction
0 5 10 15 20 25 30 0
Time (days) 0 500 1,000 1,500 2,000
Dilution factor
ii 1.2
p2
iv Autocorrelation time
v Standard deviation
p1
Recovery rate

p3 10 0.2
0.8 p4 p5 8
6
0.4 0.1
4
p6
2
0 0 0
500 700 900 1,100 0 500 1,000 1,500 0 500 1,000 1,500
Increasing light intensity Dilution factor Dilution factor

c Inferring probabilistic resilience from massive data


i 100 Intercontinental

80
Tree cover (%)

60 1 2 3 ii 40
1 1,200 mm 2 1,800 mm 3 2,400 mm
Frequency (%)

30
N = 6,586 N = 5,621 N = 2,832
40
20
20
10

0
0
0 1,000 2,000 3,000 0 10 20 40 60 80 90 0 10 20 40 60 80 90 0 10 20 40 60 80 90 100
Mean annual precipitation (mm/year) Tree cover (%) Tree cover (%) Tree cover (%)

www.annualreviews.org • Generic Indicators of Ecological Resilience 159

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

2001). Model-based indicators can be based on likelihood ratio tests between generic models with
and without tipping points (Boettiger & Hastings 2012), autoregressive models with time-varying
parameters (Ives & Dakos 2012), nonparametric models (Carpenter & Brock 2011), reconstruc-
tions of stability landscapes by potential analysis (Livina et al. 2010), or estimations of eigenvalues
using generalized models that are fit to data (Lade & Gross 2012). Extensive practical guides to
analyzing critical slowing down resilience indicators in time series (Dakos et al. 2012a) or spatial
data (Kéfi et al. 2014) are published elsewhere, and there are freely available toolboxes for applying
the indicators in practice (Boettiger & Hastings 2012, Dakos et al. 2012a, Ives & Dakos 2012).
As there is no silver bullet approach, it is important to explore more than one resilience in-
dicator, see how consistent results are, and possibly combine indicators to characterize resilience
(Dakos et al. 2012a). Obviously, rankings of resilience should be based on the same indicators esti-
mated by the same methods. Also, keep in mind that intrinsic rates of change differ between systems
(e.g., bacteria reproduce faster than trees, and air temperatures vary at a faster rate than ocean
temperatures) and that the amplitude and frequency characteristics of environmental variability
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

are reflected in the resilience indicators. Therefore, ranking resilience on the basis of indicators
of slowing down can be done only if the different instances of the system as well as the variability
of their environments are relatively similar. In any case, resilience rankings are best interpreted in
a qualitative way. For example, Kéfi et al. (2007) have analyzed the patch size distribution of veg-
etation patterns to compare the level of degradation in three Mediterranean semiarid ecosystems
with comparable environmental conditions and similar levels of grazing (Figure 5a). Similarly,
Dakos & Bascompte (2014) showed how changes in variance of population biomass may be used
to rank species resilience in simulated plant–pollinator communities in which species were similar
in terms of intrinsic growth rates.

Monitoring Change in Resilience


One of the most exciting, but also the most challenging, applications of the theory is to monitor
changes in the resilience of a single ecosystem. In the ideal scenario, the monitored data correspond
to detailed observations at discrete levels of environmental conditions along a known gradient
(Figure 5b). For instance, Dai et al. (2012) recorded the dynamics of laboratory yeast populations
at discrete, increasing levels of mortality up to a certain threshold, at which the yeast collapsed
to extinction. In most cases, however, we have a continuous record in which we typically assume
that conditions change gradually toward (or away from) a critical point. We may then estimate
resilience for different parts of the record using the same indicators as mentioned above.
For a time series, the most common way to monitor changes in resilience is to measure the
indicators within a sliding window of a given size along the time series. For spatial data, we estimate
indicators from snapshots that correspond to a gradient of conditions. For each sliding window
or spatial snapshot, we use the resilience indicators to estimate whether their trend signals loss or
gain of resilience. Similarly, in the case of model-based indicators, we fit a model to the complete
time series to estimate trends in resilience based either on model parameters or output.
Although monitoring changes in resilience is one of the most discussed applications of the
theory, it is also quite tricky. Monitoring changes in resilience requires long-term records sampled
at temporal and spatial scales relevant to the dynamics of the studied system. It is important to
first remove trends in the time series that may potentially bias the estimation of the indicators,
such as seasonal variation or longer-term trends. Failing to do that can lead to false signals. For
instance, if after some period of fluctuating around the same state the system enters a downward
or upward trend, this trend causes an increase in temporal autocorrelation and variance, even if it
is not related to slowing down of the system. Also, it is important to check for the robustness of the

160 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

results to partly arbitrary choices, such as data aggregation methods, detrending parameters, size
of the sliding window, and model-fitting conditions. Lastly, one should evaluate the significance
of the results. What is the likelihood that an observed trend in indicators occurred by chance (i.e.,
false positive)? Similarly, could a trend in indicators be absent, despite an important change in
resilience (i.e., false negative)? Some methods are associated with significance testing, but in most
cases null models based on the original data need to be developed to measure the performance of the
resilience indicators by building, for instance, receiver-operator characteristic curves (Boettiger
& Hastings 2012). In most cases, it is preferable to test a variety of methods and null models.
The main appeal of monitoring changes through resilience indicators lies in their potential
use to detect upcoming tipping points. However, even if analysis of time series is done carefully,
several fundamental limitations imply that useful early warning signals of upcoming tipping points
can be difficult to obtain. First of all, strong environmental noise may erase any signature of critical
slowing down (Perretti & Munch 2012) and can also force a system to another state far before
it reaches the tipping point. Second, change in the magnitude, color, and type of environmental
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

stochasticity can affect the indicators even if there is no change in resilience (Dakos et al. 2012b).
For instance, increasing magnitude or temporal autocorrelation of environmental shocks lead
to increasing ecosystem variance and temporal autocorrelation. Third, systems may slow down
or speed up irrespective of tipping points—for instance, due to temperature-driven changes in
growth rates. Fourth, if environmental conditions are changing fast relative to ecosystem response
rates, detecting loss of resilience before the tipping point becomes problematic (Biggs et al. 2009,
Contamin & Ellison 2009). Finally, it is also important to measure the right variable. In multivariate
systems, even if there is slowing down, some variables might yield weak signals of slowing down
(Dakos & Bascompte 2014), and other variables may provide no signal at all (Boerlijst et al. 2013).
Multivariate ecosystems may have two or more regime shifts involving different sets of interacting
variables. Interactions of multiple regime shifts can increase or decrease variance and thereby
distort signals of changing resilience (Brock & Carpenter 2010).
An important limitation of monitoring changes in resilience is the inability to interpret indi-
cators of critical slowing down in an absolute way. It is difficult to say, for instance, which level
of autocorrelation or variance corresponds to which resilience in terms of probabilities of shifting
to an alternative state. One reason is that both indicators depend on the character of the envi-
ronmental variability. Autocorrelation is theoretically expected to reach unity (1.0) at the tipping
point, but in reality all ecosystems shift before the actual tipping point due to external stochastic-
ity. Thus, the interpretation of change in indicators of resilience should be qualitative, suggesting
merely whether the ecosystem is gaining or losing resilience.

Inferring Probabilistic Resilience from Massive Data Sets


In some situations it is possible to obtain data in an automated way on numerous instances of a
system. For instance, the abundance of a thousand bacterial types can be determined in a sample
of material using a DNA probing system on a chip. Using this method, the abundance of these
different bacteria has been assessed in the intestinal ecosystem of more than a thousand healthy
people (Lahti et al. 2014). Similarly, satellite data can be used efficiently to estimate the turbidity
of large numbers of lakes (Kosten et al. 2012) or assess tree cover of every square kilometer across
the globe (Hansen et al. 2003).
In the simplest case, we may just have many observed states at one moment in time. As we have
seen in the section on flickering, the probability distribution of states allows the reconstruction of
the underlying stability landscape. With this idea in mind, satellite data of tree cover have been
analyzed to infer how stability basins for tropical forest and savanna change with mean annual

www.annualreviews.org • Generic Indicators of Ecological Resilience 161

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

precipitation (Hirota et al. 2011) (Figure 5c). Although such stability landscapes are nice visual
representations of the basins of attraction, they do not immediately provide a numerical estimate
of resilience for the different precipitation levels. One way to quantify resilience on the basis of
such data is to classify states and relate their probability to potential drivers. For instance, one can
define a threshold of tree cover that separates forest from savanna (the minimum of the probability
density) and subsequently use logistic regression to find the relationship between the probability
of forest (a measure of resilience) and environmental variables that are plausible drivers, such
as precipitation (Hirota et al. 2011). This then offers the possibility to map current and future
resilience of existing forest based on environmental variables such as rainfall.
Note that such resilience based on probability distributions is a relative indication. We can
infer the ratio of resilience of alternative states from the ratio of observed points in the alternative
states. It follows from simple conservation of mass that the mean exit times from the alternative
states (another probabilistic indicator of resilience) should follow the same ratio as the number of
points in the alternative states. However, we cannot estimate the absolute values of such mean exit
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

times from a single snapshot of each observed system or point in space. Thus, we cannot make a
statement such as this: Given a mean annual rainfall of 1,200 mm, on average a forest will shift
to savanna in 126 years. In order to derive such probabilistic estimates of the mean exit time (or
equivalently the probabilities of switching to the other state in a given time span), one needs to
examine the temporal dynamics.
A limitation of estimating probabilistic resilience as a function of environmental drivers is
that the results can be biased in different ways (van Nes et al. 2014). First of all, heterogeneity of
unobserved driving variables can lead to an overestimation of the range of hysteresis. For example,
in the case of the forest–savanna bistability, some places may have forest even under unexpectedly
dry conditions due to exceptionally good local soil conditions. Second, directionality in historical
change in conditions can be influential. For example, if the past climate was wetter, one sees more
forest than should be expected under the current conditions and overestimates forest resilience
at current precipitation levels (van Nes et al. 2014). Despite such caveats, availability of massive
data sets can obviously offer exceptional opportunities to estimate how resilience of particular
ecosystem states relates to environmental factors.

PROSPECTS
Taken together, the recent work on different classes of generic indicators of ecological resilience
has opened up a plethora of new possibilities. Although the concept of ecological resilience has
been influential since Holling’s (1973) review, ideas on how to measure this important property of
ecological systems have long been missing. Perhaps the most remarkable feature of the indicators
discussed above is that they are generic. Unlike indicators such as the occurrence of certain
species or the browning of leaves, the indicators we discuss are related to fundamental properties
of dynamical systems. This makes it possible to measure a phenomenon such as critical slowing
down in systems as different as the axon of a squid (Matsumoto & Kunisawa 1978), the climate
system (Dakos et al. 2008), a yeast population (Dai et al. 2012), or a lake (Carpenter et al. 2011).
Such generality is rather exceptional in ecology. Obviously, this is not to say that the resilience
indicators are universally applicable. As we have seen, different indicators may be useful in different
classes of systems, and there are many limitations. Even though this may lead one to “find the glass
half empty,” we take the optimistic view of finding it “half full with many unexplored possibilities
of filling it further.” A glance at studies that have used these indicators to date (Table 1) suggests
that only a few of the theoretically predicted indicators have been explored in field situations.
Large-scale field studies in which we also have good models of underlying mechanisms will be

162 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

especially important as testing grounds for our understanding of possibilities and limitations for
measuring resilience. Clearly there are many challenges in the face of the limited data on which
we often have to rely in ecology. However, the availability of massive data sets is increasing with
technical advances in remote sensing, DNA sequencing, and high-frequency automatic measuring
schedules. In addition, online data collection and long-term monitoring programs are growing
sources of information.
Such information can be used in novel ways to infer stability properties of ecological systems.
This approach adds complementary tools to the toolbox we have to infer the risk of epidemic
outbreak, population extinction, or critical transitions in ecosystems ranging from lakes to reefs
and forests. There are no silver bullet approaches in the science of complex systems. In ecology,
much as in social sciences, medicine, or other branches of science that study truly complex systems,
the best understanding comes from a combination of approaches, ranging from mathematical
modeling and controlled experiments to analysis of natural patterns, long-term observations, and
large-scale field manipulations. The approaches we discuss are not fundamentally new. They are
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

all based on long-known universal properties of dynamical systems. However, the notion that such
universal laws can be used to infer the probability of critical transitions is somehow revolutionary.
The idea of generic indicators of resilience creates entirely fresh ways to look at our data and
to design experiments. Such fundamental advances are badly needed if we are to understand the
effects of rapid global change on the stability properties of the web of complex natural systems on
which humanity ultimately depends.

DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.

LITERATURE CITED
Bailey RM. 2011. Spatial and temporal signatures of fragility and threshold proximity in modelled semi-arid
vegetation. Proc. R. Soc. B 278:1064–71
Batt RD, Brock WA, Carpenter SR, Cole JJ, Pace ML, Seekell DA. 2013a. Asymmetric response of early
warning indicators of phytoplankton transition to and from cycles. Theor. Ecol. 6:285–93
Batt RD, Carpenter SR, Cole JJ, Pace ML, Johnson RA. 2013b. Changes in ecosystem resilience detected in
automated measures of ecosystem metabolism during a whole-lake manipulation. PNAS 110:17398–403
Beaugrand G, Edwards M, Brander K, Luczak C, Ibanez F. 2008. Causes and projections of abrupt climate-
driven ecosystem shifts in the North Atlantic. Ecol. Lett. 11:1157–68
Bel G, Hagberg A, Meron E. 2012. Gradual regime shifts in spatially extended ecosystems. Theor. Ecol. 5:591–
604
Benincá E, Dakos V, van Nes EH, Huisman J, Scheffer M. 2011. Resonance of plankton communities with
temperature fluctuations. Am. Nat. 178:E85–95
Benincà E, Huisman J, Heerkloss R, Jöhnk KD, Branco P, et al. 2008. Chaos in a long-term experiment with
a plankton community. Nature 451:823–26
Berglund N, Gentz B. 2002. Metastability in simple climate models: pathwise analysis of slowly driven Langevin
equations. Stoch. Dyn. 2:327–56
Bestelmeyer BT, Duniway MC, James DK, Burkett LM, Havstad KM. 2013. A test of critical thresholds
and their indicators in a desertification-prone ecosystem: more resilience than we thought. Ecol. Lett.
16:339–45
Biggs R, Carpenter SR, Brock WA. 2009. Turning back from the brink: detecting an impending regime shift
in time to avert it. PNAS 106:826–31

www.annualreviews.org • Generic Indicators of Ecological Resilience 163

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Bjornstad ON, Grenfell BT. 2001. Noisy clockwork: time series analysis of population fluctuations in animals.
Science 293:638–43
Boerlijst MC, Oudman T, de Roos AM. 2013. Catastrophic collapse can occur without early warning: examples
of silent catastrophes in structured ecological models. PLOS ONE 8:e62033
Boettiger C, Hastings A. 2012. Quantifying limits to detection of early warning for critical transitions. J. R.
Soc. Interface 9:2527–39
Brock WA, Carpenter SR. 2006. Variance as a leading indicator of regime shift in ecosystem services. Ecol.
Soc. 11:9
Brock WA, Carpenter SR. 2010. Interacting regime shifts in ecosystems: implication for early warnings. Ecol.
Monogr. 80:353–67
Brock WA, Carpenter SR. 2012. Early warnings of regime shift when the ecosystem structure is unknown.
PLOS ONE 7:e45586
Carpenter SR, Brock WA. 2006. Rising variance: a leading indicator of ecological transition. Ecol. Lett. 9:308–
15
Carpenter SR, Brock WA. 2010. Early warnings of regime shifts in spatial dynamics using the discrete Fourier
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

transform. Ecosphere 1:10


Carpenter SR, Brock WA. 2011. Early warnings of unknown nonlinear shifts: a nonparametric approach.
Ecology 92:2196–201
Carpenter SR, Brock WA, Cole JJ, Kitchell JF, Pace ML. 2008. Leading indicators of trophic cascades. Ecol.
Lett. 11:128–38
Carpenter SR, Brock WA, Cole JJ, Pace ML. 2014. A new approach for rapid detection of nearby thresholds
in ecosystem time series. Oikos 123:290–97
Carpenter SR, Cole JJ, Pace ML, Batt R, Brock WA, et al. 2011. Early warnings of regime shifts: a whole-
ecosystem experiment. Science 332:1079–82
Chisholm RA, Filotas E. 2009. Critical slowing down as an indicator of transitions in two-species models.
J. Theor. Biol. 257:142–49
Cline TJ, Seekell DA, Carpenter SR, Pace ML, Hodgson JR, et al. 2014. Early warnings of regime shifts:
evaluation of spatial indicators from a whole-ecosystem experiment. Ecosphere 5:102
Contamin R, Ellison AM. 2009. Indicators of regime shifts in ecological systems: What do we need to know
and when do we need to know it? Ecol. Appl. 19:799–816
Cooper WS. 1984. Expected time to extinction and the concept of fundamental fitness. J. Theor. Biol. 107:603–
29
Dai L, Korolev KS, Gore J. 2013. Slower recovery in space before collapse of connected populations. Nature
496:355–58
Dai L, Vorselen D, Korolev KS, Gore J. 2012. Generic indicators for loss of resilience before a tipping point
leading to population collapse. Science 336:1175–77
Dakos V, Bascompte J. 2014. Critical slowing down as early warning for the onset of collapse in mutualistic
communities. PNAS 111:17546–51
Dakos V, Carpenter SR, Brock WA, Ellison AM, Guttal V, et al. 2012a. Methods for detecting early warnings
of critical transitions in time series illustrated using simulated ecological data. PLOS ONE 7:e41010
Dakos V, Kéfi S, Rietkerk M, van Nes EH, Scheffer M. 2011. Slowing down in spatially patterned ecosystems
at the brink of collapse. Am. Nat. 177:E153–66
Dakos V, Scheffer M, van Nes EH, Brovkin V, Petoukhov V, Held H. 2008. Slowing down as an early warning
signal for abrupt climate change. PNAS 105:14308–12
Dakos V, van Nes EH, D’Odorico P, Scheffer M. 2012b. Robustness of variance and autocorrelation as
indicators of critical slowing down. Ecology 93:264–71
Dakos V, van Nes EH, Donangelo R, Fort H, Scheffer M. 2010. Spatial correlation as leading indicator of
catastrophic shifts. Theor. Ecol. 3:163–74
Dakos V, van Nes EH, Scheffer M. 2013. Flickering as an early warning signal. Theor. Ecol. 6:309–17
DeAngelis DL. 1980. Energy-flow, nutrient cycling, and ecosystem resilience. Ecology 61:764–71
DeAngelis DL, Waterhouse JC. 1987. Equilibrium and nonequilibrium concepts in ecological models. Ecol.
Monogr. 57:1–21

164 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

DeVantier LM, De’ath G, Done TJ, Turak E. 1998. Ecological assessment of a complex natural system: a
case study from the Great Barrier Reef. Ecol. Appl. 8:480–96
Donangelo R, Fort H, Dakos V, Scheffer M, van Nes EH. 2010. Early warnings for catastrophic shifts in
ecosystems: comparison between spatial and temporal indicators. Int. J. Bifurc. Chaos 20:315–21
Drake JM, Griffen BD. 2010. Early warning signals of extinction in deteriorating environments. Nature
467:456–59
Elser JJ, Elser TJ, Carpenter SR, Brock WA. 2014. Regime shift in fertilizer commodities indicates more
turbulence ahead for food security. PLOS ONE 9:e93998
Fung T, Seymour RM, Johnson CR. 2013. Warning signals of regime shifts as intrinsic properties of endoge-
nous dynamics. Am. Nat. 182:208–22
Fussmann GF, Ellner SP, Shertzer KW, Hairston NG. 2000. Crossing the Hopf bifurcation in a live predator-
prey system. Science 290:1358–60
Gandhi A, Levin S, Orszag S. 1998. “Critical slowing down” in time-to-extinction: an example of critical
phenomena in ecology. J. Theor. Biol. 192:363–76
Gardiner CW. 2004. Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences. New York:
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

Springer
Gelcich S, Hughes TP, Olsson P, Folke C, Defeo O, et al. 2010. Navigating transformations in governance
of Chilean marine coastal resources. PNAS 107:16794–99
Grasman J, Van Herwaarden OA. 2010. Asymptotic Methods for the Fokker-Planck Equation and the Exit Problem
in Applications. Berlin/Heidelberg, Ger.: Springer-Verlag
Grimm V, Wissel C. 1997. Babel, or the ecological stability discussions: an inventory and analysis of termi-
nology and a guide for avoiding confusion. Oecologia 109:323–34
Guttal V, Jayaprakash C. 2008. Changing skewness: an early warning signal of regime shifts in ecosystems.
Ecol. Lett. 11:450–60
Guttal V, Jayaprakash C. 2009. Spatial variance and spatial skewness: leading indicators of regime shifts in
spatial ecological systems. Theor. Ecol. 2:3–12
Hansen MC, DeFries RS, Townshend JRG, Carroll M, Dimiceli C, Sohlberg RA. 2003. Global percent
tree cover at a spatial resolution of 500 meters: first results of the MODIS vegetation continuous fields
algorithm. Earth Interact. 7:1–15
Hanson FB, Tuckwell HC. 1978. Persistence times of populations with large random fluctuations. Theor.
Popul. Biol. 14:46–61
Hastings A. 2004. Transients: the key to long-term ecological understanding? Trends Ecol. Evol. 19:39–45
Hastings A, Wysham DB. 2010. Regime shifts in ecological systems can occur with no warning. Ecol. Lett.
13:464–72
Held H, Kleinen T. 2004. Detection of climate system bifurcations by degenerate fingerprinting. Geophys. Res.
Lett. 31:L23207
Hewitt JE, Thrush SF. 2010. Empirical evidence of an approaching alternate state produced by intrinsic
community dynamics, climatic variability and management actions. Mar. Ecol. Prog. Ser. 413:267–76
Hirota M, Holmgren M, van Nes EH, Scheffer M. 2011. Global resilience of tropical forest and savanna to
critical transitions. Science 334:232–35
Holling CS. 1973. Resilience and stability of ecological systems. Annu. Rev. Ecol. Syst. 4:1–23
Holling CS. 1996. Engineering resilience versus ecological resilience. In Engineering Within Ecological Con-
straints, ed. PC Schulze, pp. 31–43. Washington, DC: Natl. Acad
Hsieh CH, Reiss CS, Hunter JR, Beddington JR, May RM, Sugihara G. 2006. Fishing elevates variability in
the abundance of exploited species. Nature 443:859–62
Ives AR. 1995. Measuring resilience in stochastic systems. Ecol. Monogr. 65:217–33
Ives AR, Dakos V. 2012. Detecting dynamical changes in nonlinear time series using locally linear state-space
models. Ecosphere 3:58
Kéfi S, Dakos V, Scheffer M, van Nes EH, Rietkerk M. 2013. Early warning signals also precede non-
catastrophic transitions. Oikos 122:641–48
Kéfi S, Guttal V, Brock WA, Carpenter SR, Ellison AM, et al. 2014. Early warning signals of ecological
transitions: methods for spatial patterns. PLOS ONE 9:e92097

www.annualreviews.org • Generic Indicators of Ecological Resilience 165

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Kéfi S, Rietkerk M, Alados CL, Pueyo Y, Papanastasis VP, et al. 2007. Spatial vegetation patterns and imminent
desertification in Mediterranean arid ecosystems. Nature 449:213–17
Keitt TH, Lewis MA, Holt RD. 2001. Allee effects, invasion pinning, and species’ borders. Am. Nat. 157:203–
16
Kleinen T, Held H, Petschel-Held G. 2003. The potential role of spectral properties in detecting thresholds
in the Earth system: application to the thermohaline circulation. Ocean Dyn. 53:53–63
Kosten S, Vernooij M, van Nes EH, Sagrario MÁG, Clevers JGPW, Scheffer M. 2012. Bimodal transparency
as an indicator for alternative states in South American lakes. Freshw. Biol. 57:1191–201
Krkosek M, Drake JM. 2014. On signals of phase transitions in salmon population dynamics. Proc. R. Soc. B
281:20133221
Kuehn C. 2013. Warning signs for wave speed transitions of noisy Fisher–KPP invasion fronts. Theor. Ecol.
6:295–308
Lade SJ, Gross T. 2012. Early warning signals for critical transitions: a generalized modeling approach. PLOS
Comput. Biol. 8:e1002360
Lahti L, Salojärvi J, Salonen A, Scheffer M, De Vos WM. 2014. Tipping elements in the human intestinal
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org
Access provided by University of Otago on 10/14/15. For personal use only.

ecosystem. Nat. Commun. 5:4344


Leigh EG Jr. 1981. The average lifetime of a population in a varying environment. J. Theor. Biol. 90:213–39
Leung HK. 2000. Bifurcation of synchronization as a nonequilibrium phase transition. Physica A 281:311–17
Lever JJ, van Nes EH, Scheffer M, Bascompte J. 2014. The sudden collapse of pollinator communities. Ecol.
Lett. 17:350–59
Lewontin RC. 1969. The meaning of stability. In Diversity and Stability in Ecological Systems: Report of a Sym-
posium Held May 26–28, pp. 13–24. Springfield, VA: Clearinghouse for Federal Scientific and Technical
Information
Lindegren M, Dakos V, Gröger JP, Gårdmark A, Kornilovs G, et al. 2012. Early detection of ecosystem
regime shifts: a multiple method evaluation for management application. PLOS ONE 7:e38410
Livina VN, Kwasniok F, Lenton TM. 2010. Potential analysis reveals changing number of climate states
during the last 60 kyr. Clim. Past 6:77–82
Livina VN, Lenton TM. 2007. A modified method for detecting incipient bifurcations in a dynamical system.
Geophys. Res. Lett. 34:L03712
Maltby L, Blake N, Brock TCM, van den Brink PJ. 2005. Insecticide species sensitivity distributions: impor-
tance of test species selection and relevance to aquatic ecosystems. Environ. Toxicol. Chem. 24:379–88
Matsumoto G, Kunisawa T. 1978. Critical slowing-down near transition region from resting to time-ordered
states in squid giant axons. J. Phys. Soc. Jpn. 44:1047–48
Murray JD. 2002. Mathematical Biology: I. An Introduction. Berlin: Springer-Verlag
Oborny B, Meszéna G, Szabó G. 2005. Dynamics of populations on the verge of extinction. Oikos 109:291–96
O’Regan SM, Drake JM. 2013. Theory of early warning signals of disease emergence and leading indicators
of elimination. Theor. Ecol. 6:333–57
Perretti CT, Munch SB. 2012. Regime shift indicators fail under noise levels commonly observed in ecological
systems. Ecol. Appl. 22:1772–79
Pimm SL. 1984. The complexity and stability of ecosystems. Nature 307:321–26
Pomeau Y. 1986. Front motion, metastability and subcritical bifurcations in hydrodynamics. Phys. D Nonlinear
Phenom. 23:3–11
Rietkerk M, Dekker SC, de Ruiter PC, van de Koppel J. 2004. Self-organized patchiness and catastrophic
shifts in ecosystems. Science 305:1926–29
Rietkerk M, van de Koppel J. 2008. Regular pattern formation in real ecosystems. Trends Ecol. Evol. 23:169–75
Rinaldi S, Muratori S. 1992. Slow-fast limit cycles in predator-prey models. Ecol. Model. 61:287–308
Rinaldi S, Scheffer M. 2000. Geometric analysis of ecological models with slow and fast processes. Ecosystems
3:507–21
Rosenzweig ML. 1971. Paradox of enrichment: destabilization of exploitation ecosystems in ecological time.
Science 171:385–87
Sæther BE, Engen S, Islam A, McCleery R, Perrins C. 1998. Environmental stochasticity and extinction risk
in a population of a small songbird, the great tit. Am. Nat. 151:441–50

166 Scheffer et al.

Changes may still occur before final publication online and in print
ES46CH07-Scheffer ARI 11 September 2015 10:58

Scheffer M. 2009. Critical Transitions in Nature and Society. Princeton, NJ: Princeton Univ. Press
Scheffer M, Bascompte J, Brock WA, Brovkin V, Carpenter SR, et al. 2009. Early-warning signals for critical
transitions. Nature 461:53–59
Scheffer M, Carpenter SR, Foley JA, Folke C, Walker B. 2001. Catastrophic shifts in ecosystems. Nature
413:591–96
Scheffer M, Carpenter SR, Lenton TM, Bascompte J, Brock W, et al. 2012a. Anticipating critical transitions.
Science 338:344–48
Scheffer M, Hirota M, Holmgren M, van Nes EH, Chapin FS. 2012b. Thresholds for boreal biome transitions.
PNAS 109:21384–89
Seekell DA, Carpenter SR, Cline TJ, Pace ML. 2012. Conditional heteroskedasticity forecasts regime shift in
a whole-ecosystem experiment. Ecosystems 15:741–47
Seekell DA, Carpenter SR, Pace ML. 2011. Conditional heteroscedasticity as a leading indicator of ecological
regime shifts. Am. Nat. 178:442–51
Strogatz SH. 1994. Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry and Engineering.
Reading, MA: Addison-Wesley
Annu. Rev. Ecol. Evol. Syst. 2015.46. Downloaded from www.annualreviews.org

Suweis S, D’Odorico P. 2014. Early warning signs in social-ecological networks. PLOS ONE 9:e101851
Access provided by University of Otago on 10/14/15. For personal use only.

Suzuki K, Yoshida T. 2015. Ecological resilience of population cycles: a dynamic perspective of regime shift.
J. Theor. Biol. 370:103–15
Takimoto G. 2009. Early warning signals of demographic regime shifts in invading populations. Popul. Ecol.
51:419–26
Tirabassi G, Viebahn J, Dakos V, Dijkstra HA, Masoller C, et al. 2014. Interaction network based early-
warning indicators of vegetation transitions. Ecol. Complex. 19:148–57
Turing AM. 1952. The chemical basis of morphogenesis. Philos. Trans. R. Soc. B 237:37–72
van de Koppel J, Gascoigne JC, Theraulaz G, Rietkerk M, Mooij WM, Herman PMJ. 2008. Experimental
evidence for spatial self-organization and its emergent effects in mussel bed ecosystems. Science 322:739–42
van de Leemput IA, van Nes EH, Scheffer M. 2015. Resilience of alternative states in spatially extended
ecosystems. PLOS ONE 10:e0116859
van de Leemput IA, Wichers M, Cramer AOJ, Borsboom D, Tuerlinckx F, et al. 2014. Critical slowing down
as early warning for the onset and termination of depression. PNAS 111:87–92
van der Heide T, Bouma TJ, van Nes EH, van de Koppel J, Scheffer M, et al. 2010. Spatial self-organized
patterning in seagrasses along a depth gradient of an intertidal ecosystem. Ecology 91:362–69
van Nes EH, Hirota M, Holmgren M, Scheffer M. 2014. Tipping points in tropical tree cover: linking theory
to data. Glob. Change Biol. 20:1016–21
van Nes EH, Scheffer M. 2004. Large species shifts triggered by small forces. Am. Nat. 164:255–66
van Nes EH, Scheffer M. 2007. Slow recovery from perturbations as a generic indicator of a nearby catastrophic
shift. Am. Nat. 169:738–47
Vandermeer J. 2011. The inevitability of surprise in agroecosystems. Ecol. Complex. 8:377–82
Vandermeer J, Yodzis P. 1999. Basin boundary collision as a model of discontinuous change in ecosystems.
Ecology 80:1817–27
Veraart AJ, Faassen EJ, Dakos V, van Nes EH, Lürling M, Scheffer M. 2012. Recovery rates reflect distance
to a tipping point in a living system. Nature 481:357–59
von Hardenberg J, Meron E, Shachak M, Zarmi Y. 2001. Diversity of vegetation patterns and desertification.
Phys. Rev. Lett. 87:198101
Wang R, Dearing JA, Langdon PG, Zhang E, Yang X, et al. 2012. Flickering gives early warning signals of a
critical transition to a eutrophic lake state. Nature 492:419–22
Wissel C. 1984. A universal law of the characteristic return time near thresholds. Oecologia 65:101–7
Wouters N, Dakos V, Edwards M, Serafim MP, Valayer PJ, Cabral HN. 2015. Evidencing a regime shift
in the North Sea using early-warning signals as indicators of critical transitions. Estuar. Coast. Shelf Sci.
152:65–72

www.annualreviews.org • Generic Indicators of Ecological Resilience 167

Changes may still occur before final publication online and in print

You might also like