# Trophic coherence determines food-web stability

^{a}Warwick Mathematics Institute, and Centre for Complexity Science, University of Warwick, Coventry CV4 7AL, United Kingdom;^{b}Departamento de Electromagnetismo y Física de la Materia, and Instituto Carlos I de Física Teórica y Computacional, Universidad de Granada, 18071 Granada, Spain; and^{c}Departamento de Electrónica y Tecnología de Computadores and Centro de Investigación en Tecnologías de la Información y de las Comunicaciones, Universidad de Granada, 18071 Granada, Spain

See allHide authors and affiliations

Edited* by Robert M. May, University of Oxford, Oxford, United Kingdom, and approved November 3, 2014 (received for review May 16, 2014)

## Significance

The fact that large, complex ecosystems are particularly robust is mysterious in the light of mathematical arguments that suggest they should be unstable; i.e., susceptible to runaway fluctuations in species’ abundances. Here we show that food webs (networks describing who eats whom in an ecosystem) exhibit a property we call trophic coherence, a measure of how neatly the species fall into distinct levels. We find that this property makes networks far more linearly stable than if the links (predator–prey interactions) were placed randomly between species, or according to existing structural models. A simple model we propose to capture this feature shows that networks can, in fact, become more stable with size and complexity, suggesting a possible solution to the paradox.

## Abstract

Why are large, complex ecosystems stable? Both theory and simulations of current models predict the onset of instability with growing size and complexity, so for decades it has been conjectured that ecosystems must have some unidentified structural property exempting them from this outcome. We show that trophic coherence—a hitherto ignored feature of food webs that current structural models fail to reproduce—is a better statistical predictor of linear stability than size or complexity. Furthermore, we prove that a maximally coherent network with constant interaction strengths will always be linearly stable. We also propose a simple model that, by correctly capturing the trophic coherence of food webs, accurately reproduces their stability and other basic structural features. Most remarkably, our model shows that stability can increase with size and complexity. This suggests a key to May’s paradox, and a range of opportunities and concerns for biodiversity conservation.

In the early seventies, Robert May addressed the question of whether a generic system of coupled dynamical elements randomly connected to each other would be stable. He found that the larger and more interconnected the system, the more difficult it would be to stabilize (1, 2). His deduction followed from the behavior of the leading eigenvalue of the interaction matrix, which, in a randomly wired system, grows with the square root of the mean number of links per element. This result clashed with the received wisdom in ecology—that large, complex ecosystems were particularly stable—and initiated the “diversity–stability debate” (3⇓⇓–6). Indeed, Charles Elton had expressed the prevailing view in 1958: “the balance of relatively simple communities of plants and animals is more easily upset than that of richer ones; that is, more subject to destructive oscillations in populations, especially of animals, and more vulnerable to invasions” (7). Even if this description were not accurate, the mere existence of rainforests and coral reefs seems incongruous with a general mathematical principle that “complexity begets instability,” and has become known as May’s paradox.

One solution might be that the linear stability analysis used by May and many subsequent studies does not capture essential characteristics of ecosystem dynamics, and much work has gone into exploring how more accurate dynamical descriptions might enhance stability (5, 8, 9). However, as ever-better ecological data are gathered, it is becoming apparent that the leading eigenvalues of matrices related to food webs (networks in which the species are nodes and the links represent predation) do not exhibit the expected dependence on size or link density (10). Food webs must, therefore, have some unknown structural feature that accounts for this deviation from randomness—irrespectively of other stabilizing factors.

We show here that a network feature we call trophic coherence accounts for much of the variance in linear stability observed in a dataset of 46 food webs, and we prove that a perfectly coherent network with constant link strengths will always be stable. Furthermore, a simple model that we propose to capture this property suggests that networks can become more stable with size and complexity if they are sufficiently coherent.

## Results

### Trophic Coherence and Stability.

Each species in an ecosystem is generally influenced by others, via processes such as predation, parasitism, mutualism, or competition for various resources (11⇓⇓–14). A food web is a network of species that represents the first kind of influence with directed links (arrows) from each prey node to its predators (15⇓⇓–18). Such representations can therefore be seen as transport networks, where biomass originates in the basal species (the sources) and flows through the ecosystem, some of it reaching the apex predators (the sinks).

The trophic level of a species can be defined as the average trophic level of its prey, plus 1 (19, 20). Thus, plants and other basal species are assigned level 1, pure herbivores have level 2, but many species will have fractional values. A species’ trophic level provides a useful measure of how far it is from the sources of biomass in its ecosystem. We can characterize each link in a network with a trophic distance, defined as the difference between the trophic levels of the predator and prey species involved (it is not a true “distance” in the mathematical sense, because it can be negative). We then look at the distribution of trophic distances over all links in a given network. The mean of this distribution will always be equal to 1, and we refer to its degree of homogeneity as the network’s trophic coherence. We shall measure this degree of order with the SD of the distribution of trophic distances, *q* (we avoid using the symbol σ because it is often assigned to the SD in link strengths). A perfectly coherent network, in which all distances are equal to 1 (implying that each species occupies an integer trophic level), has *q* as an incoherence parameter. (For a technical description of these measures, see *Methods*.)

A fundamental property of ecosystems is their ability to endure over time (13, 18). “Stability” is often used as a generic term for any measure of this characteristic, including for concepts such as robustness and resilience (21). When the analysis regards the possibility that a small perturbation in population densities could amplify into runaway fluctuations, stability is usually understood in the sense of Lyapunov stability, which in practice tends to mean linear stability (22). This is the sense we shall be interested in here, and henceforth stability will mean linear stability. Given the equations for the dynamics of the system, a fixed (or equilibrium) point will be linearly stable if all of the eigenvalues of the Jacobian matrix evaluated at this point have negative real part. Even without precise knowledge of the dynamics, one can still apply this reasoning to learn about the stability of a system just from the network structure of interactions between elements (in this case, species whose trophic interactions are described by a food web) (2, 23⇓–25). In *Methods* (and, more extensively, in *SI Appendix*), we describe how an interaction matrix *W* can be derived from the adjacency (or predation) matrix *A* representing a food web, such that the real part of *W*’s leading eigenvalue, *R*, the more unstable the food web. For the simple yet ecologically unrealistic case in which the extent to which a predator consumes a prey species is proportional to the sum of their (biomass) densities, the Jacobian coincides with *W*, and *R* describes the stability for any configuration of densities (global stability). For more realistic dynamics—such as Lotka–Volterra, type II or type III—the Jacobian must be evaluated at a given point, but we show that the general form can still be related to *W* (*Methods*). Furthermore, by making assumptions about the biomass distribution, it is possible to check our results for such dynamics (*SI Appendix*). In the main text, however, we shall focus simply on the matrix *W* without making any further assumptions about dynamics or biomass distributions.

May considered a generic Jacobian in which link strengths were drawn from a random distribution, representing all kinds of ecological interactions (1, 2). Because, in this setting, the expected value of the real part of the leading eigenvalue (*R*) should grow with *S* is the number of species and *C* the probability that a pair of them be connected, larger and more interconnected ecosystems should be less stable than small, sparse ones (26). (Allesina and Tang have recently obtained stability criteria for random networks with specific kinds of interactions: although predator–prey relationships are more conducive to stability than competition or mutualism, even a network consisting only of predator–prey interactions should become more unstable with increasing size and link density; ref. 27.)

We analyze the stability for each of a set of 46 empirical food webs from several kinds of ecosystem (the details and references for these can be found in *SI Appendix*). In Fig. 1*A* we plot the *R* of each web against *B* shows *R* against *R* and *SI Appendix* we also compare the empirical *R* values to the estimate derived by Allesina and Tang for random networks in which all links are predator–prey. Surprisingly, the correlation is lower than for

Might this feature be trophic coherence? In Fig. 1*C* we plot *R* for the same food webs against the incoherence parameter *q*. The correlation is significantly stronger than with complexity – stability increases with coherence. However, there are still outliers, such as the food web of Coachella Valley. We note that although most forms of intraspecies competition are not described by the interaction matrix, there is one form which is: cannibalism. This fairly common practice is a well-known kind of self-regulation that contributes to the stability of a food web (mathematically, negative elements in the diagonal of the interaction matrix shift its eigenvalues leftwards along the real axis). In Fig. 1*D* we therefore plot the *R* and *q* we obtain after removing all self-links. Now Pearson’s correlation coefficient is *SI Appendix*). In *SI Appendix*, we measure stability according to Lotka–Volterra, type II and type III dynamics, and show that in every case trophic coherence is the best predictor of stability.

### Modeling Food-Web Structure.

Many mathematical models have been put forward to simulate various aspects of food webs (18, 26, 28, 29⇓⇓⇓⇓–34). We shall focus here on so-called structural, or static, models: those which attempt to reproduce properties of food-web structure with a few simple rules. The best known is Williams and Martinez’s niche model (35, 36). This is an elegant way of generating nontrivial networks by randomly assigning each species to a position on a “niche axis,” together with a range of axis centered at some lower niche value. Each species then consumes all other species lying within its range of axis, and none without. The idea is that the axis represents some intrinsic hierarchy among species that determines who can prey on whom. The niche model is itself based on Cohen and Newman’s cascade model, which also has an axis, but species are randomly assigned prey from among all those with lower niche values than themselves (37). Stouffer et al. proposed the generalized niche model, in which some of a species’ prey are set according to the niche model and the rest ensue from a slightly refined version of the cascade model, the proportion of each being determined by a contiguity parameter (38). [The generalized cascade model is like the original cascade model except that the numbers of prey species are drawn from the beta distribution used in the niche model and subsequent niche-based models (39). This is the version of the model used throughout this paper, as explained in *SI Appendix*.] The minimum potential niche model of Allesina et al. is similar, but includes (random) forbidden links within species’ ranges, instead of extra ones, as a way of emulating the effects of more than one axis—with the advantage that all of the links of real food webs have a nonzero probability of being generated by this model (40). Meanwhile, the nested hierarchy model of Cattin et al. takes into account that phylogenetically close species are more likely to share prey than unrelated ones (41). (For details of the models, see *SI Appendix*.)

These models produce networks with many of the statistical properties of food webs (36, 39, 40). However, as we go on to show below, they tend to predict significantly less trophic coherence (larger *q*) than we observe in our dataset. We therefore propose the preferential preying model (PPM) as a way of capturing this feature. We begin with *B* nodes (basal species) and no links. We then add new nodes (consumer species) sequentially to the system until we have a total of *S* species, assigning each their prey from among available nodes in the following way. The first prey species is chosen randomly, and the rest are chosen with a probability that decays exponentially with their absolute trophic distance to that initial prey species (i.e., with the absolute difference of trophic levels). This probability is set by a parameter *T* that determines the degree of trophic specialization of consumers. The number of prey is drawn from a beta distribution with a mean value proportional to the number of available species, just as the other structural models described use a mean value proportional to the niche value. (For a more detailed description, see *Methods*.)

The PPM is reminiscent of Barabási and Albert’s model of evolving networks (42), but it is also akin to a highly simplified version of an “assembly model” in which species enter via immigration (29, 32). It assumes that if a given species has adapted to prey off species A, it is more likely to be able to consume species B as well if A and B have similar trophic levels than if not. It may seem that this scheme is similar in essence to the niche model, with the role of niche axis being played by the trophic levels. However, whereas the niche values given to species in niche-based models are hidden variables, meant to represent some kind of biological magnitude, the trophic level of a node is defined by the emerging network architecture itself. We shall see that this difference has a crucial effect on the networks generated by each model.

### The Origins of Stability.

Fig. 2*A* shows three networks with varying degrees of trophic coherence. The one on the left was generated with the PPM and *B* shows how trophic coherence varies with *T* in PPM networks. At about *q* for networks generated with the generalized niche model against “diet contiguity,” *c*, its only free parameter (38). At *SI Appendix*). However, diet contiguity has little effect on trophic coherence.

Fig. 2*C* shows the stability—as measured by *R*, the leading eigenvalue of the interaction matrix—for the networks of Fig. 2*B*. For the PPM networks, stability closely mirrors trophic coherence: as *T* decreases, the networks become more stable (smaller *R*) as well as more coherent (smaller *q*). The empirical value of *R* is obtained at about the same *T* that best approximates the empirical *q*. The generalized niche model also generates more stable networks as diet contiguity is increased, but this effect cannot be due to trophic coherence, which remains nearly constant. The origin of increasing stability in this model is revealed when we measure *R* after removing all self-links from the networks): the generalized niche model now displays only a very small dependence of stability on diet contiguity. In contrast, the behavior of *T* in the PPM networks remains qualitatively the same as in the previous case, and the empirical stability continues to be obtained at *R* and

We perform this analysis for each of the 46 food webs in our dataset, obtaining the value of *T* that best captures the empirical trophic coherence according to the PPM. We then compute the ensemble averages of *R* and *T*, for comparison with the empirical values. Similarly, we compute the average values of these measures predicted by each of the niche-based models described above: the cascade, niche, nested hierarchy, generalized niche, and minimum potential niche models. The last two models have free parameters, but as these do not have a significant effect on trophic coherence, we use the values published as optimal in refs. 39 and 40, respectively (or the mean optimal values for those food webs that were not analyzed in these papers). Fig. 3 *A–C* shows the average absolute deviations from the empirical values for trophic coherence and stability, before and after removing self-links, for each model. In Fig. 3*A* we observe that, as mentioned above, the niche-based models fail to capture the trophic coherence of these food webs. Stability, with or without considering self-links, is predicted by the PPM significantly better than by any of the other models, as shown in Fig. 3 *B* and *C*. This is in keeping with Allesina and Tang’s observation that current structural models cannot account for food-web stability (27). In *SI Appendix* we show the results of similar model comparisons for several other network measures: modularity, mean chain length, mean trophic level, and numbers of cannibals and of apex predators. The PPM does as well as any of the other models in regards to the numbers of cannibals and apex predators, and is significantly better at predicting the other measures. Allesina et al. have developed a likelihood-based approach for comparing food-web models (40). We have not yet been able to obtain the corresponding likelihoods for the PPM, but if this is done in the future it would provide a firmer basis from which to gauge the models’ relative merits, and perhaps to build a more realistic model drawing on each one’s strengths.

Why does the trophic coherence of networks determine their stability? The case of a maximally coherent structure, with *A*), is amenable to mathematical analysis. In *SI Appendix* we consider the undirected network that results from replacing each directed link of the predation matrix with a symmetric link, the nonzero eigenvalues of which always come in pairs of real numbers

Although a general, analytical relationship between trophic coherence and stability remains elusive, it is intuitive to expect that a deviation from maximal coherence will drive the real part of the leading eigenvalue toward the positive values established for random structures, as is indeed observed in our simulations.

### May’s Paradox.

As we have seen, the PPM can predict the stability of a food web quite accurately just with information regarding numbers of species, basal species and links, and trophic coherence. But what does this tell us about May’s paradox—the fact that large, complex ecosystems seem to be particularly stable despite theoretical predictions to the contrary? To ascertain how stability scales with size, *S*, and complexity, *K*, in networks generated by different models, we must first determine how *K* scales with *S*; i.e., if

In Fig. 3*D* we show how stability scales with *S* in each of the niche-based models when complexity increases with size according to *S*. This scaling is indeed closely matched by the cascade model. The behavior of the other models is similar (except for the nested hierarchy model, in which *R* increases more rapidly at high *S*), and, as expected, networks always become less stable with increasing size and complexity. In Fig. 3*E* we show how the stability of PPM networks scales in the same scenario. For high *T*, their behavior is similar to that of the cascade model: *T* is lowered, until, for sufficiently large and coherent networks, it becomes negative; in other words, stability increases with size and complexity. Fig. 3*E*, *Inset*, shows the exponent γ obtained against *T*, for different values of α. The smaller α, the larger the range of *T* that yields a positive complexity–stability relationship. [Plitzko et al. recently showed that there exists a range of parameters (in a generalized modeling framework; ref. 46) for which niche model networks can increase in stability with complexity (47). However, for this study networks were rejected unless they were stable and had exactly four trophic levels. This selection may have screened for trophic coherence, cannibalism, or other structural features.]

In *SI Appendix* we extend this analysis to specific dynamics—Lotka–Volterra, type II and type III—by assuming an exponential relationship between biomass and trophic level, which can be described as a pyramid. The positive complexity–stability relationship does not appear to depend on the details of dynamics. However, the slope of the

## Discussion

The predation matrices corresponding to real ecosystems are clearly peculiar in some way, because their largest eigenvalues do not depend solely on their size or complexity, as we would expect both from random graph theory and structural food-web models. This is in keeping with the empirical observation that large, complex ecosystems are particularly stable, but challenges current thinking on food-web architecture. We have shown that the structural property we call trophic coherence is significantly correlated with food-web stability, despite other differences between the ecosystems and the variety of empirical methods used in gathering the data. In fact, cannibalism and trophic coherence together account for most of the variance in stability observed in our dataset. Furthermore, we have proved that a maximally coherent food web with constant interaction strengths will always be stable.

We have suggested the preferential preying model as a simple algorithm for generating networks with tunable trophic coherence. Although this model does not attempt to replicate other characteristic features of food webs, such as a phylogenetic signal or body-size effects, it reproduces the empirical stability of the 46 webs analyzed quite accurately once its only free parameter has been adjusted to the empirical degree of trophic coherence. Most remarkably, the model predicts that networks should become more stable with increasing size and complexity, as long as they are sufficiently coherent and the number of links does not grow too fast with size. Although this result should be followed up with further analytical and empirical research, it suggests that we need no longer be surprised at the high stability of large, complex ecosystems.

We must caution that these findings do not imply that trophic coherence was somehow selected for by the forces of nature to improve food-web function. It seems unlikely that there should be any selective pressure on the individuals making up a species to do what is best for their ecosystem. Rather, many biological features of a species are associated with its trophic level. Therefore, adaptations which allow a given predator to prey on species A are likely to be useful also in preying on species B if A and B have similar trophic levels. This leads to trophic coherence, which results in high stability.

If stability decreased with size and complexity, as previous theoretical studies have assumed, ecosystems could not grow indefinitely, for they would face a cutoff point beyond which they would become unstable (26). On the other hand, if real ecosystems are coherent enough that they become more stable with size and complexity, as our model predicts, then the reverse might be true. We must also bear in mind, however, that our results are only for linear stability, whereas structural stability, for instance, may depend differently on size and coherence, and could become the limiting factor (18). In any case, ascertaining whether the loss of a few species would stabilize or destabilize a given community could be important for conservation efforts, particularly for averting tipping points (14).

The findings we report here came about by studying food webs. However, directed networks of many kinds transport energy, matter, information, capital, or other entities in a similar way to how food webs carry biomass from producers to apex predators. It seems likely that the relation between a network’s trophic coherence and its leading eigenvalue will be of consequence to other disciplines, and perhaps the preferential preying model, although overly simplistic for many scenarios, may serve as a first approximation for looking into these effects in a variety of systems.

## Methods

### Measuring Stability.

Let us assume that the populations of species making up an ecosystem (each characterized by its total biomass) change through time according to some set of nonlinear differential equations, the interactions determined by the predation matrix, *A* (whose elements *i* preys on species *j*, and 0 otherwise). If the system persists without suffering large changes it must, one assumes, find itself in the neighborhood of a fixed point of the dynamics. We can study how the system would react to a small perturbation by expanding the equations of motion around this fixed point and keeping only linear terms. The subsequent effect of the perturbation is then determined by the corresponding Jacobian matrix, and the system will tend to return to the fixed point only if the real parts of all its eigenvalues are negative (22).

Even without knowledge of the details of the dynamics, it is possible to draw some conclusions about the stability of a food web solely from its predation matrix (23). Independently of the exact interaction strengths, we know that not all of the biomass lost by a prey species when consumed goes to form part of the predator; in fact, this efficiency is relatively low (48). It is therefore natural to assume that the effect of species *j* on species *i* will be mediated by *W*, thereby shifting its set of eigenvalues (or spectrum) leftward along the real axis. Therefore, the eigenvalue with largest real part of *W*, as defined above, can be seen as a measure of the minimum intraspecies competition required for the system to be stable. Thus, the lower this value,

In *SI Appendix*, we describe this analysis in more detail. Beginning with a general consumer–resource differential equation for the biomass of each species, we obtain the Jacobian in terms of the function *i* consumes species *j*. For the simple (and unrealistic) case *W* as given above, independently of the fixed point. For more realistic dynamics, the Jacobian depends on the fixed point. For instance, for the Lotka–Volterra function *h* the Hill coefficient), we have either type II

The Jacobians for Lotka–Volterra, type II and type III dynamics are all similar in form to the matrix *W*, although for an exact solution we require the fixed point. In the main text we therefore use the leading eigenvalue of *W* as a generic measure of stability. However, in *SI Appendix* we consider the effects that different kinds of biomass distribution have on each of these more realistic dynamics. The results are qualitatively the same as those for the matrix *W*, although we find that both the squatness of a biomass pyramid and the level of noise in this structure affect the strength of the diversity–stability relationship described in the main text.

This measure of stability depends on the parameter η. In *SI Appendix* we show that the results reported here remain qualitatively unchanged for any

### Trophic Levels and Coherence.

The trophic level *i* is defined as the average trophic level of its prey, plus 1 (19). That is,

where *i* (or *i*’s in degree), and *A*. Basal species (those with *A* are ones and zeros), we omit the link strength term usually included in Eq. **1** (19).

We can write Eq. **1** in terms of a modified graph Laplacian matrix, **s** is the vector of trophic levels, **v** is the vector with elements **1**.

We define the trophic distance spanned by each link *i* the average over its incoming links is *L* is the total number of links,

Trophic coherence bears a close resemblance to Levine’s measures of trophic specialization (19). However, our average is computed over links instead of species, with the consequence that we need not consider the distinction between resource and consumer specializations. It is also related to measures of omnivory: in general, the more omnivores one finds in a community, the less coherent the food web.

### The Preferential Preying Model.

We begin with *B* nodes (basal species) and no links. We then add, sequentially, *i* is first awarded a random node *j* from among all those available when it arrives. Then another *l* are chosen with a probability *j* and *l*. Specifically, we use the exponential form

where *j* is the first node chosen by *i*, and *T* is a parameter that sets the degree of trophic specialization of consumers.

The number of extra prey, *i* arrives, and

where *L* is the expected number of links. In this work, we only consider networks with a number of links within an error margin of 5% of the desired *L*; thus, for all of the results reported, we have imposed this filter on the PPM networks and those generated with the other models.

To allow for cannibalism, the new node *i* is initially considered to have a trophic level *i* has been assigned all its prey,

## Acknowledgments

We are grateful to J. Dunne and U. Jacob for providing food-web data and to J. Hidalgo for technical help. Many thanks also to N. S. Jones, D. C. Reuman, D. B. Stouffer, J. Bascompte, I. Mendoza, J. A. Bonachela, S. A. Levin, O. Al Hammal, A. Thierry, A. Maritan, N. Guisoni, G. Baglietto, E. Albano, B. Moglia, J. J. Torres, and J. A. Johnson for conversations and comments on versions of the manuscript. We acknowledge financial support from the Spanish Ministerio de Ciencia e Innovación, and the European Regional Development Fund, under Project FIS2009-08451 and from Junta de Andalucía Proyecto de Excelencia P09FQM-4682. S.J. is grateful for support from the Oxford Centre for Integrative Systems Biology, University of Oxford, and from the European Commission under the Marie Curie Intra-European Fellowship Programme PIEF-GA-2010-276454.

## Footnotes

↵

^{1}S.J. and V.D.-G. contributed equally to this work.- ↵
^{2}To whom correspondence should be addressed. Email: S.Johnson.2{at}warwick.ac.uk.

Author contributions: S.J., V.D.-G., and M.A.M. designed research; S.J., V.D.-G., L.D., and M.A.M. performed research; S.J., V.D.-G., and M.A.M. analyzed data; L.D. devised the proof for maximally coherent networks; and S.J., V.D.-G., and M.A.M. wrote the paper.

The authors declare no conflict of interest.

↵*This Direct Submission article had a prearranged editor.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1409077111/-/DCSupplemental.

## References

- ↵
- ↵.
- May RM

- ↵
- ↵
- ↵
- ↵.
- Dunne JA,
- Brose U,
- Williams RJ,
- Martinez ND

*Aquatic Food Webs: An Ecosystem Approach*(Oxford Univ Press, Oxford), pp 117–129 - ↵.
- Elton CS

- ↵
- ↵
- ↵.
- Jacquet C, et al.

*arXiv:1307.5364* - ↵.
- Elton CS

- ↵.
- Pimm SL

- ↵.
- Solé RV,
- Bascompte J

- ↵
- ↵.
- Pimm SL

- ↵
- ↵.
- Drossel B,
- McKane AJ

- ↵.
- Rossberg AG

- ↵
- ↵.
- Pauly D,
- Christensen V,
- Dalsgaard J,
- Froese R,
- Torres F Jr

- ↵
- ↵
- ↵
- ↵
- ↵.
- Logofet DO

- ↵
- ↵
- ↵
- ↵
- ↵.
- Loeuille N,
- Loreau M

- ↵
- ↵.
- McKane AJ,
- Drossel B

- ↵
- ↵.
- Rossberg AG,
- Brännström A,
- Dieckmann U

- ↵
- ↵
- ↵.
- Cohen JE,
- Newman CM

- ↵.
- Stouffer DB,
- Camacho J,
- Amaral LAN

- ↵
- ↵.
- Allesina S,
- Alonso D,
- Pascual M

- ↵
- ↵.
- Barabási AL,
- Albert R

- ↵
- ↵.
- Rossberg AG,
- Farnsworth KD,
- Satoh K,
- Pinnegar JK

- ↵
- ↵.
- Gross T,
- Rudolf L,
- Levin SA,
- Dieckmann U

- ↵
- ↵
- ↵

## Citation Manager Formats

## Article Classifications

- Biological Sciences
- Ecology

- Physical Sciences
- Applied Mathematics