q-bio0401025/jtb6.tex
1: \documentclass[12pt,a4paper]{article}
2: 
3: %\renewcommand{\baselinestretch}{1.735}
4: \addtolength{\textwidth}{3cm}
5: \addtolength{\oddsidemargin}{-1.5cm}
6: \addtolength{\evensidemargin}{-1.5cm}
7: \addtolength{\topmargin}{-1.5cm}
8: %\setlength{\textwidth}{16cm}
9: \usepackage[dvips]{graphicx}
10: 
11: \begin{document}
12: \thispagestyle{empty}
13: 
14: \mbox{}
15: \vspace{1cm}
16: \begin{center}
17: {\Large{\bf The impact of non-linear functional responses on the}}
18: 
19: \smallskip
20: 
21: {\Large{\bf long-term evolution of food web structure}}
22: 
23: \vspace{0.5cm}
24: 
25: {\it Barbara Drossel$^{\rm a}$, Alan J. McKane$^{\rm b}$, Christopher 
26: Quince$^{{\rm c}\,\dag}$} 
27: \\
28: \bigskip
29: $^{\rm\, a}$Institut f\"ur Festk\"orperphysik, Technische Universit\"at 
30: Darmstadt, Hochschulstr.~6, 64289 Darmstadt, Germany\\
31: $^{\rm\, b}$Department of Theoretical Physics, University of Manchester, \\
32: Manchester M13 9PL, UK \\
33: $^{\rm\, c}$Department of Physics and Astronomy, Arizona State University, \\
34: Tempe, AZ 85287-1504, USA \\
35: \end{center}
36: 
37: \vspace{2cm}
38: 
39: \begin{abstract}
40: We investigate the long-term web structure emerging in evolutionary food web 
41: models when different types of functional responses are used. We find that 
42: large and complex webs with several trophic layers arise only if the 
43: population dynamics is such that it allows predators to focus on their best 
44: prey species. This can be achieved using modified Lotka-Volterra or 
45: Holling/Beddington functional responses with effective couplings that depend 
46: on the predator's efficiency at exploiting the prey, or a ratio-dependent 
47: functional response with adaptive foraging. In contrast, if standard 
48: Lotka-Volterra or Holling/Beddington functional responses are used, long-term 
49: evolution generates webs with almost all species being basal, and with 
50: additionally many links between these species. Interestingly, in all cases 
51: studied, a large proportion of weak links result naturally from the evolution 
52: of the food webs.
53: \end{abstract}
54: 
55: \vspace{2cm}
56: 
57: Keywords: Food webs, functional responses, population dynamics, 
58: weak links
59: 
60: \vspace{2cm}
61: 
62: $^{\dag}\,$The authors are listed in alphabetical order.
63: 
64: \newpage
65: 
66: \section{Introduction}
67: \label{intro}
68: The debate over what stabilises complex food webs has been an active
69: one for over a quarter of a century (McCann, 2000).  While it is found
70: that the stability of a randomly linked food web model tends to decrease 
71: with the proportion of links and the number of species (May, 1972), real 
72: food webs display a high degree of stability, in spite of being very 
73: complex. Clearly, real food webs have features that are ignored in randomly 
74: linked models. For example, the distribution and strength
75: of links in real food webs are far from random. More realistic
76: approaches use link values taken from real food webs (Yodzis, 1981),
77: or assemble a food web by repeated addition of species from a large
78: species pool that contains different types of species such as
79: ``plants'', ``herbivores'', ``carnivores'', and ``top
80: predators''(Morton \& Law, 1997; Law, 1999). Using Lotka-Volterra type
81: dynamics, assembly models of this type lead to large complex
82: webs. These Lotka-Volterra type models lack nonlinear effects such as predator
83: saturation, which many consider important for food web dynamics. The 
84: species pool is usually composed of a large number of species which are 
85: generated in an ad hoc manner, by pre-defining the possible types and 
86: value ranges of links. A complementary approach to food web stability 
87: focuses on more realistic dynamics (that include predator saturation, for
88: instance) of small systems consisting only of a few species. 
89: These small systems can be viewed as a small part of a food web. It has 
90: been found that the presence of many weak links and the use of nonlinear 
91: growth rates, that prevent predators from feeding successfully on many prey 
92: at the same time, stabilise the dynamics of these small systems by reducing 
93: population fluctuations (McCann {\it et al.}, 1998). Again, the possible 
94: types and values of links are pre-defined in these studies.
95: 
96: In real food webs, the link strengths and the linkage pattern can
97: change with time through invasion and replacement with related
98: species, and through adaptive evolutionary processes (Thompson,
99: 1998). It is therefore desirable to investigate food web models that
100: implement such long-term changes due to modification of existing link
101: strengths and linkage patterns. In the following, we call this
102: long-term change of the food web ``evolution'', with the meaning of
103: ``development of the food web structure in time'', while the ``true''
104: biological evolution is of course happening on the level of
105: individuals.  Our approach is different from the species assembly
106: models insofar that the new species are not picked irrespective of the
107: web composition, but are similar to existing species.
108: 
109: Considering the long-term change of food webs adds a new dimension to
110: the complexity-stability debate: even though species die out and are
111: replaced with others during the course of time, ecosystems as a whole
112: persist in time. Except for rare catastrophic events due to external
113: causes, food webs preserve their general structure. This means that
114: typical structural features of food webs, such as the total number of
115: species or the number of trophic layers, do not collapse to zero, but
116: rather fluctuate around some mean value far away from zero. This
117: definition of stability is perhaps more relevant for understanding the
118: complexity of ecosystems, than some of the other definitions found in
119: the literature. Persistence in time of a complex ecosystem of course
120: also requires a certain kind of stability of the population dynamics:
121: while population dynamics does not need to reach fixed points,
122: oscillations must be small enough such that the population sizes of
123: the majority of species remain positive after the introduction of a
124: mutant or invading individual. If this were not so, each change in the 
125: web composition would lead to a collapse of the web, and complex
126: ecosystems could never arise. 
127: 
128: In this paper, we therefore study evolutionary food web models and
129: investigate under which conditions these models show a persistent
130: complex structure with several trophic layers. The evolutionary model
131: that we use was studied previously with specific types of population
132: dynamics (Caldarelli {\it et al.}, 1998; Drossel {\it et al.}, 2001). 
133: Here, we explore the effect of changing the functional response used 
134: in the model. We chose a representative selection of
135: functional responses found in the literature and modified them such
136: that they are suitable for many-species communities.  Our aim here is
137: to show that biologically realistic requirements, in particular the
138: capability of predators to adjust their feeding rates and focus on
139: their best prey, are essential to the construction of large stable
140: complex webs. 
141: 
142: \section{The model}
143: 
144: The governing equation for the population $N_{i}(t)$ of species $i$ is taken
145: to have the form
146: \begin{eqnarray}
147: \frac{dN_{i}(t)}{dt} &= &\lambda\sum_{j}N_{i}(t)g_{ij}(t) - 
148: \sum_{j} N_{j}(t)g_{ji}(t) \nonumber \\ && - d_{i}N_{i}(t)
149:  -\sum_j \alpha_{ij}N_i(t)N_j(t)\,,  
150: \label{balance}
151: \end{eqnarray}
152: which naturally accounts for the four processes which lead to a change
153: in the population number $N_{i}(t)$. The function $g_{ij} (t)$ is the
154: functional response, i.e.~the rate at which an individual of species
155: $i$ feeds on species $j$; it depends on the population sizes, and its
156: analytical form will be specified below.  The constant $\lambda$ is
157: the ecological efficiency at which consumed prey are converted into
158: predator offspring.  The first term describes population growth due to
159: food consumption, while the second term describes population decline
160: due to predation. The death rate, $d_{i}$, will be assumed to be equal
161: to 1 for all species. The last term describes direct interference
162: competition between two predators $i$ and $j$. It is needed to
163: represent intra-specific interference competition, which ensures
164: bounded solutions when the functional response is independent of
165: predator density. This term also allows the incorporation of
166: inter-specific interference competition into such models, which
167: facilitates comparison with the predator dependent functional
168: responses.
169: 
170: On a larger time scale, species undergo changes, either by replacement
171: with invading similar species, or by evolutionary change. In order to
172: implement these changes, each species was characterised by a set of 10
173: out of 500 possible features (Caldarelli {\it et al.}, 1998). This
174: representation gives a measure of similarity between species (the
175: number of features they have in common) and allows for ``mutations''
176: or ``invasions'' by randomly replacing one feature of one individual
177: with another. Scores between features were assigned in a random and
178: asymmetric fashion. They are a measure of how useful a feature is for
179: its carrier at feeding on a species carrying another feature.  The
180: interactions, or scores, $a_{ij}$, between species were obtained by
181: adding the mutual scores between all pairs of features carried by the
182: two species.  Positive scores indicate that the first species can feed
183: on the second species and negative scores mean that the first species
184: is consumed by the second. In the latter case the score was set to zero, 
185: since this information was already contained in the (positive) score where 
186: the two species are interchanged. Therefore all scores were such that 
187: $a_{ij} \geq 0$. The external resources were represented as an additional 
188: species of fixed (and large) population size, which does not feed on any 
189: species. The evolutionary dynamics consisted of the following steps: 
190: starting from some initial species configuration (usually the external 
191: resources plus one basal species feeding on it), one individual was 
192: picked at random, and one of its features was changed. Under the dynamics 
193: (\ref{balance}), this new species either died out, or added to the system, 
194: or drove one or several other species to extinction. When the dynamics 
195: had reached an equilibrium, the next ``mutation'' occurred, and the 
196: process was repeated.
197: 
198: In the following, we list the different functional responses used in our 
199: computer simulations. Lotka-Volterra dynamics are recovered by taking 
200: \begin{equation}
201: g_{ij}= a_{ij}N_j\, .
202: \label{LV}
203: \end{equation}
204: 
205: A  Holling type II functional response, which implies saturation of
206: consumption rates at high prey abundance, is given by
207: \begin{equation}
208: g_{ij} = \frac{a_{ij} N_j}{1+\sum_k b_{ik}N_k}\,,
209: \label{generalholling}
210: \end{equation}
211: where the sum in the denominator is taken over all prey $k$ of species
212: $i$. 
213: 
214: More complicated functional responses for multi-species systems can
215: only be found in the recent literature.  Arditi and Michalski (1996)
216: suggest the following generalised Beddington form:
217: \begin{equation}
218: g_{ij} = \frac{a_{ij} N_j}{1+\sum_k b_{ik}N_k+\sum_lc_{il}N_l}\,,
219: \label{generalbeddington}
220: \end{equation}
221: where the first sum is again taken over all prey $k$ of species $i$,
222: and the second sum is taken over all those predator species $l$ that
223: share a prey with $i$. We chose the $c_{il}$ such that individuals
224: belonging to the same species competed more strongly with each other
225: than individuals belonging to different species.
226: 
227: Arditi and Michalski (1996) also suggest the following ratio-dependent
228: functional response, which implements the idea that predators share
229: the prey:
230: \begin{equation}
231: g_{ij} = \frac{a_{ij} N_j^{r(i)}}{N_i+\sum_{k\in R(i)}b_{ik}N_k^{r(i)}}\,,
232: \label{generalratio}
233: \end{equation}
234: with the self-consistent conditions 
235: $$N_j^{r(i)}=\frac{\beta_{ji}N_i^{C(j)}N_j}{\sum_{k\in
236: C(j)}\beta_{jk}N_k^{C(j)}},\quad
237: N_k^{C(j)}=\frac{h_{jk}N_j^{r(k)}N_k}{\sum_{l\in R(k)}h_{lk}N_l^{r(k)}}\,
238: .$$
239: Here $\beta_{ij}$ is the efficiency of  predator $i$ at consuming
240: species $j$, $h_{ij}$ is the relative preference of predator
241: $i$ for prey $j$, $R(i)$ are the prey species for predator $i$,
242: $C(i)$ are the species predating on prey $i$, $N_j^{r(i)}$ is the part
243: of species $j$ that is currently being accessed as resource by species
244: $i$ and $N_k^{C(j)}$ is the part of species $k$ that is currently
245: acting as consumer of species $j$.
246: 
247: Finally, in a recent paper (Drossel {\it et al.}, 2001), we used the 
248: ratio-dependent expression
249: \begin{equation}
250: g_{ij}(t) = \frac{a_{ij}f_{ij}(t)N_j(t)}{bN_j(t)
251: +\sum_k \alpha_{ki}a_{kj}f_{kj}(t)N_k(t)}\,,
252: \label{ourgij}
253: \end{equation}
254: where $f_{ij}$ is the fraction of its effort (or available searching time)
255: that species $i$ puts into preying on species $j$. These efforts are
256: determined self--consistently from the condition
257: \begin{equation}
258: f_{ij}(t) = \frac{g_{ij}(t)}{\sum_k g_{ik}(t)}.\label{eff}
259: \end{equation}
260: This condition is such that no individual can increase its energy
261: intake by putting more effort into a different prey. The sum in 
262: (\ref{ourgij}) is over all species $k$ which are predators of $j$.
263: The competition strength $\alpha_{ik}$ is set equal to one only for 
264: $i=k$, and is smaller than 1 otherwise. For this model, or for any 
265: other model with a functional response which depends on predator 
266: density, the last term in Eq.(\ref{balance}) should be dropped. 
267: 
268: \section{Simulation results: food web structure}
269: 
270: Let us first present results for the simplest type of population
271: dynamics, which is Lotka-Volterra dynamics, Eq.~(\ref{LV}). Starting
272: with one species and the external resources, input at a rate $R$, we 
273: allowed for enough time for the system to evolve towards its characteristic 
274: stationary structure. A typical food web resulting after sufficiently 
275: long time is shown in Fig.~\ref{fig1}. All species feed on the external
276: resources, and additionally on other species.  Only for short
277: transient periods of time  does the web have species on the second
278: trophic level, and virtually never on the third. We also performed
279: simulations that started from a complex web with several trophic
280: layers, which was stable under the population dynamics. However, the
281: evolutionary dynamics caused this structure to collapse, and the
282: resulting structure was again similar to that shown in Fig.~\ref{fig1}. 
283: These results were obtained without inter-specific direct 
284: competition, i.e. with $\alpha_{ij}=c\delta_{ij}$ where $\delta_{ij}$
285: is the Kronecker delta defined by $\delta_{ij} = 0$ for $i \neq j$
286: and $\delta_{ii} = 1$. They are still true when direct inter-specific 
287: competition is included, as in a different evolutionary model described 
288: by L\"assig {\it et al} (2001) and also in our model if we set  
289: $\alpha_{ij} = c\rho_{ij}(1 + q_{ij})/2$ in Eq.~(\ref{balance}). Here 
290: $\rho_{ij}$ is equal to one if $i$ and $j$ share at least one prey and 
291: is zero otherwise, and $q_{ij}$ is the fraction of discrete features 
292: shared by the two species. 
293: 
294: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
295: 
296: \begin{figure}
297: \begin{center}
298: \includegraphics*[width=5cm]{fig1.eps}
299: \caption{Example of a food web resulting from an evolutionary model
300: with Lotka-Volterra population dynamics. The arrow direction indicates
301: the flow of resources, and the arrow thickness is a measure of link
302: strength. Links are only drawn if a species obtains more than 1
303: percent of its food through that link. The radius of the circles
304: increase logarithmically with population size. \label{fig1}}
305: \end{center}
306: \end{figure}
307: 
308: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
309: 
310: Our interpretation of these results is as follows. When a modified
311: species that can feed on more prey than a parent species is introduced
312: into the model, the modified species takes over. This is due to the
313: linear growth rates of Lotka-Volterra systems that allow a species to
314: feed on each of several prey just as easily as it would feed
315: on one prey, if only this prey was present. If allowed to evolve, an
316: artificially constructed Lotka-Volterra food web will therefore tend
317: to a simple web structure where each species feeds on many other
318: species and in particular on the most abundant resource, the external
319: environment. Clearly, this feature of the model is unrealistic as
320: there are evolutionary trade-offs that do not allow species to feed
321: very easily on an arbitrarily large number of species. In particular,
322: species cannot normally feed on external resources and on many other 
323: types of prey. 
324: 
325: In the light of these arguments, a Holling or Beddington type
326: functional response is more realistic, since it naturally limits the
327: total amount of prey taken by a predator. Indeed, the need for 
328: nonlinear growth rates in the context of multispecies communities 
329: has often been discussed in the literature (Pimm, 1991; Rosenzweig, 1995; 
330: Vandemeer {\it et al.}, 2002). Functional responses of the form
331: (\ref{generalholling}) were used by McCann
332: \emph{et al} (1998) in their investigation of the effects of weak
333: links. They have the biologically meaningful feature that predators
334: cannot maintain a high feeding efficiency on many prey at the same
335: time, and they have a stabilising effect on the dynamical equations,
336: at least for subsystems consisting of up to four species (McCann {\it et al.}, 
337: 1998). Additional direct inter-specific competition can be introduced in a
338: natural way by using functional responses of the Beddington form
339: (\ref{generalbeddington}). However, when evolutionary dynamics by
340: modification of existing species are added to either of these models,
341: we find again that complex webs with several trophic layers are not
342: stable and that after some time the webs consist merely of basal
343: species, with a structure very similar to that shown in Fig.~\ref{fig1}. 
344: 
345: It appears that even for the Holling or Beddington type functional
346: response too many species can feed on the same food source, and that
347: the model does not have a mechanism that prevents possible links (i.e. 
348: $a_{ij}>0$) from being realised. In the models we are considering 
349: here, there is always a possible predator-prey interaction between any 
350: two species $i$ and $j$, either $a_{ij} > 0$ and $a_{ji} = 0$ or 
351: $a_{ji} > 0$ and $a_{ij} = 0$. Thus the ecosystems are potentially 
352: highly connected with one half of all inter-specific links realised. 
353: However, species usually focus on those prey to which they are 
354: best adapted. We therefore must introduce a mechanism
355: which would limit the number of species a predator actually
356: feeds on, with potential links becoming active only when the preferred 
357: prey is scarce or unavailable, or after a change in the composition of 
358: the web.  One way to (artificially) realise this feature is to allow 
359: only the best predators to feed on a given prey. We did this by replacing the
360: interactions $a_{ij}$ with adjusted interactions $a_{ij}'=a_j^{\rm
361: max}(1-(a_j^{\rm max}-a_{ij})/\delta)$ (Caldarelli {\it et al.}, 1998), 
362: with $\delta$ being a small parameter and $a_j^{\rm max}$ being the largest
363: interaction against $j$. Negative $a_{ij}'$ are set to zero. Fig.~\ref{fig2} 
364: shows a food web obtained from Lotka-Volterra dynamics with
365: this artificial constraint. Table 1 shows the mean number of species, of 
366: links per species, and the mean occupation numbers of the trophic levels 
367: for this model, for $R = 1\times10^4$, $\lambda = 0.1$, $c = 1.0$ and 
368: $\delta = 0.2$. For comparison, the results for the original Lotka-Volterra 
369: model with the parameters for $R = 1\times10^4$, $\lambda = 0.1$, 
370: and $c = 3.0$ are also shown. Similar web structures were obtained 
371: including direct inter-specific competition or using Holling and 
372: Beddington forms with the same type of adjusted interactions.
373: 
374: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
375: 
376: \begin{table}
377: \begin{center}
378: \begin{tabular}{|c||c|c|c|c|c|c|c|c|} \hline
379:  &  & & \multicolumn{6}{|c|} {Trophic level} \\ 
380: \cline{4-9}
381: Model & Number of & Links per & 1 & 2 & 3 & 4 & 5 & 6\\
382: \cline{4-9}
383: & species & species & \multicolumn{6}{|c|}{Number of species} \\
384: \cline{4-9}
385: & & & \multicolumn{6}{|c|}{Frequency of occupation} \\
386: \hline\hline
387: Lotka-Volterra without & 40.6(2.5) & 2.3(0.1) & 40.5 & 1.6 & --- & --- 
388: & --- & --- \\\cline{4-9}
389: adjusted interactions &           &          & 1.000 & 0.061 
390: & --- & --- & --- & ---  \\ 
391: \hline
392: Lotka-Volterra with & 69.8(20.4) & 1.4(0.1) & 20.9 & 22.1 & 19.1 & 7.5 & 1.4 & 
393: 1.0 \\\cline{4-9}
394: adjusted interactions &            &          & 1.000 & 1.000 & 
395: 1.000 & 0.987 & 0.140 & 0.003 \\ 
396: \hline
397: \hline
398: \end{tabular}
399: \end{center}
400: \caption[]{Food web statistics for the two Lotka-Volterra models: without 
401: and with adjusted interactions. The results are averaged over ten different 
402: simulations (lasting 200000 iterations) and over the last 20000 iterations 
403: of each simulation. The quantities in brackets give standard deviations over 
404: the ten runs for the number of species and links per species. Only links 
405: between non-environment species that constituted greater than 1\% of the 
406: predator's diet were counted in the calculation of the links per species.}
407: \end{table}
408: 
409: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
410: 
411: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
412: 
413: \begin{figure}[t]
414: \begin{center}
415: \includegraphics*[width=6cm]{fig2.eps}
416: \caption{Example of a food web resulting from an evolutionary model with 
417: Lotka-Volterra population dynamics and adjusted interactions ($\delta=0.2$). 
418: The same conventions apply as for Fig.~\ref{fig1} except that vertical 
419: position is now proportional to the average path length from the species 
420: to the environment weighted by diet proportions. Notice that the population
421: sizes vary much less across trophic levels than when a ratio-dependent 
422: functional response is used --- as in Fig.~\ref{fig3}, for instance.
423: \label{fig2}}
424: \end{center}
425: \end{figure}
426: 
427: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
428: 
429: This shows that the ability of predators to concentrate on the prey
430: that they are best suited to exploit, rather than on all possible
431: prey, is essential for the production of realistic food web
432: structures. 
433: 
434: Finally, let us discuss evolutionary food web models with
435: ratio-dependent functional responses.  Ratio-dependent functional
436: responses naturally limit the number of actual prey of a predator by
437: dividing the available prey among its predators. One such
438: ratio-dependent form, given by (\ref{generalratio}), was suggested 
439: by Arditi and Michalski (1996). However, it leads to such
440: strong competition that only one predator species can feed on a given
441: prey species and thus cannot give rise to complex food web
442: structures. (One can conclude this directly from the published webs
443: (Arditi \& Michalski, 1996) or from an analytical calculation with two
444: predator species and one prey species). In reality, however, different
445: species exploit a food source in different ways and therefore
446: competition between individuals of different species is less than
447: between individuals of the same species.
448: 
449: This feature is implemented in the functional response (\ref{ourgij}),
450: which was introduced by some of us (Drossel {\it et al.}, 2001). 
451: Furthermore, equations (\ref{ourgij}) and (\ref{eff}) have
452: several other biologically meaningful features: when prey is very
453: abundant, the consumption rates saturate due to the term containing
454: $b$ in the denominator. Additionally, predators divide their effort
455: (or time) among the available prey in order to obtain the maximum
456: possible amount of food. The condition (\ref{eff}) is an
457: evolutionarily stable strategy and allows species to swiftly adjust to
458: a modified situation, thus endowing the food web model with the
459: flexibility found in real webs. Finally, equations (\ref{ourgij}) and
460: (\ref{eff}) are invariant if identical species are aggregated into a
461: single species. This feature is shared by the other models used in
462: this paper if the last term in equation (\ref{balance}) is either set
463: to zero or present for any pair of species but not if, for instance, 
464: only inter-specific competition is included.
465: 
466: Using the functional response (\ref{ourgij}) in the evolutionary
467: model, we rapidly obtain large and complex food webs with a stable
468: structure, although there is an ongoing turnover of species in the
469: system. The dynamics (\ref{balance}) converge quickly to a fixed
470: point.  The food webs generated by this model share many features with
471: real food webs, such as the fraction of top, intermediate and basal
472: species, and the mean number of links per species.  This is discussed
473: in detail elsewhere (Drossel {\it et al.}, 2001; Quince {\it et al.}, 2004a).  
474: 
475: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
476: 
477: \begin{figure}
478: \begin{center}
479: \includegraphics*[width=4.5cm]{fig3a.eps} 
480: \hskip -5mm \includegraphics*[width=4.5cm]{fig3b.eps}
481: \caption{A model web generated using the functional response (\ref{ourgij}) 
482: before (left) and after (right) deletion of basal species 5. 
483: The two shaded species feeding on it managed to survive by feeding at a 
484: higher rate on other species.
485: \label{fig3}}
486: \end{center}
487: \end{figure}
488: 
489: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
490: 
491: The webs generated by this model often  preserve their structure under 
492: changes in the species composition. Fig.~\ref{fig3} shows a web before and 
493: after a basal species was deleted. One can see that in this case all other 
494: species manage to survive, with the strength of several links being modified,
495: illustrating the flexibility and resilience of the model food webs. Of
496: course, the deletion of a randomly chosen species is not always
497: buffered in this perfect way. A useful measure of the ecosystem stability
498: to deletion is the fraction of species that can be removed without
499: causing further extinctions (Pimm, 1979). This statistic depends on 
500: the model parameters; for the values used to generate 
501: Fig.~\ref{fig3} it is 62\%. For the deletions which do cause further 
502: extinctions the distribution of event sizes decays roughly exponentially with
503: a characteristic size of just a few species (Quince {\it et al.}, 2004b).
504: 
505: \section{Simulation results: link strength distributions}
506: 
507: There has been an increasing realisation that food webs have a
508: large proportion of weak links (Paine, 1992; Tavares-Cromar \&
509: Williams, 1996; Berlow {\it et al.}, 1999; Neutel {\it et al.}, 2002), 
510: and that weak links tend to stabilise population dynamics 
511: (McCann {\it et al.}, 1998).  For this reason, we evaluated the link 
512: strength distribution in our model webs. We found that for all functional 
513: responses that generated large ecosystems either the link strength 
514: distributions were skewed towards zero or a large fraction of links were 
515: zero. This is a highly non-trivial result as in contrast to other work on 
516: the topic of weak links, the link strength distributions in our model 
517: are an emergent property of the system and not put in by hand. It is a 
518: strong indication that weak links are the natural outcome of long-term 
519: ecosystem evolution coupled to population dynamics. 
520: 
521: In the literature there are different definitions of link strength,
522: ranging from the biomass flowing through a link per unit time to the
523: response of one population size to a small change in another
524: population size (Laska \& Wootton, 1998). Here, we evaluated link
525: strength distributions based on the following two definitions: (i) the
526: proportion of the prey in the predator's diet after the population
527: dynamics had equilibrated,
528: \begin{equation} 
529: l_{ij}^{(1)}= \frac{g_{ij}}{\sum_k g_{ik}}\,,
530: \end{equation}
531: and (ii) the per-capita interaction strength $l_{ij}^{(2)}$. These are 
532: sometimes known as the elements of the ``community matrix'' (Laska \& Wootton, 
533: 1998), although this phrase is also used to refer to the unnormalised matrix 
534: of partial derivatives (May 1973). It quantifies the strength of all direct 
535: interactions, both predator-prey interactions and interference competition. 
536: It is defined as:
537: \begin{equation}
538: l_{ij}^{(2)} = \frac{1}{N_{i}}\left(\frac{\partial \dot N_{i}}
539: {\partial N_{j}}\right).
540: \end{equation}
541: Definition (i) typically leads to a U-shaped link-strength distribution, 
542: with a peak at weak links and another one near the maximum link strength 1, 
543: indicating that many predators have one main prey. (Note that the second peak 
544: does not occur if the link strength is defined as the biomass passing through 
545: a link, since in this case the link strength is not normalised to 
546: $\sum_j l_{ij} = 1$). Definition (ii) typically leads to broad distributions 
547: with or without a peak at zero.
548: 
549: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
550:  
551: \begin{figure}
552: \begin{center}
553: \includegraphics*[width=9cm]{fig4.eps}
554: \caption{The distribution of link strengths $l_{ij}^{(1)} > 0$ for the 
555: Lotka-Volterra model without adjusted scores (Fig.~A) with adjusted scores
556: (Fig.~B) and the ratio-dependent functional response (Fig.~C).
557: The $y-$axes on all graphs are scaled logarithmically. The insets show the 
558: data with both axes scaled logarithmically, for Figs.~B and C these include 
559: power-law fits which had exponents -0.74 and -0.29 respectively.
560: \label{fig4}}
561: \end{center}
562: \end{figure}
563: 
564: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
565: 
566: Fig.~\ref{fig4} shows the distribution of link strengths $l_{ij}^{(1)}$
567: for the two Lotka-Volterra models (without and with adjusted scores),
568: and for the model with the ratio-dependent functional response, averaged
569: over many evolutionary time steps and several different model
570: webs. For small link strength, the distribution for the
571: ratio-dependent functional response resembles a power law with an exponent
572: around $-0.74$, while the Beddington model with adjusted scores shows an 
573: exponent around $-0.9$, and the Lotka-Volterra model with adjusted scores 
574: shows an exponent around $-0.29$ (if a power law fit should be attempted at
575: all).  The Lotka-Volterra model without adjusted scores does not have several
576: trophic layers, and its link strength distribution appears to decay
577: exponentially at small values. For the Beddington functional response without 
578: adjusted scores, the decay is much steeper than with adjusted scores, but not 
579: exponential. All the above mentioned models are skewed towards
580: small values except for the Lotka-Volterra model with adjusted scores where
581: the approach to the origin is quite flat. For this model however only a small
582: proportion, ~3\%, of links have $l_{ij}^{(1)} > 0$. We conclude that
583: for the models capable of generating large ecosystems only a small fraction
584: of links were realised or $l_{ij}^{(1)}$ had a large weight at small values.
585: We can summarise this by calculating the fraction of links with 
586: $l_{ij}^{(1)} < 0.01$ including zero, which gives 91\% and 97\% for the 
587: Lotka-Volterra models (without and with adjusted scores) and 97\% for the 
588: ratio-dependent model.
589: 
590: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
591: 
592: \begin{figure}
593: \begin{center}
594: \includegraphics*[width=9cm]{fig5.eps}
595: \caption{The distribution of link strengths $l_{ij}^{(2)} > 0$ for the 
596: Lotka-Volterra models without adjusted scores (Fig.~A) and with 
597: adjusted scores (Fig.~B).
598: \label{fig5}}
599: \end{center}
600: \end{figure}
601: 
602: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
603: 
604: We can investigate these ideas further by examining the link strength 
605: according to the definition (ii); this can be either positive or negative. 
606: In the case of the Lotka-Volterra models, the only direct inter-specific 
607: interactions are between predators and prey with 
608: $l_{ij}^{(2)} = \lambda a_{ij}$ and $l_{ji}^{(2)} = - a_{ij}$ if $i$ 
609: predates $j$. Thus in Fig.~\ref{fig5} we only need to show the positive 
610: half of the $l_{ij}^{(2)}$ distribution. From this we see
611: that, whereas the Lotka-Volterra model without adjusted scores has a 
612: distribution of $l_{ij}^{(2)}$ values that is skewed towards zero,
613: the model with adjusted scores has a maximum at an intermediate value of 
614: $l_{ij}^{(2)}$. However, as was the case for $l_{ij}^{(1)}$, we
615: find a much higher proportion of non-zero links in the model without
616: adjusted interactions (51\% as compared to 3\%).
617: 
618: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
619: 
620: \begin{figure}[t]
621: \begin{center}
622: \includegraphics*[width=9cm]{fig6.eps}
623: \caption{The distribution of non-zero $-\alpha_{ij}$ for the Lotka-Volterra 
624: models with direct inter-specific competition. The results for the model 
625: without adjusted scores is shown in Fig.~A and with adjusted scores in Fig.~B.
626: The fraction of inter-specific interaction with non-zero $\alpha_{ij}$ was 
627: 100\% for the former model and 23\% for the latter. Note that the non-zero
628: $\alpha_{ij}$ are clustered near to the smallest allowed magnitude of 2.5 
629: in Fig.~A and near to the largest allowed magnitude of 5 in Fig.~B.
630: \label{fig6}}
631: \end{center}
632: \end{figure}
633: 
634: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
635: 
636: If we introduce direct inter-specific competition into the Lotka-Volterra 
637: models as in Section 3 by setting $\alpha_{ij} = c\rho_{ij}(1 + q_{ij})/2$ 
638: in Eq.~(\ref{balance}), where $\rho_{ij}$ is equal to one if $i$ and $j$ share 
639: at least one prey and is zero otherwise and where $q_{ij}$ is the fraction of 
640: discrete features shared by the two species, we find that, even if adjusted 
641: interactions are used, only small webs can be evolved. The reason for this 
642: seems to be that the resulting direct competitive interactions which are 
643: allowed by the adjusted interactions are both strong and non-zero for a 
644: large fraction of species pairs in the web. This is illustrated by 
645: Fig.~\ref{fig6} where the distributions of the quantity $-\alpha_{ij}$ 
646: when $c = 5.0$ are plotted for the model with and without adjusted scores. 
647: We plot $-\alpha_{ij}$ because it is the contribution of direct competition 
648: to $l_{ij}^{(2)}$. Note that because $q_{ij}$ is discrete so is $\alpha_{ij}$
649: and that the range of non-zero $\alpha_{ij}$ is 
650: $-5 \leq -\alpha_{ij} \leq -2.5$. 
651: 
652: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
653: 
654: \begin{figure}[t]
655: \begin{center}
656: \includegraphics*[width=9cm]{fig7.eps}
657: \caption{The distribution of $l_{ij}^{(2)}$ for competitive interactions 
658: where the shared prey comprised greater than 1\% of both competitors diet 
659: estimated for the ratio-dependent model. This condition was satisfied by 
660: 5\% of all species pairs.
661: \label{fig7}}
662: \end{center}
663: \end{figure}
664: 
665: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
666: 
667: The model with the ratio-dependent functional response, Eq.~(\ref{ourgij}),
668: also includes direct inter-specific competition, as a result of the implicit
669: sharing of prey between predators. However this does not lead to the strong 
670: competitive interactions seen in Fig.~\ref{fig6}, probably because 
671: this competition occurs through the predator-prey terms, which as was shown
672: in Fig.~\ref{fig4}, are themselves very diffuse. This can be seen in 
673: Fig.~\ref{fig7}, where the distribution of $l_{ij}^{(2)}$ values for all 
674: competitive interactions and where the shared prey comprised greater than 1\% 
675: of both competitors diet, is shown. This distribution, although bimodal, is 
676: heavily weighted towards the origin. This then explains why the ratio-dependent
677: model with inter-specific competition is, in contrast to the Lotka-Volterra 
678: models discussed immediately above, capable of evolving large complex 
679: ecosystems.
680: 
681: This finding of ours is complemented by a recent paper by Kondoh (2003), 
682: who investigates randomly linked food web models and the cascade model based 
683: on Lotka-Volterra dynamics with adjustable foraging efforts. He finds that 
684: incorporating adaptive foragers turns a negative relationship between 
685: stability and complexity, defined in terms of species number and potential 
686: connectance, into a positive one. This result had already been found, albeit 
687: over a more limited range of parameters, by Pelletier (2000). These results 
688: are strictly different to ours in that the stability criterion, community 
689: persistence defined as fraction of species surviving for some long time in 
690: a stochastic environment, is different. However the mechanism that allows 
691: the generation of large webs with high potential connectance, the reduction 
692: in the number of non-zero realised links and the reduction in strength of 
693: those links that are realised due to the foraging dynamics, is very similar 
694: to that proposed above. In fact this mechanism is arguably just May's 
695: original hypothesis, that strong links or a large fraction of links will 
696: destabilise ecosystems, placed in an adaptive context (May, 1972).
697: 
698: \section{Discussion}
699: 
700: The form that a realistic functional response might take has been the subject 
701: of a large number of papers in the literature but few, if any, of these 
702: suggestions were implemented in a model of multispecies communities in order 
703: to test their effectiveness. In this paper we have used an existing model 
704: to investigate the effect that different choices for the functional response 
705: have on food web structure. We studied models with a wide range of functional 
706: responses: Lotka-Volterra with and without direct inter-specific competition,
707: Holling and Beddington forms, all of these but with a mechanism to limit the 
708: number of species a predator actually feeds on, and ratio-dependent 
709: functional responses. We found that in the first two cases complex webs could 
710: not be built up, but in the last two cases they often could. More specifically,
711: unless an (artificial) mechanism was introduced which restricted predator
712: choice to prey which they were best suited to exploit or alternatively a 
713: ratio-dependent functional response was used, then stable webs consisting of
714: more than one trophic level could not be built up.  
715: 
716: Our second major conclusion was that, given we chose a functional response
717: which gave rise to complex, stable webs as described above, the link strength
718: distributions were skewed towards zero. For the Beddington functional response 
719: with adjusted scores or ratio-dependence, the link strength distribution
720: followed a power-law, but this was much less clear in the case of the 
721: Lotka-Volterra model, where only a small fraction of the links were realised
722: i.e. a large fraction of the links were zero. Some insight into the mechanisms 
723: involved in these effects was also gained by looking at the strength of a 
724: direct inter-specific competition introduced into the Lotka-Volterra equations.
725: Without adjusted scores, interactions were present, but not particularly 
726: strong. However with adjusted scores, about three-quarters of the interactions 
727: (for the parameter values we used) were zero, but the remaining quarter were 
728: strong. In this case we found that, even with adjusted scores, large complex
729: webs could not be grown. We concluded that the existence of strong competitive
730: interactions might also destabilise food webs.
731: 
732: In summary, we found that the type of functional response used in the 
733: population dynamics of multispecies communities has to be chosen carefully 
734: if a large complex community is to be sustained. If this is achieved, then 
735: a large proportion of weak links arises naturally from the evolution of the 
736: food webs. These conclusions were arrived at within the context of a class
737: of evolutionary food web models. It would be interesting to investigate if 
738: these results could be obtained from other starting points or with different 
739: model assumptions.
740: 
741: \vspace{0.9cm}
742: 
743: \noindent{\bf Acknowledgements}: We wish to thank Paul Higgs for useful
744: discussions. CQ thanks the EPSRC (UK) for financial support during the 
745: initial stages of this work.
746: 
747: \newpage
748: 
749: \section*{References}
750: 
751: \noindent Arditi, R., Michalski, J., 1996. Nonlinear food web
752: models and their responses to increased basal productivity. In
753: Food webs: Integration of patterns and dynamics. Polis, G.~A.,
754: Winemiller, K.~O. (eds), pp\,122--133. Chapman and Hall, New York.\hfill\break
755: \noindent Berlow, E.~L., Navarrete, S.~A., Briggs, C.~J., Power, M.~E.,
756: Menge, B.~A., 1999. Quantifying variation in the strengths of species
757: interactions. Ecology 80, 2206--2224.\hfill\break
758: \noindent Caldarelli, G., Higgs, P.~G., McKane, A.~J., 1998. Modelling 
759: coevolution in multispecies communities. J. Theor. Biol. 193, 345--358.
760: \hfill\break
761: \noindent Drossel, B., Higgs, P.~G., McKane, A.~J., 2001. The influence 
762: of predator-prey population dynamics on the long-term evolution of food 
763: web structure. J. Theor. Biol. 208, 91--107.\hfill\break
764: \noindent Kondoh, M.~K., 2003. Foraging adaptation and the relationship 
765: between food web complexity and stability. Science 299, 1388--1391.
766: \hfill\break
767: \noindent Laska, M.~S., Wootton, T.~J., 1998. Theoretical concepts and 
768: empirical approaches to measuring interaction strengths. Ecology 79, 
769: 461--476. \hfill\break
770: \noindent L\"assig, M., Bastolla, U., Manrubia, S.~C., Valleriani, A., 2001.
771: Shape of ecological networks. Phys. Rev. Lett. 86, 4418--4421.
772: \hfill\break
773: \noindent Law, R., 1999. Theoretical aspects of community assembly. In:
774: Advanced ecological theory: principles and applications. McGlade, J. (ed).
775: pp\,143--171. Blackwell, Oxford.\hfill\break
776: \noindent May, R.~M., 1972. Will a large complex system be stable?
777: Nature 238, 413--414.\hfill\break
778: May, R.~M., 1973. Stability and complexity in model ecosystems. 
779: Princeton University Press, Princeton. \hfill\break
780: \noindent McCann, K., 2000. The diversity-stability debate. Nature
781: 405, 228--233.\hfill\break
782: \noindent McCann, K, Hastings, A., Huxel, G.~R., 1998. Weak trophic 
783: interaction and the balance of nature. Nature 395, 794--798.
784: \hfill\break
785: \noindent Morton, R.~D., Law, R., 1997. Regional species pools and the 
786: assembly of local ecological communities. J. Theor. Biol. 187, 
787: 321--331. \hfill\break
788: \noindent Neutel, A-M., Heesterbeek, J.~A.~P., de Ruiter, P.~C., 2002.
789: Stability in real food webs: weak links in long loops. Science
790: 296, 1120--1123.\hfill\break
791: \noindent Paine, R.~T., 1992. Food-web analysis through field measurements of 
792: per capita interaction strength. Nature 355, 73--75.
793: \hfill\break
794: \noindent Pelletier, J.~D., 2000. Are large complex ecosystems more unstable? 
795: A theoretical reassessment with predator switching. Math. Biosc. 
796: 163, 91--96. \hfill\break 
797: \noindent Pimm, S.~L., 1979. Complexity and stability: another look at 
798: MacArthur's original hypothesis. Oikos 33, 351--357.
799: \hfill\break
800: \noindent Pimm, S.~L., 1991. The balance of nature. The University of Chicago 
801: Press, Chicago. \hfill\break
802: \noindent Quince, C., Higgs, P.~G., McKane, A.~J., 2004a. Topological 
803: structure and interaction strengths in model food webs. Submitted for 
804: publication to Ecol. Model. \hfill\break
805: \noindent Quince, C., Higgs, P.~G., McKane, A.~J., 2004b. Deleting species 
806: from model food webs. Submitted for publication to Oikos. 
807: \hfill\break
808: \noindent Rosenzweig, M.~L., 1995. Species diversity in space and time. 
809: Cambridge University Press, Cambridge. \hfill\break
810: \noindent Tavares-Cromar, A.~F., Williams, D.~D., 1996. The importance of 
811: temporal resolution in food web analysis: evidence from a detritus-based 
812: stream. Ecol. Monogr. 66, 91--113.\hfill\break
813: \noindent Thompson, J.~N., 1998. Rapid evolution as an ecological process, 
814: TREE 13, 329--332. \hfill\break
815: \noindent Vandemeer, J. \emph{et al}., 2002. Increased competition may promote
816: species coexistence. Proc. Nat. Acad. Sci. 99, 8731--8736. \hfill\break
817: \noindent Yodzis, P., 1981. The stability of real ecosystems. Nature 
818: 289, 674--676.
819: 
820: \end{document}
821: