Pair approximations in spatial biology Tim C.D. Lucas† P air-approximation methods, as used in ecology, epidemiology and evolutionary biology are examined. A number of different moment closures are presented, analysed and compared to individual-based simulations. The closures examined are the unspatial mean-field approximation and then the ordinary pair approximation, a more generic version of this closure and the Kirkwood closure. Parallels to the continuous-space literature and the ways discrete models have been developed and used within the biological literature are reviewed and areas of profitable research are identified. 1. Introduction T he complexity of ecological systems has made them difficult to study with traditional analytical approaches. The arrival of fast, widely accesible computers has ushered in an era of individual-based, spatially-explicit ecological models. However, while the utility of simulations is not in doubt, the results derived from analytically treated models can be more generic and easier to understand. Therefore, the interface between analytical models and simulations is an important area of research. A large proportion of ecological models are formulated under the mean-field assumption; spatial effects are ignored as the population is assumed to be well-mixed. However, it is clear that aggregation is a common property of both animal 1–5 and plant populations. 6–8 Whereas mean-field models assume that individuals are distributed through space randomly, in reality individual dispersal might be reletively short causing individuals to occur in clusters. Therefore individuals experience density higher than they would in a well-mixed population. Whenever this is the case, models should be rooted in a spatial context. Spatial models are, however difficult to analyse. Spatial dynamics in ecology are commonly studied using lattice models. 9 This modelling paradigm treats space as a regular network of spaces. Commonly, networks with a constant degree of four (i.e. square arrays) are used as this is convenient for computing. As these individual based simulations are hard to interpret attempts have been made to create analytical models of these fully specified models. To fully examine the spatial dynamics of populations, one needs to follow population density (ratio of individuals to lattice spaces) as well as pair density (density of pairs of individuals within a certain distance from each other) which describes how aggregated the population is. However, analytical expressions of the density of pairs depend on the density of triplets which in turn depends on higher order interactions. An intuitive (but incomplete) way to picture this hierarchy is to consider concentric circles around a focal individual. Only Supervised by David Murrell and Stephen Baigent. †Email: timecdlucas@gmail.com. 1 Pair approximations Tim Lucas the inner circle directly effects the focal individual, but the state of the inner circle depends on the circle outside it, whose state is in turn dictated by the circle even further out. This hierarchy of dependent moments is known as a cascade. One way of dealing with this problem is to truncate the cascade at a certain level. For example, we can approximate the triplet (and higher) spatial-correlations as a function of pair density 10 which simplifies spatial dynamics without completely removing spatial effects. This is an example of a moment closure and the intended result is a closed system of equations that approximate higher order implementations. This method is known as pair approximation. 10,11 However, moment closure is not an exact science and there are many ways to close a system of equations. 12 Many closures are good purely because of their heuristic properties and therefore searching for closures by considering intuitive approximations is not always effective. Whatever the closure used 12 the equations should still satisfy a number of conditions; as the spatial scale of interactions approaches infinity, the equations should reduce to a non-spatial system for example. Being an approximation method, all solutions will fail at some point. The major trade off between different moment closures is between simplicity and accuracy. Another consideration is whether accuracy at steady state is more important than accuracy during the transition towards the steady state. I study three different moment closures of a spatial logistic model on a square lattice. The first closure is the ordinary pair approximation (OPA). 10,13 This closure turns out to be a specific example of a more general closure which has been considered in continuous-space 14 but not discrete space. This is the second closure I examine; I refer to this closure as the extended pair approximation (EPA). Finally, the Kirkwood closure (KPA), 15 which has been rarely used in the discrete literature (but see Filipe et al. 16 ) is considered. This report examines some of the details of these closures, specifically identifying when they are good and when they break down. The paper is organised as follows. First, the ways in which pair approximation models have been used is reviewed (Section 2). Secondly, a lattice IBM (Section 3.1) and a pair approximation anologue with different closures (Section 3.2) are described. Finally a discussion (Section 4) is included. 2. Literature Review Pair approximation has found applications in many areas of ecology, disease biology and evolutionary biology. Any process in which interactions occur on a scale smaller than the entire environment should be considered in a spatial context. As mentioned before, while simulations are often the only way to incorporate all facets of a system in a model, pair approximation can provide analytically tractable models which are often easier to interpret and generalise. Ordinary pair approximation is the most commonly used closure. It rests on the assumption that the effects of neighbour-of-neighbours is week. Therefore the expected local density of an individual next to a full-empty pair is assumed to be the same as an individual next to a full lattice space. Many extensions to the ordinary pair approximation model have been examined. A purely heuristic solution to the innaccuracies of the ordinary pair approximation, known as the improved pair approximation, is to estimate from lattice simulations a correction parameter. 13,17 However, this approach removes some of the advantages of pair approximation. If the point of pair approximation is to avoid timely simulations, then a closure that requires a simulation to learn from is not very useful. More importantly, if we assess pair approximation models by comparing them to simulations, how do we assess the improved pair approximation? One common extension to the ordinary pair approximation is to blur the boundary between mean-field and spatial models by allowing a process to occur both on the local scale and on a global scale. 11,18–20 Steady state results can become very complicated. 11 This model can also be used to study evolutionary stability of long vs. short dispersers. 20,21 The examination of other or generalized network topologies is an ongoing research topic 22,23 while using the information from topological assumptions can also be used to improve moment closures. 22,24–26 Commonly, 2 Pair approximations Tim Lucas A) B) = C) * * * Figure 1. A) Pairs between an individual (red) and any of its pairs (blue) are considered equal and contribute to ρ11 equally. B) Rotational and reflective symmetry. All reflected and rotated examples of a configuration between a focal (red) individual, its neighbour (blue) and neighbour-but-one (light blue) are equal. C) All possible triplet configurations. In the closures used here, all configurations are considered equal even though starred triplets are within the red individual’s neighbourhood while the unstarred individuals are not. Other closures use the information from this geometry to make closures more accurate, but they are not considered in detail here. all configurations of triplets are considered equal — on a square lattice, a triplet in a line is considered the same as an L-shaped triplet. However, individuals of an L-shaped triplet are all neighbours of each other (in an eight-tile neighbourhood) while the two end members of a linear triplet are not in each others neighbourhood (see Figure 1). An obvious extension to the pair approximation model is to truncate the spatial terms at triplet densities instead of pair densities thus making a triplet-approximation model. 23,25–27 While this increases accuracy, the expressions become unweildy, although the results can sometimes be usefully simple, even if their derivations are difficult. The application of pair approximation to a number of biological topics are reviewed below. Although modelling different systems, the models are often very similar. For example, examining the population growth of a forest with suitable and unsuitable habitats, is completely analogous to the spread of a disease through a population of resistant and non-resistant individuals. 2.1. Ecology. Possibly the most obvious use of pair approximations and the area in which it was first applied is spatial ecology. While many populations can be considered well-mixed, species where movement is slow enough and dispersal short enough will have spatial structure. Pair approximation can add a spatial element to non-spatial models as in the case of the distinction between mean-field and pair approximations. It is also used to examine explicitely spatial questions — such as forest gap size — that lose meaning without a spatial model. As the focus of this paper is to examine the validity of different moment closures, I use a simple logistic equation. This model was first proposed by Matsuda et al. 10 This most simple pair approximation model predicts population densities much lower than mean-field models as individuals constrained in space suffer density dependent death more accutely than individuals in a well mixed population. Pair approximation models of population growth, as affected by spatial considerations, have been improved and extended in a number of ways. One common way to extend the model given here is to allow a proportion of offspring to be dispersed globally 3 Pair approximations Tim Lucas instead of locally. 11,19,20 Another approach is to allow density-dependent death and dispersal to act on different scales 28 (the death rates in this report are considered density independent, which relates to a density dependent length scale of zero.) In parallel to the development of pair approximations in discrete spaces, continuous space models have also been studied. Both the fully defined simulations and analytical approximations are significantly more complex in continuous space. 14,29 Interesting dynamics have been found in continuous space logistic models such as extinction of populations with reletively high birth rates and steady-state population sizes much higher than mean field predictions due to uniform dispersal. 14 The examination of closures in continuous space seems to be more in-depth. More closures are analysed, and the relative advantages of each are examined more carefully. 12 The allee effect is any form of negative density dependence that occurs at low population sizes. Many non-spatial models have been proposed to model the allee effect. 30 Three main classes of mechanism can be considered: demographic stochasticity and mate finding, inbreeding and cooperative behaviour. 31 All of these are affected by local as well as global density and are therefore suitable for study with pair approximation models. Stewart-Cox et al. explicitely include allee effects into pair approximation models and consider densities of empty, unpollinated and pollinated spaces and allow pollination to be global or local. 32 If either pollination or seed dispersal is global, the model becomes equivelent to mean-field models. However, if there is a local element to pollination or dispersal, allee effects are weaker or absent due to the local clumping of individuals. Satō reaches similar conclusions 33 by studying a meta-population, in which each space in the lattice is a small, transient sub-population rather than an individual. Heterogeneous environments have been modelled by considering densities and pairs of full, empty and unsuitable sites 20,34,35 which in the simplest case only requires three state variables. 34,35 These models yield results useful for conservation. A reduced proportion of habitable space reduces the steady state population size, as does reducing environemental correlation i.e. many small patches can support a smaller population than a few large patches. However, odd results can occur from the simplest models; Hiebeler found that only the structure of unsuitable habitat, and not the proportion, changed the population density. 35 2.2. Disease Biology. There is a large literature on epidemiology of infectious diseases that uses lattice models. Usually, the models are based around the class of SEIRS compartmental models that were developed under mean-field assumptions. However, when examined with pair approximation the large number of equations usually restricts analyses to either SI(S) models and SIR(S) models. Thus, each node in the lattice is either empty or a susceptible, infected or resistant host. The first epidemiological pair approximation models came soon after the first ecological models. 13 These first models obtained results quantitatively different from mean-field models, such as pathogen driven extinction in large areas of parameter space. While the authors did not make strong claims that their model reflected reality, it was clear that spatially modelled disease dynamics were very different to mean-field dynamics and that the spatial dimension of epidemiology needed to be studied. STDs have attracted a lot of attention due partially to the rise of high-profile diseases such as HIV, but also due to the clear cut nature of network edges; edges join individuals in a sexual relationship. The topology of lattices has been an important area of research with dynamic lattices 23,36 and heterogenous lattices 23 recieving particular attention. Intuitive results are often found; short, frequent sexual relationships and concurrent relationships both increase the size of an epidemic. 23,36,37 3. Methods 3.1. Discrete-space model. I simulated continuous-time spatial ecology on a discrete-space s × s lattice (see Table 1 for a list of symbols) where each space is either empty (0) or contains an individual (1). Periodic boundaries prevent edge effects. For comparison with pair approximations 4 Pair approximations Tim Lucas q1 1 Ρ1 0.3 0.2 0.2 0.1 0.1 40 80 Time 120 40 (a) Density Time 120 (b) Local Density q1 1 Ρ1 æ æ æææ ææææ æ ææææ ææ ææ ææ 0.3 0.3 ææ æ æ æ ææææ æ 0.2 ææææ æ æ ææ æ æ æ æ æ æ æ ææææææææææææææææ 1.2 1.4 1.6 0.2 æ æ 0.1 0.1 1.0 80 b 1.8 æææææææææææææææ 1.0 (c) Density 1.2 1.4 1.6 b 1.8 (d) Local Density Figure 2. A–B) Mean trajectories through time from 15 simulations with b = 0.9, 1.1, 1.3, 1.5 and 1.7. A) Density ρ1 . B) Local density q1|1 . Other parameters as z = 8, s = 20. q1|1 becomes very variable when ρ1 is small. A birth N = 1 → N = 2 causes q1|1 = 0 → q1|1 = 1/z C–D) Steady state global (C) and local (D) densities for varying birth rates (b). For b < 1.3 the population goes to extinction so ρ1 , q1|1 = 0. the density of individuals is the ratio of population size (N ) and available space so that ρ1 = N/s2 . Each iteration of the simulation entails either a birth or a death. If an individual dies the site it occupies changes from 1 → 0. If an individual creates an offspring, one of its neighbouring sites changes from 0 → 1. If the chosen site is already full the birth is aborted, thus creating density dependance. I use square neighbourhoods with odd number radii. The size of the neighbourhood is denoted z (z = 8 unless z is the only paremeter being varied). The number of full sites in an individual’s neighbourhood is zi1 . As the neighbourhood increases, the spatial structure of the system breaks down and density is explained by a mean-field birth death process. To speed up simulations, instead of computing aborted births, the likelihood that an individual will create an offspring is weighted by the proportion of free sites in its neighbourhood. Thus, an individual and event is randomly selected in one go by choosing from a weighted vector of all birth and death events. Deaths are scaled to a weighting of 1 making a unit of time equal to 1 generation. The birth weighting wi of an individual is wi = bzi1 /z and the probability of a birth or death during an iteration is wi i=1 (wi + 1) 1 P (i dies) = PN i=1 (wi + 1) P (i gives birth) = PN (1) (2) To make time continuous I used a Gillespie algorithm. 38 Unlike discrete time simulations in which a variable number of events happen with a fixed-length time step, in the Gillespie algorithm, 5 Pair approximations Tim Lucas Table 1. List of parameter and variable names. Symbol Description ρi Frequency of sites of type i qi|j Conditional probability of a type-i neighbour given a type-j site ρij Frequency of type i - type j pairs t Time b Birth rate d Death rate (scaled to 1) z Number of sites in neighbourhood j zi Number of state-j sites in neighbourhood of individual i N Absolute population size τ Time-step between events. wi Event weighting s Length of one side of the lattice Type Model State variable Both State Variable Both Variable Variable Parameter Parameter Parameter Variable PA Both Both Both Both IBM Variable Variable Variable Parameter IBM IBM IBM IBM one event occurs in each time step, but the length of the time step is variable. The length of each time step is drawn from an exponential distribution scaled by the summed rate of all possible events. Thus, in each iteration of the model, one random event is selected. I then draw a time-step from a probability distribution based on all the possible events. The two possible events are birth and death and the summed ‘any-event’ rate is given by Equation 3. The time between events, τ , is by Equation 5. Note that d = 1 due to rescaling and has no coefficient as it is density independent. E= N X zi1 b + d) (3) R ∼ U(0, 1) 1 τ = − Log(R) E (4) ( i=1 z (5) Time-series from this IBM are shown in Figure 2. It can be seen that only when the birth rate is above about 1.5 does the population avoid extinction. As the population becomes small, q1|1 becomes very variable. In the limit of a change from a population of 1 to a population of a 11 pair, q1|1 changes from 0 to 1/z. This variability cannot be captured by any of the moment closures. 3.2. Pair approximation model. We want to examine the density of individuals ρ1 and the density of pairs ρ11 . However, the final equations are easier in terms of conditional probabilities, q1|1 . This is the probability that a space is full, given that its neighbour is full and is analogous to the mean of zi1 for all i. A number of identities connect these quantites. This model is constucted under the assumption of rotational symmetry (see Figure 1) where a 01 pair and a 10 pair are equivelent in all contexts; they contribute equally to ρ01 and a 101 triplet and a 110 triplet would contribute equally to q1|01 . Equation 6 explicitely states this assumption. Equations 7 – 10 are true by definition. 6 Pair approximations Tim Lucas ρ01 ≡ ρ10 (6) ρ1 ≡ 1 − ρ0 (7) ρ11 ≡ ρ1 q1|1 (8) ρ10 ≡ ρ1 q0|1 ≡ ρ0 q1|0 (9) q1|1 ≡ 1 − q0|1 (10) We can write a simple birth-death process of the proportion of lattice-space filled, ρ1 with death rate d and birth rate b (Equation 11). Dividing by d rescales time so that 1 unit of time is 1 generation and reduces the number of parameters to one. In the IBM, density dependance is included by aborting births that land on a full tile. Births therefore occur at a rate of bzi0 ρ1 . Thus in the analytical model, births occur at a rate bq0|1 ρ1 (Equation 12) as q0|1 is analogous to the population mean of zi0 . dρ1 = −dρ1 + bρ1 dt dρ1 = −ρ1 + bq0|1 ρ1 dt (11) (12) As ρ1 is dependent on local density q0|1 , we must construct an equation for the second spatial moment. Even though we want q1|1 as our state variable, we write an equation for ρ1 1 as it is more intuitive. dρ11 b b = −2ρ11 + 2 ρ10 + 2 (z − 1) q1|01 ρ10 dt z z (13) where z is neighbourhood size as in the IBM. The first term describes the loss of pairs. Pairs of full spaces are lost when either of the pairs die i.e. 11 → 10 or 11 → 01; hence the coefficient of 2. New pairs are created either by an individual creating an offspring next to itself (10 → 11 or 01 → 11) or by a neighbour-but-one individual creating an offspring next to the focal individual (101 → 111). These pair-birth rates are described by the second and third term in Equation 13 respectively. This last term includes the third spatial moment q1|01 which is, in turn, dependent on higher order moments. This is a cascade of moments as discussed above and means the system is not closed and cannot be solved. Thus an approximation is needed to close the system; this is the role of the closures discussed in Sections 3.3 – 3.6. We want to examine the state variables ρ1 and q1|1 (equations for q1|1 turn out to be simpler than for ρ11 ). As the density of pairs is equal to the product of singlets and local pairs (from ρ11 = ρ1 q1|1 ) we can write an expression for the dq1|1 /dt as in Equation 14. dq1|1 d(ρ11 /ρ1 ) ρ11 dρ1 1 dρ11 = =− 2 + dt dt ρ1 dt ρ1 dt (14) For each closure, we find an approximate expression for q1|01 which is substituted into Equation 13. Equations 12 and 13 are then substitued into Equation 14 to find an equation for dq1|01 /dt in terms of only state variables and parameters. This and Equation scaled are then a closed system of ordinary differential equations. 7 0.00 0.15 1 q1 0.15 0.00 ρ1 0.30 Tim Lucas 0.30 Pair approximations 2 4 6 8 2 Radius (r) 4 6 8 Radius (r) Figure 3. Simulation results (mean after 30 events burn-in) for population density and spatial second moment for different radii with mean and 95% confidence intervals shown. Dashed red lines show the predictions from mean field equations. Mean field models the system well at r = 6. Parameters used: b = 1.1, s = 20, z = 8. 3.3. Mean field approximation. In a mixed population, local density is equal to global density. Therefore, the mean field assumption can be explicitely stated as q1|1 ≈ ρ1 (15) Thus Equation 12 becomes a standard logistic birth-death process with density dependent births dρ1 = −ρ1 + bρ1 (1 − ρ1 ) dt (16) As this equation does not depend on any second-order terms we can neglect equations for second-order dynamics. Figure 3 shows that this model succesfully predicts population density for IBMs run with sufficiently large nieghbourhood as such large dispersal breaks all spatial correlation. 3.4. Ordinary pair approximation closure. The ordinary pair approximation assumption is that the indirect effect of neighbours-of-neighbours is small and therefore their density be considered as equal to the density of direct neighbours. Thus q1|01 ≈ q1|0 = (1 − q1|1 )ρ1 ρ10 = ρ0 1 − ρ1 (17) Identities 7 – 10 can be used to find an exppression for this approximation in terms of triplet densities as shown below. Either form (but rarely both) are used in the literature. While Equation 17 is more intuitive, Equation 18 leads more easily to the other approximations used and so is useful to include. ρijk ≈ ρij ρjk ρj Substituting Equation 17 into Equation 13 yields the approximation 8 (18) Pair approximations Tim Lucas dρ11 ρ10 ρ1 b b = −2ρ11 + 2 ρ10 + 2 (z − 1)ρ10 dt z z 1 − ρ1 (19) Substituting Equations 12 and 19, along with identities 7 – 10 gives us a closed set of differential equations in terms of only the state variables ρ1 and q1|1 and the parameters z and b. dρ1 = − ρ1 + b(1 − q1|1 )ρ1 dt h i dq1|1 = − q1|1 −1 + b(1 − q1|1 ) dt " # ) ( 1 1 (1 − q1|1 )ρ1 + 2 −q1|1 + b (1 − q1|1 ) + 1− z z 1 − ρ1 (20) (21) These equations give the phase plots in Figure 4. The steady states are given by ρ∗1 = ρ∗1 b(z − 1) − z 1 ∗ , q1|1 =1− b(z − 1) − 1 b ∗ q1|1 = 0, = 2b + z(1 + b) ± (22) q −8b2 z − [2b + z(1 + b)]2 2bz (23) If we set ρ∗1 = 0 in Equation 22 and solve for b, we see that this positive steady state becomes zero when b = z/z − 1. For values of b lower than this, there is only the steady state in Equation 23. This steady state is charecterised by extinction (ρ1 = 0) but positive q1|1 . This is a failing in the closure as, from Equation 8, q1|1 = 0 when ρ1 = 0. This relates to the fact that in an empty lattice, an individual will have no individuals in its neighbourhood. To the best of my knowledge, this fact is not mentioned in any of the many papers that use this closure. 10,20,34,35,39 The phase plots in Figure 4 show that the positive steady state in (Equation 22) is globally stable when it exists. Thus whenever birth rate is sufficiently larger than death rate, the population should persist. 3.5. Extended pair approximation. The ordinary pair approximation is just one of many possible closures. It is in fact a special form of a more general class of closures that have been used in the continuous literature. 14 ρijk 1 ≈ α+β+γ ρij ρjk ρjk ρki ρki ρij α +β +γ ρj ρk ρi ! (24) When i = j = k (i.e. ρ111 or ρ000 ) or β = γ = 0 and α > 0 this closure equals Equation 18. In this analysis I use ijk = 101 and α = β = γ = 1. However, studies of continuous systems have heuristically found β = γ = 1, α = 4 to be a good form of this closure. 14 As these parameters simply define the proportional influence of the difference focal individual in each triplet, setting one parameter to 1 can always reduce the number of parameters. Getting this into an approximate value for q1|01 and only in terms of the two state variables and constants, gives q1|01 ≈ ρ1 + 2q1|1 − 3ρ1 q1|1 3(1 − ρ1 ) Substituting this into Equations 13 and 14 and rearranging gives 9 (25) Tim Lucas 1.0 1.0 0.8 0.8 0.8 0.6 0.6 0.6 0.4 0.4 0.4 0.2 0.2 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 (a) OPA, b = 1 (b) EPA, b = 1 (c) KPA, b = 1 1.0 1.0 0.8 0.8 0.8 0.6 0.6 0.6 0.4 q1 1 1.0 q1 1 q1 1 q1 1 1.0 q1 1 q1 1 Pair approximations 0.4 0.4 0.2 0.2 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Ρ1 (d) OPA, b = 1.6 (e) EPA, b = 1.6 (f) KPA, b = 1.6 Figure 4. Stream plots for z = 8 and ABC) b = 1, DEF) b = 1.6 using the three closures. Steady states are marked with a black dot. With b = 1, all three populations yield a population that goes to extinction. For b = 1.6 there is positive steady states in OPA and EPA but not the kirkwood closure. dρ1 = − ρ1 + b(1 − q1|1 )ρ1 dt dq1|1 1 1 2bρ1 (1 − q1|1 ) = − (ρ1 + b(1 − q1|1 ))q1|1 − dt ρ1 ρ1 z 2b(z − 1)ρ1 (1 − q1|1 )(ρ1 + 2q1|1 − 3ρ1 q1|1 ) − 2ρ1 q1|1 + 3z(1 − ρ1 ) (26) (27) This system of ODEs yields the steady states ρ∗1 ∗ q1|1 p 2 + b − 2z + 2bz ± 4 + 4b(z − 1)2 + 10z − 5z 2 = 6 − 3z + b(z + 2) b−1 = b (28) Unlike the other two closures, this system has only two steady states. When b < 1 the steady states are complex, making this a rather unuseful closure. Furthermore, at b > 1 the steady state ∗ is greater than zero. As mentioned before, this is an inconsistency. However, at local density q1|1 b = 1 this system has a steady state at ρ1 = 0, q1|1 = 0. This is the only system to have a steady state at this, true extinction, point. The phase diagrams for this system are shown in Figure 4. 10 Pair approximations Tim Lucas 3.6. Kirkwood closure. A rarely used closure in discrete space (Filipe et al. 16 apply the closure to discrete space epidemiology models while Diekmann and Law 12 use the closure in continuous space) that has been found to have useful properties are Kirkwood style closures. 15 While the ordinary pair approximation is known as a power-2 closure as it contains pair densities squared, the Kirkwood closure is a power-3 closure. It takes the form ρij ρjk ρki ρi ρj ρk (29) q1|1 q1|1 = 1 − ρ1 ρ0 (30) ρijk ≈ which yields q1|01 ≈ This can be seen as a corrected form of the ordinary pair approximation. When population density is high, the probability of have a neighbour-but-one is higher, thus in this closure q1|01 increases as ρ1 increases. Again we substitute this into Equations 13 and 14 and rearrange to find an equation for the state variables. h i 3z(ρ1 − 1)q1|1 + b(q1|1 − 1) z(ρ1 − 3)q1|1 + 2(ρ1 + q1|1 − 1) dq1|1 = dt z(ρ1 − 1) (31) This system of ODEs yields the steady states b−1 1 ∗ , q1|1 = b + z − bz b √ √ 3bz − 3z − 4b − 3 8bz + 3z 2 − 6z 2 b + 3b2 z 2 ∗ ∗ ρ1 = 0, q1|1 = 2(6 + 2b − 3z + bz) ρ∗1 = (32) (33) These steady states are never positive. Therefore this system never has a positive population density and cannot describe the lattice model which has stable populations for b > ∼1.3. 4. Discussion Figure 5 shows the trajectories of all three moment closures as well as simulated time series from the IBM. As the Kirkwood closure always predicts population extinction, the population density when b = 1.6 is qualitatively incorrect. The extended pair approximation grossly overestimates population density at positive population values (Pane B) and underestimates the speed at which extinction is reached (Pane A). No closures accurately predicts local density during extinction (Pane C), although the extended pair approximation is probably closest. The ordinary pair approximation is the best closure, although it overestimates population density significantly at stable positive populatiaon densities. As discussed, the extinction stable state is also qualitatively incorrect in its prediction of steady state local density. This can be seen in the positive value of this in Pane C. Previous work in the continuous literature has suggested that the extended pair approximation with different values for α, β and γ might give more accurate approximations. 14 This approximation should be studied in more depth in discrete space. It is internally consistant, unlike the Kirkwood close, but has more flexibility than the ordinary pair approximation. As mentioned before, 1 of the 3 parameters defining the relative influence of the three individuals in each triplet (α, β, and γ) can be set to zero as the closure as a whole is normalised by the 1/(α + β + γ) 11 Pair approximations Tim Lucas Ρ1 Ρ1 0.4 0.4 0.2 0.2 0 10 20 30 Time 0 (a) Density. b = 1 20 30 Time (b) Density. b = 1.6 q1 1 q1 1 0.4 0.4 0.2 0.2 0 10 10 20 30 (c) Local Density. b = 1 Time 0 10 20 30 Time (d) Local Density. b = 1.6 Figure 5. Plot showing numerical solutions to ODE’s for OPA (blue), EPA (red) and KPA (purple) as well as an average of 20 simulated time series (black). b = 1.3, z = 8, s = 20. term. However, this still leaves a large amount of parameter space to search. If these parameters are set by training them against simulations, we arrive at the same problem of validation as discussed for the improved pair approximation. Therefore, a number of different parameter sets should be tried, but training should be avoided. Although using the geometry of a lattice can make models more accurate when compared with IBMs on that specific lattice, they suffer from a lack of generality. When accurately describing a specific simulation is the aim, the motivation for these closures is clear. However when the aim is to describe generic lattice simulations, continuous-space simulations or real world situations, these models may reduce accuracy rather than increase it while incurring the cost of increased complexity. Similarly, the improved pair approximation is ‘improved’ only in the sense that it closer approximates a specific lattice simulation. This heuristic correction is likely to be more robust to applications on other geometries and continuous space, but whether it applies to real life situations is debatable. The most glaring gap in the literature on pair approximation models is the lack of empirical tests of the models. However, some examples occur in the epidemiology literature. For example, Xiao et al. formulated a meta-population model of Salmonella transmission in cattle. 39 In their model each node can be a susceptible, infected or resistant herd (not individual) of cattle or a clean or infected environment (a novel way to model free-living bacteria.) The model yields interesting oscillations as well as useful indications into efficient intervention strategies. Filipe et al. statistically fitted a pair approximation model to time-series data collected from experimental radish-Rhizoctonia solani systems. 40 They recover biologically plausible parameter estimates from the time-series and also obtain qualitatively correct dynamical behaviour. Interestingly, the parameter estimates with largest standard error are those for primary infection rates. It 12 Pair approximations Tim Lucas has been noted that ordinary pair approximation is least accurate at very low densites 25,41 such as when a new disease is entering a population. Other experiments in mesocosms of benthic invertebrates, plants or microbial species would likely be reletively easy to perform, if potentially time consuming. It is worth remembering that pair approximation is only one approach to studying spatial systems. For example, Snyder and Nisbet 41 present a model based on the observation that spatial structure (ρ10 /ρ1 (1 − ρ1 )) increases almost linearly with an increase in population density. They use this to construct a model that is simpler and more accurate than pair approximations. Their model is less flexible than pair approximation as the relationship that is the foundation of the model falls apart if dispersal is based on a more complex kernal than simple nearest neighbour interactions. A mix of nearest neighbour and global dispersal can however be included. While pair approximation is a very useful tool, especially at equilibrium, other approaches should continue to be studied. Similarly, while discrete space, square lattice pair approximation makes comparison to simulations easy, it is not obvious how results from these models will compare with continuous space or empirical results. These models are generally formulated in highly regular spatial arenas, whose geometry is decided based on computational, and not biological, considerations. Only comparisons to empirical datasets can truly validate these models. 4.1. Conclusions. I have presented and analysed a suite of moment closures for a pair approximation of the spatial logistic model. I have not found any closures that improve on the ordinary pair approximation. However, I have highlighted some failings with this closure that have not been commented on. Further study, especially on the extended pair approximation closure and in empirical tests of these models, is advised. References 1. Dennis, P., Young, M., and Gordon, I. Distribution and abundance of small insects and arachnids in relation to structural heterogeneity of grazed, indigenous grasslands. Ecological Entomology 23(3), 253–264 (1998). 2. Adler, P. and Lauenroth, W. Livestock exclusion increases the spatial heterogeneity of vegetation in colorado shortgrass steppe. Applied Vegetation Science 3(2), 213–222 (2000). 3. Andrew, N. Spatial heterogeneity, sea urchin grazing, and habitat structure on reefs in temperate Australia. Ecology 74, 292–302 (1993). 4. Peltonen, M., Liebhold, A., Bjørnstad, O., and Williams, D. Spatial synchrony in forest insect outbreaks: roles of regional stochasticity and dispersal. Ecology 83(11), 3120–3129 (2002). 5. Thomson, J., Weiblen, G., Thomson, B., Alfaro, S., and Legendre, P. Untangling multiple factors in spatial distributions: lilies, gophers, and rocks. Ecology 77(6), 1698–1715 (1996). 6. Nicotra, A., Chazdon, R., and Iriarte, S. Spatial heterogeneity of light and woody seedling regeneration in tropical wet forests. Ecology 80(6), 1908–1926 (1999). 7. Hubbell, S. Tree dispersion, abundance, and diversity in a tropical dry forest. Science 203(4387), 1299 (1979). 8. Law, R., Herben, T., and Dieckmann, U. Non-manipulative estimates of competition coefficients in a montane grassland community. Journal of Ecology 85, 505–517 (1997). 9. Durrett, R. and Levin, S. Stochastic spatial models: a user’s guide to ecological applications. Proc. R. Soc. B. 343(1305), 329–350 (1994). 10. Matsuda, H., Ogita, N., Sasaki, A., and Satō, K. Statistical mechanics of population. Prog. Theor. Phys. 88(6), 1035–1049 (1992). 11. Harada, Y. and Iwasa, Y. Lattice population dynamics for plants with dispersing seeds and vegetative propagation. Researches on Population Ecology 36(2), 237–249 (1994). 13 Pair approximations Tim Lucas 12. Diekmann, U. and Law, R. Relaxation projections and the method of moments. In The geometry of ecological interactions: simplifying spatial complexity, Dieckmann, U., Law, R., and Metz, J., editors, 288–411. Cambridge Univ. Press, (2000). 13. Satō, K., Matsuda, H., and Sasaki, A. Pathogen invasion and host extinction in lattice structured populations. Journal of Mathematical Biology 32(3), 251–268 (1994). 14. Law, R., Murrell, D., and Dieckmann, U. Population growth in space and time: spatial logistic equations. Ecology 84(1), 252–262 (2003). 15. Kirkwood, J. Statistical mechanics of fluid mixtures. The Journal of Chemical Physics 3, 300 (1935). 16. Filipe, J., Maule, M., and Gilligan, C. On ‘analytical models for the patchy spread of plant disease’. Bulletin of Mathematical Biology 66(5), 1027–1037 (2004). 17. Harada, Y., Ezoe, H., Iwasa, Y., Matsuda, H., and Satō, K. Population persistence and spatially limited social interaction. Theoretical Population Biology 48(1), 65–91 (1995). 18. Boots, M. and Sasaki, A. ‘small worlds’ and the evolution of virulence: infection occurs locally and at a distance. Proc. R. Soc. B. 266(1432), 1933–1938 (1999). 19. Iwasa, Y. Lattice models and pair approximation in ecology. In The geometry of ecological interactions: simplifying spatial complexity, Dieckmann, U., Law, R., and Metz, J., editors, 227–251. Cambridge Univ. Press, (2000). 20. Hiebeler, D. Competing populations on fragmented landscapes with spatially structured heterogeneities: improved landscape generation and mixed dispersal strategies. Journal of Mathematical Biology 54(3), 337–356 (2007). 21. Harada, Y. Short- vs. long-range disperser: the evolutionarily stable allocation in a latticestructured habitat. Journal of Theoretical Biology 201(3), 171–187 (1999). 22. Petermann, T. and De Los Rios, P. Cluster approximations for epidemic processes: a systematic description of correlations beyond the pair level. Journal of Theoretical Biology 229(1), 1–11 (2004). 23. Bauch, C. A versatile ODE approximation to a network model for the spread of sexually transmitted diseases. Journal of Mathematical Biology 45(5), 375–395 (2002). 24. Van Baalen, M. Pair approximations for different spatial geometries. In The geometry of ecological interactions: simplifying spatial complexity, Dieckmann, U., Law, R., and Metz, J., editors, 359–387. Cambridge Univ. Press, (2000). 25. Bauch, C. The spread of infectious diseases in spatially structured populations: An invasory pair approximation. Mathematical Biosciences 198(2), 217–237 (2005). 26. House, T., Davies, G., Danon, L., and Keeling, M. A motif-based approach to network epidemics. Bulletin of Mathematical Biology 71(7), 1693–1706 (2009). 27. Hiebeler, D. and Millett, N. Pair and triplet approximation of a spatial lattice population model with multiscale dispersal using markov chains for estimating spatial autocorrelation. Journal of Theoretical Biology 279 (2011). 28. Ellner, S. Pair approximation for lattice models with multiple interaction scales. Journal of Theoretical Biology 210(4), 435–447 (2001). 29. Bolker, BM amd Pacala, S. and Levin, S. Moment methods for ecological processes in continuous space. In The geometry of ecological interactions: simplifying spatial complexity, Dieckmann, U., Law, R., and Metz, J., editors, 288–411. Cambridge Univ. Press, (2000). 30. Boukal, D. and Berec, L. Single-species models of the allee effect: extinction boundaries, sex ratios and mate encounters. Journal of Theoretical Biology 218(3), 375–394 (2002). 31. Courchamp, F., Clutton-Brock, T., and Grenfell, B. Inverse density dependence and the allee effect. Trends in Ecology & Evolution 14(10), 405–410 (1999). 32. Stewart-Cox, J., Brittona, N., and Mogie, M. Pollen limitation or mate search need not induce an allee effect. Bulletin of Mathematical Biology 67(5), 1049–1079 (2005). 33. Satō, K. Allee threshold and extinction threshold for spatially explicit metapopulation dynamics with allee effects. Population Ecology 51(3), 411–418 (2009). 14 Pair approximations Tim Lucas 34. Ovaskainen, O., Satō, K., Bascompte, J., and Hanski, I. Metapopulation models for extinction threshold in spatially correlated landscapes. Journal of Theoretical Biology 215(1), 95–108 (2002). 35. Hiebeler, D. Populations on fragmented landscapes with spatially structured heterogeneities: landscape generation and local dispersal. Ecology 81(6), 1629–1641 (2000). 36. Kim, J. and Koopman, J. Hiv transmissions by stage in dynamic sexual partnerships. Journal of Theoretical Biology 298, 147–153 (2012). 37. Bauch, C. and Rand, D. A moment closure model for sexually transmitted disease transmission through a concurrent partnership network. Proc. R. Soc. B. 267(1456), 2019–2027 (2000). 38. Gillespie, D. Exact stochastic simulation of coupled chemical reactions. The journal of Physical Chemistry 81(25), 2340–2361 (1977). 39. Xiao, Y., French, N., Bowers, R., and Clancy, D. Pair approximations and the inclusion of indirect transmission: Theory and application to between farm transmission of Salmonella. Journal of Theoretical Biology 244(3), 532–540 (2007). 40. Filipe, J., Otten, W., Gibson, G., and Gilligan, C. Inferring the dynamics of a spatial epidemic from time-series data. Bulletin of Mathematical Biology 66(2), 373–391 (2004). 41. Snyder, R. and Nisbet, R. Spatial structure and fluctuations in the contact process and related models. Bulletin of Mathematical Biology 62(5), 959–975 (2000). 15