Main

Understanding the mechanisms that maintain high species diversity in plant communities such as tropical forests has long challenged ecologists1 and has stimulated major efforts in field and theoretical ecology2,3,4,5. However, despite a multitude of coexistence mechanisms that have been proposed6 and recent advances in coexistence theories7,8,9,10,11,12,13,14,15,16,17,18, this fundamental question has not been fully resolved8,9,14. For example, theoretical models indicate that stable coexistence is difficult to reach in large communities11,13. We argue that consideration of spatial patterns of plant individuals, such as intraspecific clustering and interspecific segregation, may allow for a better understanding of mechanisms of coexistence in species-rich communities17.

Although many studies suggest that spatial patterns and neighbourhood effects may play an important role in diversity maintenance17,18,19,20,21, the integration of spatial patterns into coexistence theories of species-rich communities is difficult. A major difficulty is linking spatial processes at the individual level to community dynamics. One reason for this is a scale mismatch. The analytical models that form the basis of most coexistence theories7,8,11,12,13,14,15,16,22 have state variables that operate at the macroscale (that is, the population or community-level abundances), use parameters that describe average ‘mesoscale’ properties of the individuals (such as population-level interaction coefficients and demographic rates) and often rely on ‘mean-field’ approximations18,23 where spatial patterns are neglected. However, spatial patterns and population-level interaction coefficients emerge at the mesoscale from the microscale behaviour of individuals and their interactions with other individuals and the environment. Therefore, studying the impact of spatial patterns on species coexistence requires multiscale approaches such as spatial moment equations18,23,24 that incorporate pattern-forming processes operating at the level of individuals and translate these into population and community dynamics.

We propose here such a multiscale approach. To this end, we first derive population-level interaction coefficients αfi from individual-level interaction coefficients βfi and neighbourhood crowding indices19,21 that are commonly used to describe interactions among tree individuals at the microscale, and then incorporate the emerging coefficients αfi into analytical macroscale models. Our approach is based on separation of timescales (adiabatic approximation25), given that mesoscale spatial patterns usually build up quickly and approach a quasi-steady state whereas the macroscale state variables (for example, abundances) change slowly23. Therefore, we do not need to describe the dynamics of the quick mesoscale patterns explicitly (as, for example, is done in approaches based on moment equations18,23,24) but concentrate instead on spatial patterns that transport the critical information from the microscale into macroscale models. This approach requires information on mesoscale spatial patterns that can be obtained from fully mapped forest plots such as those of the Forest Global Earth Observatory (ForestGEO) network4.

More specifically, we (1) derive species-level interaction coefficients from individual-level interactions using empirical information on spatial patterns in nine ForestGEO megaplots4, (2) integrate the resulting species-level interaction coefficients into analytical macroscale multispecies models and (3) study their consequences for multispecies dynamics and coexistence.

Results and discussion

Species-level interaction coefficients

We first derive species-level interaction coefficients from individual-level neighbourhood crowding indices19,21,26 that quantify how the performance of a focal individual depends on interactions with its neighbours. To this end, we describe the survival rate of a focal individual k of species f in dependence on the local number of neighbours as

$$s_{kf} = s_f \exp \left( - \beta _{ff} \left(n_{kff} + \underbrace {\mathop {\sum}\nolimits_{i \ne f} {(\beta _{fi}/\beta _{ff})n_{kfi}} }_{n_{kf\beta }} \right) \right),$$
(1a)

where sf is a density-independent background survival rate of species f; the crowding indices nkff, nkfi and nkfh are the number of conspecifics, neighbours of species i and heterospecific neighbours within distance R of a focal individual k, respectively; the subscript ‘h’ indicates all heterospecifics together (that is, nkfh = ∑i≠f nkfi) and the crowding index nkfβ weights each heterospecific neighbour by its relative competition strength βfi/βff (equation (1a)), with βfi being the individual-level interaction coefficients between species f and i (Fig. 1a–c). The corresponding population-level survival rate is given by

$${\mathrm{surv}}_f = s_f\exp \left( - \alpha _{ff} \left(N_f(t) + \mathop {\sum}\limits_{i \ne f} {(a_{fi}/ \alpha _{ff})} N_i(t) \right) \right),$$
(1b)

where Ni(t) is the abundance of species i at time step t and αfi is the population-level interaction coefficient between species f and i. To estimate survf we average the survival rates skf (equation (1a)) of all individuals k of species f:

$${\mathrm{surv}}_f = \frac{1}{{N_f\left( t \right)}}\mathop {\sum}\limits_{k = 1}^{N_f\left( t \right)} {s_{kf} = s_f} \mathop {\int}\limits_x {\mathop {\int}\limits_y {\exp } } \left( { - \beta _{ff}\left( {x + y} \right)} \right)p_x p_y {\mathrm{d}}x {\mathrm{d}}y$$
(2)

where px and py are the distributions of the crowding indices x = nkff and y = nkfβ for individuals of species f, respectively.

Fig. 1: Neighbourhood crowding indices describe individual-level interactions and their intraspecific variability.
figure 1

a, The conspecific crowding index nkff is the number of conspecific neighbours (filled red circles) within distance R (black circle) of the focal individual k (filled red square). b, The heterospecific crowding index nkfh is the number of heterospecific neighbours (filled grey circles) within distance R of the focal individual k. c, The heterospecific interaction crowding index nkfβ additionally weights heterospecifics by their relative competitive effect βfi/βff, symbolized by the arrows. Different colours indicate different species. d, Distribution of the number nkff of conspecific neighbours with diameter at breast height (dbh) ≥ 10 cm of the species Castanopsis cuspidata of the 25 ha Fushan plot. e, Corresponding distribution of heterospecific neighbours nkfh. f, Corresponding distribution of the crowding index nkfβ. In d–f, blue lines show gamma distributions with the same means and variance-to-mean ratios as the observed distributions, and the vertical black line indicates the mean value. See Supplementary Data Table1 for additional examples.

Full size image

To determine the distributions px and py, we analysed forest inventory data from nine 20–50 ha forest dynamics plots (Supplementary Table 1) in the ForestGEO network4. We used phylogenetic similarity between tree species as a surrogate for the relative competition strength βfi/βff because it is available for the species in the nine plots (Methods). This is an established approach in species-rich communities19,26,27 to approximate niche differences in the absence of other data.

The number of con- and heterospecific neighbours and the heterospecific interaction index nkfβ vary widely among conspecifics and can be described by gamma distributions (Fig. 1 and Extended Data Figs. 1 and 2). Detailed analysis of the empirical crowding indices reveals additional relationships that are relevant for our subsequent analysis. First, we find that the crowding indices nkff and nkfβ are not, or are only weakly, correlated for a given species f (Extended Data Fig. 3a). Second, we find for trees of a given species f high correlations between the two crowding indices nkfh and nkfβ (Extended Data Fig. 3b) with a common factor Bf (that is, nkfβ ≈ Bfnkfh). This result suggests operation of diffuse neighbourhood competition, in which the competition strength of heterospecifics is on average a factor Bf lower than that of conspecifics.

The integral of equation (2) can be solved analytically for independent gamma distributions px and py and yields

$${\mathrm{surv}}_f = s_f \exp \left( { - \beta _{ff}\left( {\gamma _{ff}\bar n_{ff} + \gamma _{f\beta }\bar n_{f\beta }} \right)} \right)$$
(3)

where \(\bar n_{ff}\) and \(\bar n_{f\beta }\) are the average values of the crowding indices nkff and nkfβ, respectively, and γff and γfβ contain the variance-to-mean ratios of the gamma distributions px and py, respectively, but in our case have values close to one (Methods).

The last step in deriving pairwise population-level interaction coefficients is to relate the averages of the different crowding indices to the macroscale population abundances Nf(t). We accomplish this by taking advantage of connections between crowding indices and the summary functions of spatial point process theory21. The mean of the crowding index nkff (that is, the mean number of further conspecific neighbours within distance R) is proportional to Ripley’s K, a well-known quantity in point process theory28,29:

$$\bar n_{ff} = K_{ff}\left( R \right)N_f(t)/A$$
(4a)

where Kff(R) is the univariate K function for species f and A is the area of the observation window. The K function describes the spatial pattern of conspecifics within a neighbourhood distance R, indicating clustering if Kff(R) > πR2, a random pattern if Kff(R) = πR2 and regularity if Kff(R) < πR2. In the following, we use the normalized K function \(k_{ff}\left( R \right) = K_{ff}\left( R \right)/\uppi R^2\) to quantify the spatial neighbourhood patterns, and therefore \(\bar n_{ff} = k_{ff}\left( R \right)\frac{{\uppi R^2}}{A}N_f(t)\).

Analogously, the mean number of heterospecific neighbours is given by

$$\bar n_{f{\mathrm{h}}} = k_{f{\mathrm{h}}}\left( R \right)\frac{{\uppi R^2}}{A}\mathop {\sum }\limits_{i \ne f} N_i(t),$$
(4b)

where the bivariate normalized K function kfh(R) indicates segregation to heterospecifics (subscript ‘h’) within distance R if kfh(R) < 1. Independent placement occurs if kfh(R) = 1, and attraction if kfh(R) > 1.

Motivated by the finding nkfß ≈ Bf nkfh (Extended Data Fig. 3b) we rewrite the mean crowding index \(\bar n_{f\beta }\) as

$$\bar n_{f\beta } = \underbrace {B_f}_{\frac{{\bar n_{f\beta }}}{{\bar n_{f{\mathrm{h}}}}}}\bar n_{f{\mathrm{h}}},$$
(5)

where the point process summary function Bf indicates how much the competition strength of one heterospecific neighbour differs on average from that of one conspecific neighbour. The values of Bf depend mainly on the individual-level interaction coefficients βfi but also on the spatial pattern of the different species and their relative abundances (equation (12)).

Inserting the expressions for \(\bar n_{ff}\) and \(\bar n_{f\beta }\) into equation (3) and comparing with equation (1b) leads to our first main result, the analytical expressions of the population-level interaction coefficients:

$$\alpha _{ff} = c\,\gamma _{ff}\,k_{ff}\beta _{ff} \qquad {\mathrm{intraspecific}}\,{\mathrm{interactions}}\,{\mathrm{of}}\,{\mathrm{species}}\,f$$
(6a)
$$\alpha _{fi} = c\gamma _{f\beta }k_{f{\mathrm{h}}}\beta _{ff}B_f \, \quad {\mathrm{interspecific}}\,{\mathrm{interactions}}\,{\mathrm{with}}\,{\mathrm{species}}\,i$$
(6b)

with scaling constant c = πR2/A. Notably, equation (6b) indicates that the emerging population-level interaction coefficients αfi are the same for all heterospecifics (that is, αfi = αfh for i ≠ f). Thus, even if the individual-level interactions coefficients βfi differ among species pairs, the emerging population-level interaction coefficients αfi have a substantially simpler structure. This phenomenon is an example of simplicity emerging from complex species interactions30 and is likely to occur only in species-rich communities9,12,31.

The population-level interaction coefficients αfi depend on several factors that can influence the macroscale balance between intra- and interspecific competition: (1) intraspecific clustering kff and interspecific segregation kfh (Fig. 2a,b), (2) the relative competition strength Bf of one heterospecific neighbour (Fig. 2c) and (3) the shape of the response of survival to crowding (γff and γfβ, which contain the variance-to-mean ratios of the distribution of the crowding indices). Note that absence of spatial patterns (that is, kff = 1 and kfh = 1) and a linear approximation of equation (3) lead to γff = γfβ = 1 and direct proportionality αfi = cβfi of the individual- and population-level interaction coefficients, as assumed by Lotka–Volterra models.

Fig. 2: Emergent spatial patterns in the nine forest dynamics plots.
figure 2

a–c, Distribution of the values of the different measures of spatial patterns, taken over the focal species of the different forest plots. Boxplots show the 10th, 25th, 50th, 75th and 90th percentiles; outliers are indicated by filled black points. Intraspecific clustering is indicated by kff > 1 and interspecific segregation by kfh < 1, and the less a heterospecific neighbour competes on average relative to a conspecific neighbour, the more Bf decreases. The neighbourhood radius used was R = 10 m. For the analysis, we used all individuals with dbh ≥ 10 cm and included focal species with more than 50 individuals. For forest plot names, see Supplementary Data Table 1.

Full size image

The information on spatial patterns extracted from the inventory data of our nine forests allows us to estimate the relative population-level interaction coefficients αfi/αff for all pairs of species i and f (Fig. 3 and Extended Data Fig. 4). For example, at BCI, the values of αfi/αff differ substantially from the corresponding individual-level coefficients βfi/βff (Fig. 3a,c), and for 83% of all species pairs, we find αfi/αff < βfi/βff. Thus, the mesoscale spatial patterns can reduce, at the population level, the strength of heterospecific interactions relative to conspecific interactions by ‘diluting’ encounters with heterospecific neighbours relative to conspecific neighbours. Spatial patterns therefore have a strong potential to alter the outcome of deterministic individual-level interactions.

Fig. 3: The emergent population-level interaction coefficients.
figure 3

a, Relationship between the relative population-level interaction coefficients αfi/αff and the corresponding relative individual-level interaction coefficients βfi/βff, (equation (6)) for the 75 focal species of the BCI plot. The βfi/βff were based on phylogenetic dissimilarity. b, Distribution of the values of αfi/αff, taken over the 289 focal species of the different forest plot. Boxplots show the 10th, 25th, 50th, 75th and 90th percentiles; outliers are indicated by filled black points. For full, separate distributions for tropical, subtropical and temperate forests, see Extended Data Fig. 4. c, Example for the distribution of the relative individual- and population-level interaction coefficients for the BCI plot. The neighbourhood radius used was R = 10 m. For the analysis, we used all individuals with dbh ≥ 10 cm and included focal species with more than 50 individuals. For forest plot names, see Supplementary Data Table 1.

Full size image

Conditions for coexistence in the multiscale model

To study the consequences of the emerging spatial patterns for community dynamics and coexistence we insert the population-level interaction coefficients αfi (equation (6)) into a simple macroscale model

$$\begin{array}{l}\frac{{N_f\left( {t + {\Delta}t} \right) - N_f\left( t \right)}}{{{\Delta}t}} = N_f\left( t \right)\\\left[ {\left( {r_f - 1} \right) + s_f \exp \left( - \alpha _{ff}N_f\left( t \right) - \alpha _{f{\mathrm{h}}}\mathop {\sum }\limits_{i \ne f} N_i(t)\right)} \right]\end{array}$$
(7)

In this model we assume that survival is governed by neighbourhood competition with αfi = αfh, and the number of recruits of species f during a time step Δt is given by rfNf(t), where rf is the per capita reproduction rate of species f.

The carrying capacity of species f (that is, the equilibrium of equation (7) with Ni(t) = 0 for all i ≠ f) is given by \(K_f = - \ln \left( {\frac{{1 - r_f}}{{s_f}}} \right) {\alpha _{ff}}^{-1}\). Note that our theory also applies, after redefinition of the carrying capacity, to alternative macroscale models (Supplementary Table 3). From equation (7) we find \(K_f = N_f^{\ast} + \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}\mathop {\sum }\limits_{i \ne f} N_i^{\ast} = N_f^{\ast} \left( {1 - \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}} \right) + \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}J^{\ast}\), where \(N_f^{\ast}\) is the abundance of species f in equilibrium and J* the equilibrium community size (that is, \(J^{\ast} = \sum _i N_i^{\ast}\); see also equation (14)). This leads, under the assumption that the population-level interaction coefficients αfh are constant (Supplementary text), to a single equilibrium of the macroscale model for species f

$$N_f^{\ast} = \frac{{\overbrace {\alpha _{ff}K_f}^{ - \ln \left( {\frac{{1 - r_f}}{{s_f}}} \right)} - \alpha _{f{\mathrm{h}}}J^{\ast}}}{{\alpha _{ff} - \alpha _{f{\mathrm{h}}}}}$$
(8)

that is positive if denominator and numerator are both positive or both negative. However, the invasion criterion (equation (18)) is only fulfilled if both are positive. In this case, equation (8) suggests two different ways a species can go extinct. First, the denominator indicates that a species with strong clustering kff will show a small equilibrium abundance since in this case αff ≫ αfh (equation (6)). Large values of kff can be expected for species of low abundance under dispersal limitation, where recruitment happens close to conspecific adults.

Second, the numerator of equation (8) indicates a positive abundance of species f if αffKf > αfhJ* and αfh/αff < 1. Therefore, we introduce a new feasibility index

$$\mu _f = \frac{{\alpha _{f{\mathrm{h}}}J^{\ast}}}{{\alpha _{ff}K_f}}$$
(9)

that indicates a positive abundance if µf < 1 given that heterospecific interactions at the population level are weaker than conspecific interactions (that is, αfh/αff < 1). The invasion criterion7,8 that tests whether a species with low abundance can invade the equilibrium community of all other species turned out to be basically the same as the feasibility condition (equation (9)) if the invading species does not show strong clustering (Methods and equation (18)). Note that we did not assume Allee effects8.

Further analysis that considers the dependency of J* on the values of Kf and αfh/αff shows that the values of µf must be similar for all species f to fulfil the condition µf < 1, and that µf can show larger interspecific variability if the species richness S is smaller and/or if the mean of \(\frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}} \left( {1 - \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}} \right)^{-1}\) is smaller (equations (14) and (15)). The feasibility index µf therefore governs species assembly by determining the subset of species of a larger species pool that can persist13, but any addition of a new species changes µf and may lead to reassembly of the community.

Using the observed abundances in the forest plots (and assuming equilibrium) allows us to test our theory. We can estimate from the observed abundances the carrying capacities Kf and therefore also the indices µf for all focal species of our nine plots (Fig. 4). For 282 of our 289 focal species, we found µf < 1 and αfh/αff < 1, which means that the two conditions for stable coexistence are indeed satisfied for nearly all species. However, this is not a given, as shown by the seven species from BCI with µf > 1 and αfh/αff > 1. Thus, our theory is compatible with the observed coexistence of most species at our nine forest plots if the assumption of approximately constant population-level interaction coefficients holds.

Fig. 4: The feasibility index µf for the 289 focal species of different forest plots.
figure 4

a, The distribution of the index µf (equation (9)) for the 289 analysed tree species. The vertical dashed line indicates the median of the distribution. b, Distribution of µf, taken over the 289 analysed species of the different forest plot. Boxplots show the 10th, 25th, 50th, 75th and 90th percentiles; outliers are represented by black points. c, Relationship between the feasibility index µf and the ratio αfh/αff of heterospecific to conspecific population-level interaction coefficients. Values of µf < 1 indicate a positive abundance. The red lines show the dependence of µf on αfh/αff expected for communities without interspecific variability in µf (equation (10a)) with 18 focal species (CBS plot, lower line) and 75 focal species (BCI plot, upper line). The neighbourhood radius used was R = 10 m. For the analysis, we used all individuals with dbh ≥ 10 cm and included tree species with more than 50 individuals. For plot names, see Supplementary Data Table 1.

Full size image

In addition, we get information on the typical values of µf and αfh/αff that allows for insight into the stability of the communities. In agreement with the predictions of our theory, we find that µf tends to be smaller if αfh/αff is smaller (Fig. 4c). Furthermore, the values of µf were, for most species, larger than the expectation of µf for the corresponding communities without interspecific variability in µf (equation (10a)) but with the same number of species and the same mean values of αfh/αff (Fig. 4c), but all of the values were relatively close to the critical value of 1 (the median of all 289 species was 0.938; Fig. 4a).

Consequences of spatial patterns for coexistence

Our theory predicts that coexistence requires, in the limit of high species richness, that species approach functional identity with respect to the feasibility index µf (Methods). This resembles neutral theory2,32, but our theory allows for trade-offs among demographic parameters and emerging spatial patterns to reach this equivalence (equation (9)). To study the consequences of spatial patterns for coexistence, we analysed a symmetric33 version of our model where all species have the same parameters and follow the same stochastic rules and where all individuals compete identically (leading to Bf = 1). Thus, we eliminate any potential coexistence mechanism other than that resulting from spatial patterns.

If the mesoscale patterns kff and kfh converge to a stochastic equilibrium, we find that the feasibility and invasion criteria are always fulfilled if αfh/αff < 1 since

$$\mu _f = \frac{{S\frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}}}{{1 + \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}\left( {S - 1} \right)}} < 1$$
(10a)
$$\mu _f^i = \frac{{(S - 1)\frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}}}{{1 + \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}\left( {S - 2} \right)}} < 1$$
(10b)

Equations (10a) and (10b) follow from equations (16) and (18), respectively, if αfh/αff is the same for all species f. Thus, the spatial patterns that emerge at the mesoscale from the individual-level interactions can stabilize if αfh/αff < 1. The underlying mechanism is a positive fitness–density covariance34 (Methods).

To reveal the conditions that can lead to coexistence in a spatially explicit context, we use a spatially explicit and individual-based35 implementation of the symmetric version of the multiscale model (equation (7) and Methods). Models of this type are able to produce realistic spatial patterns consistent with mapped species distributions of large forest plots17,35. While our analytical approach in equation (7) only allows us to make simplified assumptions about the spatial component of the recruitment process, the simulation model allows us to explore the role of the spatial component of recruitment in more detail.

Indeed, the way recruits were placed was critical for coexistence. Randomly placed recruits produced unstable dynamics (Extended Data Fig. 5a) characterized by regularity (mean of kff = 0.92) and segregation (mean of kfh = 0.92), both caused by competition18, and the instability was caused by con- and heterospecifics competing equally at the population level (that is, αfh/αff ≈ 1) (Extended Data Fig. 5d,g). When we followed the common approach of placing recruits with a kernel around conspecific adults17,18,35,36,37,38,39 to mimic dispersal limitation, we again found unstable dynamics (Extended Data Fig. 5b), despite intraspecific clustering and interspecific segregation (that is, αfh/αff < 1; Extended Data Fig. 5n). The reason for the instability was high clustering of rare species24,35 (Extended Data Fig. 6) that completely negated the potentially positive effects of αfh/αff < 1.

In contrast, community dynamics can be stabilized if recruits are placed in small clusters but independent of the location of conspecific adults (Extended Data Fig. 5c). With this mechanism we mimic canopy gaps40, animal seed dispersal41 or other mechanisms that can generate clustering independent of parent locations, as found at BCI42. Decoupling clustering from the parent locations does not lead to the negative relationship between clustering and abundance, and all measures of spatial patterns converged quickly into quasi-equilibrium (Extended Data Fig. 5f,i,o).

The simulation data reveal the spatial coexistence mechanism underlying the positive fitness–density covariance34 (Extended Data Figs. 7 and 8). We find that the emerging spatial patterns lead to a situation where individuals of a common species are more likely to be near more neighbours and tend therefore to experience stronger competition (Extended Data Fig. 7). While the number of heterospecific neighbours remains approximately constant, the number of conspecific neighbours decreases with decreasing abundance if clustering does not change with abundance (equation (4a)). However, if clustering increases with decreasing abundance, the rare-species advantage is weakened and the dynamics become unstable24.

The data of several ForestGEO forest plots were compatible with a positive fitness–density covariance (Extended Data Fig. 8f–i) as they show that, when a species becomes rare, areas of higher conspecific crowding tend to have fewer total competitors. Comparison of the results of the stable versus unstable simulation showed that even relatively weak tendencies in this relationship are sufficient to stabilize the dynamics (Extended Data Figs. 8a,b). However, this was not the case for three temperate forest plots where the power-law clustering–abundance relationship showed exponents of b < −0.5 (that is, species tended to have high clustering at low abundances), but the other plots showed b > −0.5 (Extended Data Fig. 8n–p).

The apparent contradiction with previous theoretical studies18,24,35,37 where intraspecific clustering and interspecific segregation could not stabilize community dynamics thus arises as a consequence of the assumption of placing recruits close to their parents. This finding has important consequences for ecological theory because it shows, in contrast to the prevalent view36,43,44, that spatial patterns alone can lead to coexistence of multiple species. This is even more important since the specific spatial patterns required for this coexistence mechanism also exist in real forests.

Conclusions

Understanding the mechanisms that maintain high species diversity in communities such as tropical forests is at the core of ecological theory, but these mechanisms are not yet fully resolved. Here, we argue that spatial patterns may play an important role in species coexistence of high diversity plant communities17. To test this hypothesis, we introduced a multiscale framework that reveals how pattern-forming processes operating at the level of individuals translate into mesoscale spatial patterns and how those patterns influence macroscale community dynamics.

We showed that the population-level interaction coefficients αfi can have, for a broad range of common circumstances, a simpler structure than the underlying individual-level interaction coefficients βfi. This simplicity, which emerged from spatially explicit species interactions30, allowed for an analytical treatment of equilibrium, feasibility and invasion conditions of the corresponding macroscale models (equations (8–10)). Inserting the emerging αfi coefficients into macroscale community models (for example, equation (7); Supplementary Table 3) should, in principle, allow us to take advantage of macroscale theory9,11,12,13,45. However, our results also indicate that the population-level interaction coefficients may not be temporally constant as commonly assumed but depend on spatial patterns that may change with abundance. This is especially likely if recruitment is mainly located close to the parents.

It is also possible to expand our framework to take into account more detailed neighbourhood crowding indices that consider not only the number of neighboured trees of a given species but also their distance and size19,21,26. This requires redefinition of the quantities kff and kfh that describe intraspecific clustering and interspecific segregation, respectively, but does not change the overall structure of our equations. A special strength of our approach is that the population-level interaction coefficients contain measures of spatial neighbourhood patterns that can be directly estimated from fully mapped forest plots4. Together with additional information, this may allow for estimating network structures as well as stability of the whole community.

Our analysis revealed that communities of competing species can show a stable mode where the mesoscale patterns converge quickly into quasi-equilibrium and an unstable mode where negative relationships between species clustering and abundance emerge (Extended Data Figs. 5 and 6). The two modes are governed by the way species clustering is generated: the well-known unstable mode is related to clustering of recruits around their parents17,18,35,36,37,38,39 whereas the stable mode is related to clustering in locations that are independent from the parent locations, due, for example, to animal seed dispersal41 or canopy gaps40. This result calls for a closer examination of the spatial relationship between the recruits and adults. Indeed, independent placement of recruits from conspecific large trees may not be unusual. For example, Getzin et al.42 found in detailed analyses of the BCI forest that recruits were for most species spatially independent of large conspecific trees. For the stable mode we could identify conditions for coexistence, and forthcoming work may extend to quantifying the ability of additional mechanisms such as niche differences7,8, habitat associations46, spatial and temporal relative nonlinearity7,8 and storage effects7,8 to alleviate the destabilizing increase of clustering if species become rare.

This study explicitly incorporates spatial patterns in theoretical models of plant communities and combines analytical theory with spatial simulations and field data analysis. Our finding that species with similar attributes may show stable coexistence has profound implications for ecological theory. Furthermore, the multiscale framework we propose here opens exciting new avenues to explain species coexistence through a spatial lens.

Methods

Study areas

Nine large forest dynamics plots of areas between 20 and 50 ha were used in the present study (Supplementary Table 1). The forest plots are part of the ForestGEO network4 and are situated in Asia and the Americas at locations ranging in latitude from 9.15° N to 45.55° N. Tree species richness among the plots ranges from 36 to 468. All free-standing individuals with diameter at breast height (dbh) ≥1 cm were mapped, size measured and identified. We focused our analysis here on individuals with dbh ≥ 10 cm (resulting in a sample size of 131,582 individuals) and focal species with more than 50 individuals (resulting in 289 species). The 10 cm size threshold excludes most of the saplings and enables comparisons with previous spatial analyses20,35,47,48. Shrub species were also excluded.

Some of our analyses require estimation of the ratio βfi/βff that describes the relative individual-level competitive effect18 of individuals of species i on an individual of the focal species f. We used for this purpose phylogenetic distances49 based on molecular data, given in Myr, that assume that functional traits are phylogenetically conserved19,26,27. In this case, close relatives are predicted to compete more strongly or to share more pests than distant relatives26. To obtain consistent measures among forest plots, phylogenetic similarities were scaled between 0 and 1, with conspecifics set to 1, and a similarity of 0 was assumed for a phylogenetic distance of 1,200 Myr, which was somewhat larger than the maximal observed distance (1,059 Myr). This was necessary to avoid discounting crowding effects from the most distantly related neighbours26.

Observed spatial patterns at species-rich forests

Figure 1 and Supplementary Data Table 1 show the intraspecific variation in our three crowding indices nkff, nkfh and nkfβ that can be approximated by gamma distributions. To assess how well the gamma distribution described the observed distribution, we used an error index defined as the sum of the absolute differences of the two cumulative distributions divided by the number of bins (spanning the two distributions). The maximal value of the error index is one, and a smaller value indicates a better fit.

Equations (6, 8 and 9) relate the measures of the emerging spatial patterns (that is, kff, kfh and Bf) to macroscale properties and conditions for species coexistence. Even though our multiscale model (equation (7)) is simplified, it allows for a direct comparison with the emerging patterns in our nine fully stem-mapped forest plots. We estimate the key quantities of equations (8) and (9) directly from the forest plot data (Fig. 4), with the exception of the carrying capacities Kf, which were indirectly estimated from the observed species abundances (assuming approximate equilibrium; equation (8) and Supplementary Data Table 1). This allowed us to estimate the feasibility index µf (equation (9)). Because statistical analyses with individual-based neighbourhood models19,26 based on neighbourhood crowding indices have shown that the performance of trees depends on their neighbours for R between 10 and 15 m, we estimate all measures of spatial neighbourhood patterns with a neighbourhood radius of R = 10 m. Analyses with R = 15 or R = 20 gave similar results.

The spatial multispecies model and equilibrium

We use a general macroscale model to describe the dynamics of a community of S species:

$$\frac{{N_f\left( {t + {\Delta}t} \right) - N_f\left( t \right)}}{{{\Delta}t}} = N_f\left( t \right)\left[ {\left( {r_f - 1} \right) + s_f\exp \left( { - \alpha _{ff}N_f\left( t \right) - \mathop {\sum }\limits_{i \ne f} \alpha _{fi}N_i(t)} \right)} \right]$$
(11)

where rf is the mean number of recruits per adult of species f within time step Δt, sf is a density-independent background survival rate of species f and the αfi are the population-level interaction coefficients, yielding αff = c γff kff βff and αfi = c γfβ kfh βff Bf (equation (6)). The βfi are the assumed individual-level interaction coefficients between individuals of species i and f; kff = Kff(R) / π R2 and kfh = Kfh(R) / π R2 measure intraspecific clustering and interspecific segregation, respectively, with Kff(R) being the univariate K function for species f and Kfh(R) the bivariate K function describing the pattern of all heterospecifics ‘h’ around individuals of species f. A is the area of the observation window.

Following equation (5), Bf can be estimated as

$$B_f = \frac{{\bar n_{f\beta }}}{{\bar n_{f{\mathrm{h}}}}} = \frac{{\mathop {\sum }\nolimits_{i \ne f} \left[ {ck_{fi}N_i\left( t \right)} \right]\frac{{\beta _{fi}}}{{\beta _{ff}}}}}{{\mathop {\sum }\nolimits_{i \ne f} \left[ {ck_{fi}N_i\left( t \right)} \right]}} = \frac{{\mathop {\sum }\nolimits_{i \ne f} k_{fi}N_i\left( t \right)\frac{{\beta _{fi}}}{{\beta _{ff}}}}}{{k_{f{\mathrm{h}}}\mathop {\sum }\nolimits_{i \ne f} N_i\left( t \right)}},$$
(12)

and is the weighted average of the relative individual-level interaction coefficients βfi/βff between species i and the focal species f, weighted by the mean number of individuals of species i in the neighbourhoods of the individuals of the focal species (that is, c kfi Ni(t)). For competitive interactions, Bf ranges between zero and one; Bf = 1 indicates that heterospecific and conspecific neighbours compete equally, and smaller values of Bf indicate reduced competition with heterospecific neighbours. The denominator can be rewritten in terms of segregation kfh to all heterospecifics and the total number of heterospecifics ∑i≠f Ni(t).

The analytical expression of the equilibrium (equation (8)) relies on the assumption that the values of Bf are approximately constant in time. This assumption may not apply in our model during the initial burn-in phase of the simulations if the βfi/βff show large intraspecific variability (Supplementary Text and Figs. 1–5). The underlying mechanism is the central niche effect introduced by Stump45 where a species has reduced average fitness if it has high niche overlap with competitors.

Finally, the factors γff = ln(1 + bff βff) (bff βff)−1 and γfβ = ln(1+ bfβ βff) (bfβ βff)−1 describe the influence of the variance-to-mean ratios bff and bfβ of the gamma distribution of the crowding indices nkff and nkfβ, respectively. For high survival rates during one time step (for example, >85%), the values of γff and γfβ are close to one; in this case the exponential function in equation (1a) can be approximated by its linear expansion and γff = γfβ = 1.

In equilibrium we have (Nf(t + Δt) ‒ Nf(t))/Δt = 0, which leads, with equation (7), to:

$$N_f^{\ast} = \left( {K_f - \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}J^{\ast} } \right) \left( {1 - \frac{{\alpha _{f{\mathrm{h}}}}}{{\alpha _{ff}}}} \right)^{-1}$$
(13)

with \(K_f = - {\mathrm{ln}}\left( {\frac{{1 - r_f}}{{s_f}}} \right) \left( \alpha _{ff} \right)^{-1}\) and the total number of individuals being \(J^{\ast} = \sum _iN_i^{\ast}\). Rewriting equation (13) yields \(\frac{K_f}{J^{\ast}} = \left( \frac{N_f^{\ast}}{J^{\ast}} \right) \left(1- \frac{\alpha_{f{\mathrm{h}}}}{\alpha_{ff}}\right) + \frac{\alpha_{f{\mathrm{h}}}}{\alpha_{ff}}\). For αfh/αff < 1 we therefore find Kf < J*, which indicates that a multispecies forest would host more individuals than a monoculture. To estimate J* we sum equation (13) over all species i and find \(J^{\ast} = \mathop {\sum }\limits_i \frac{{K_i}}{{1 - \alpha _{i{\mathrm{h}}}/\alpha _{ii}}} - J^{\ast} \mathop {\sum }\limits_i \frac{{\alpha _{i{\mathrm{h}}}/\alpha _{ii}}}{{1 - \alpha _{i{\mathrm{h}}}/\alpha _{ii}}}\). Therefore, we obtain

$$J^{\ast} = \mathop {\sum }\limits_{i = 1}^S \frac{{K_i}}{{1 - \alpha _{i{\mathrm{h}}}/\alpha _{ii}}} \left( {1 + \mathop {\sum }\limits_{i = 1}^S \frac{{\alpha _{ih}/\alpha _{ii}}}{{1 - \alpha _{ih}/\alpha _{ii}}}} \right)^{-1} = \frac{{Sm_K}}{{1 + Sm_\alpha }}$$
(14)

with \(m_K = \frac{1}{S}\mathop {\sum }\limits_i^{\,} \frac{{K_i}}{{1 - \alpha _{i{\mathrm{h}}}/\alpha _{ii}}}\) and \(m_\alpha = \frac{1}{S}\mathop {\sum }\limits_i^{\,} \frac{{\alpha _{i{\mathrm{h}}}/\alpha _{ii}}}{{1 - \alpha _{i{\mathrm{h}}}/\alpha _{ii}}}\) being averages over the S species of the community.

All species have positive abundances at equilibrium if the two conditions µf = αfhJ*/αffKf < 1 and αfh/αff < 1 are met (see equation (9)). We now show that the chance that these conditions are satisfied for all species is larger if the values of µf show little intraspecific variability. To understand this, we assume that the µf can be approximated by their mean \(\bar \mu\). In this case \(J^{\ast} /\bar \mu\) is also approximately constant and we can replace Ki in equation (14) by \((J^{\ast} /\bar \mu ) (\alpha _{i{\mathrm{h}}}/\alpha _{ii})^{-1}\) and obtain

$$J^{\ast} = \mathop {\sum }\limits_{i = 1}^S \frac{{\left( {\frac{{J^{\ast} }}{{\bar \mu }}} \right)\frac{{\alpha _{i{\mathrm{h}}}}}{{\alpha _{ii}}}}}{{1 - \frac{{\alpha _{i{\mathrm{h}}}}}{{\alpha _{ii}}}}} \left( {1 + \mathop {\sum }\limits_{i = 1}^S \frac{{\frac{{\alpha _{i{\mathrm{h}}}}}{{\alpha _{ii}}}}}{{1 - \frac{{\alpha _{i{\mathrm{h}}}}}{{\alpha _{ii}}}}}} \right)^{-1} = \frac{{J^{\ast} }}{{\bar \mu }}\frac{{Sm_\alpha }}{{1 + Sm_\alpha }}$$
(15)

where S is the number of species in the community, and therefore

$$\bar \mu = \frac{{Sm_\alpha }}{{1 + Sm_\alpha }} < 1$$
(16)

Thus, in the case of a perfect interspecific balance in µf we always have a feasible equilibrium if αfh/αfh < 1, and species can go extinct only if the intraspecific variability in µf becomes too large. The smaller the mean value of µf, the more variability in µf is allowed. Equation (16) shows that \(\bar \mu\) is smaller if the number S of species in the community is smaller and/or if the mean value of mα is smaller.

Equation (16) also suggests that communities with more species need to show stronger species equivalence in µf because the term S mα(1 + S mα)−1 approaches a value of one for a large number of species S. This finding mirrors the results of analyses of Lotka–Volterra models with random interaction matrices11 that showed that the larger the number of species S, the more difficult it becomes to generate a feasible community.

Invasion criterion

Using the population-level interaction coefficients (equation (6)) in the macroscale model, we now derive conditions for coexistence based on the invasion criterion7,8 for a species m. The growth rate of an invading species m with low density M(t) into the equilibrium community of all other S – 1 species Ni(t) should be positive; thus, with equation (7), we have

$$\left( {r_m - 1} \right) + s_m \exp\left( { - \alpha _{mm}M\left( t \right) - \alpha _{m{\mathrm{h}}}\mathop {\sum }\limits_{i = 1}^{S - 1} N_i^{\ast} } \right) > 0.$$
(17)

Considering that \(J_m^{\ast} = \mathop {\sum}\nolimits_{i = 1}^{S - 1} {N_i^{\ast}}\) and \(\alpha_{mm} M(t) \ll \alpha_{f{\mathrm{h}}} J_m^{\ast}\) (that is, the invading species m is at low abundance and does not show strong clustering) we find \(- \ln \left( \frac{1-r_m}{s_m} \right) > \alpha_{m{\mathrm{h}}} J_m^ \ast\), and by dividing by αmm we obtain the invasion condition

$$\mu _m^i = \frac{{\alpha _{m{\mathrm{h}}}J_m^{\ast} }}{{\alpha _{mm}K_m}} < 1$$
(18)

which is basically identical to the condition for feasibility (equation (9)), but here the community size \(J_m^ \ast\) of the reduced community appears instead of the equilibrium community size J* of all species, including species m. Thus, a new species m is more likely to invade if it has a high value of the carrying capacity Km and if it more strongly reduces heterospecific interactions relative to conspecific interaction (that is, αmh/αmm is smaller). However, if the species is too efficient (that is, has too large a capacity Km and/or too low an αmh/αmm) it may increase the value of J* too much (equation (14)), thereby causing the extinction of the weakest species with the highest values of µf (that is, a too-low value of Km and a too-high value of αfh/αff). Equation (18) also suggest that an equilibrium with µf > 1 and αmh/αmm > 1 will be unstable.

Fitness–density covariance

To place our new spatial coexistence mechanism in the context of existing coexistence theory, we apply scale transition theory34 to our model version where spatial effects are the only potential coexistence mechanism (that is, all species have the same parameters and all individuals compete equally; βfi/βff = 1, Bf = 1).

Following equation (1a), the expected fitness of an individual k of a focal species f (that is, its expected contribution to the population after some defined interval of time Δt) in the macroscale model (equation (7)) is

$$\lambda _{kf} = r_f + s_{f}f\left( {W_k} \right) = r_f + s_f \exp\left( { - \beta _{ff}\left( {n_{kff} + \mathop {\sum }\limits_{i \ne f} n_{kfi}} \right)} \right)$$
(19)

where Wk = nkff + ∑i≠fnkfi is the fitness factor of individual k, f(Wk) = exp(−βff  Wk) is the fitness function and nkff and ∑i≠f nkfi are the number of conspecific and heterospecific neighbours, respectively, of individual k within distance R. The spatial average of the fitness factor over the entire plot is

$$\bar W_k = cN_f\left( t \right) + c\mathop {\sum }\limits_{i \ne f} N_i(t) = cJ(t)$$
(20)

where c = πR2 / A, and J(t) = ∑iNi(t) is the total number of individuals in the plot. Given that J(t) converges very quickly into equilibrium J* (Extended Data Fig. 5 and Supplementary Figs. 1 and 2), we find for the spatial average fitness \(\bar \lambda _f = 1\).

The average individual fitness \(\tilde \lambda _f(t)\) of a focal species f is the average of λk,f over all individuals k of species f and can be estimated for the macroscale model (equations (6 and 7)) as \(\tilde \lambda _f\left( t \right) = N_f\left( {t + {\Delta}t} \right)/N_f\left( t \right)\). A key ingredient of scale transition theory34 is that the fitness–density covariance is given by \({\mathrm{cov}} = \tilde \lambda _f\left( t \right) - \bar \lambda _f\). With equation (3) and γff ≈ γfβ ≈ 1 and \(\bar n_{f\beta } = \bar n_{f{\mathrm{h}}}\) we find

$$\tilde \lambda _f\left( t \right) - \bar \lambda _f = (r_f - 1) + s_f \exp( - \beta _{ff}(\bar n_{ff} + \bar n_{f{\mathrm{h}}}))$$
(21)

where the mean of the crowding indices is given by \(\bar n_{ff}(N_f) = ck_{ff}(N_f)N_f\) and \(\bar n_{f{\mathrm{h}}}(N_f) = ck_{f{\mathrm{h}}}(J^{\ast} - N_f)\) (equation (4)). Therefore, if clustering kff and segregation kfh are independent from abundance Nf, more abundant species have more neighbours, since

$$\bar n_{ff} + \bar n_{f{\mathrm{h}}} = c(k_{ff} - k_{f{\mathrm{h}}})N_f + ck_{f{\mathrm{h}}}J^{\ast}$$
(22)

Thus, a positive fitness–density covariance in our model means that individuals of a common species are more likely to be near more trees in total.

Extended Data Fig. 7 shows the quantities \(\bar n_{ff} + \bar n_{f{\mathrm{h}}}\), \(\bar n_{ff}\), \(\bar n_{f{\mathrm{h}}}\) and \(\tilde \lambda _f - \bar \lambda _f\) plotted over abundance Nt for data generated by our spatially explicit simulation model for the scenarios of stable and unstable dynamics (Extended Data Fig. 5b,c). Indeed, the stable simulations show a positive fitness–density covariance, however, there is no such trend for the dynamics of the unstable community (Extended Data Fig. 7g,h).

Spatial patterns will act as positive fitness–density covariance if, when a species becomes rare, areas of high conspecific crowding have fewer competitors. We tested this for the data generated by our simulation model and for the nine forest plots (Extended Data Fig. 8). We could estimate for each focal species f the covariance between the number of conspecific neighbours (that is, nkff) and the total number of neighbours (that is, nkff + nkfh) and demand that the covariance should be mostly positive and larger for more abundant species. However, since the quantity nkff appears in this test on both sides, a positive covariance can be expected. To compensate for this artefact, we instead used the covariance between the local dominance of conspecifics in the neighbourhood of individuals k (that is, dkff = nkff (nkff + nkfh)−1) and total number of neighbours (that is, nkff + nkfh) (Extended Data Fig. 8).

Spatially explicit simulation model

The model is a spatially explicit and stochastic implementation of the spatial multispecies model (equation (7)), similar to that of May et al.35,37 and Detto and Muller-Landau17, and simulates the dynamics of a community of S tree species in a given plot of a homogeneous environment (for example, 50 ha) in 5 yr time steps adapted to the ForestGEO census interval (Extended Data Fig. 5 and Supplementary Figs. 1 and 2). Only reproductive (adult) trees are considered, but size differences between them are not considered. During a given time step the model first simulates stochastic recruitment of reproductive trees and placement of recruits, and second, stochastic survival of adults that depends on the neighbourhood crowding indices for conspecifics (nkff) and heterospecifics (nkfβ) (but excluding recruits). In the next time step, the recruits count as reproductive adults and are subject to mortality. No immigration from a metacommunity is considered. To avoid edge effects, torus geometry is assumed.

The survival probability of an adult k of species f is given by \(s_f \exp \left(-\beta_{ff} \left( n_{kff}+n_{kf\beta}\right)\right)\) (equation (1a)). The two neighbourhood indices nkff and nkfβ describe the competitive neighbourhood of the focal individual k and sum up all conspecific and heterospecific neighbours, respectively, within distance R, but weight them with the relative individual-level interaction coefficients βfi/βff (refs. 19,21,26).

Each individual produces on average rf recruits, and their locations are determined by a type of Thomas process28 to obtain clustering. To this end, the spatial position of the recruits is determined by two independent mechanisms. First, a proportion 1 – pd of recruits is placed stochastically around randomly selected conspecific adults by using a two-dimensional kernel function (here a Gaussian with variance σ2). This is the most common way to generate species clustering in spatially explicit models17,18,35,36,37,38,39. Specifically, we first randomly select one parent for each of these recruits among the conspecific adults and then determine the position of the recruit by sampling from the kernel. Second, the remaining proportion pd of recruits is distributed in the same way around randomly placed cluster centres that are located independently of conspecific adults. This mode mimics spatial clustering of recruits independent of the parent locations42 in a simple way, such as contagious seed dispersal by animals50 or forest gaps that may imprint clumped distributions of recruits of pioneer species40. For each species we assume a density λfc of randomly distributed cluster centres, which have, at each time step, a probability pfp of changing location. For each of these recruits, we first randomly select one cluster centre among the cluster centres of the corresponding species and then determine the position of the recruit by sampling from the kernel. For the simulation shown in Extended Data Fig. 5a, the recruits were located at random positions within the plot.

Parameterization of the simulation model

Extended Data Fig. 5 shows simulations of the individual-based model conducted in a 200 ha area containing approximately 83,000 trees with, initially, 80 species. There was no immigration. The model parameters were the same for all species, and all species followed exactly the same model rules. We selected βfi = βff to obtain no differences in con- and heterospecific interactions and sf = 1 (no background mortality), and we adjusted the parameters βff = 0.0075 and rf = 0.1 to yield tree densities (415 ha−1) and an overall 5 yr mortality rate (10%) similar to those of trees with dbh ≥ 10 cm in the BCI plot51.

The Gaussian kernel used to place recruits around conspecific adults or around random cluster centres had a parameter σ = 10 m. There were 40 random cluster centres in total for each species that had a probability of pfp = 0.3 of changing location within one census interval. The only difference between the simulation shown in Extended Data Fig. 5b and the one shown in Extended Data Fig. 5c is that in the former, we used a proportion pd = 0.05 of recruits to be placed around randomly distributed cluster centres (that is, 95% of the recruits were placed close to their parents), but in the latter, we selected pd = 0.95 (that is, 95% of the recruits were placed around randomly distributed cluster centres). In our simulations, on average, one of these cluster centres received four recruits per time step, which were scattered within a radius of approximately 30 m, and received approximately 13 recruits during its lifetime (at each time step it had a probability of 0.3 of changing location). In contrast, in Extended Data Fig. 5a recruits were placed at random locations within the plot.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.