Coordination of physiological and structural traits in Amazon forest trees

Coordination of physiological and structural traits in Amazon forest trees S. Patiño, N. M. Fyllas, T. R. Baker, R. Paiva, C. A. Quesada, A. J. B. Santos3,4,†, M. Schwarz, H. ter Steege, O. L. Phillips, and J. Lloyd Max-Planck-Institut für Biogeochemie, Postfach 100164, 07701, Jena, Germany School of Geography, University of Leeds, LS2 9JT UK Institito Nacional de Pesquisas da Amazônia, Manaus, AM, Brazil Departamento de Ecologia, Universidade de Brası́lia, DF, Brazil Dept. of Plant Ecology and Biodiversity, Utrecht University, The Netherlands James Cook University, Cairns, Qld 4871, Australia †deceased

1 Introduction 5 Plant functional traits are widely used at both the ecology-evolution and ecologybiogeochemistry modelling interface. Sets of functional characters can serve as the basis of identifying important evolutionary adaptations that improve the success of different taxa at different environments as well as for obtaining a mechanistic basis of plant and ecosystem functioning. Over the last decade significant advances have been 10 made in terms of our understanding of plant trait inter-relationships and associated trade-offs (Reich et al., 1997;Westoby et al., 2002), especially in terms of the so called "leaf economic spectrum" (Wright et al., 2004) with well documented systematic and co-ordinated changes in leaf nitrogen and phosphorus concentrations, leaf mass per unit area, M A and leaf lifetimes. 15 Attention has also been paid to the relationships between physiological and structural characteristics of leaves and other plant traits. For example, it has been reported that leaf size declines with wood density, ρ w (Pickup et al., 2005;Wright et al., 2006Malhado et al., 2009). It has been suggested that this is because the ratio of leaf area to sapwood area (Φ LS ) should also decline with increasing wood den-20 sity due to hydraulic constraints ). Nevertheless, although Φ LS may decline with ρ w for trees in some ecosystems that are clearly water-limited (Ackerly, 2004; Cavender- Bares et al., 2004), Φ LS sometimes actually increases with ρ w Meinzer et al., 2008). The latter study also found that associated with these higher Φ LS and high wood density stems were lower stem hydraulic con-Introduction Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | be characterised by adaptations allowing them to function at more severe water deficits than is the case for low wood density species. The Panama study of Meinzer et al. (2008) also found that higher ρ w species tended to have higher M A . Although similar positive correlations between M A and ρ w have also been reported for other ecosystems (e.g. for sclerophyllous forest: Ishida et al.,5 2008) when examining the bivariate relationship between ρ w and M A across a range of tropical forest sites,  observed no significant relationship. Likewise, when examining variation in leaf and stem traits for 17 dipterocarp species growing in a common garden in southern China, Zhang and Cao (2009) also found no significant correlation between ρ w and M A . 10 Variations in M A may also be related to a suite of additional plant physiological characteristics (Poorter et al., 2009), varying negatively with dry-weight foliar nitrogen and phosphorus concentrations (Wright et al., 2004;Fyllas et al., 2009) as well as tending to increase with increasing tree height (Thomas and Bazzaz, 1999;Kenzo et al., 2006;Lloyd et al., 2010). Potential tree height, H max , has also been related to a number 15 of wood traits (Chave et al., 2009) with taller plants tending to have bigger conduits in their trunks, but fewer conduits overall (Preston et al., 2006), this reflecting longer path lengths to the transpiring tissue and therefore a need for wider vessels to maintain a high hydraulic conductance, K P (Coomes et al., 2007). Although lumen fraction and/or hydraulic conductivity are also sometimes negatively correlated with wood den-20 sity (e.g. Santiago et al., 2004a;Russo et al., 2010), this providing one explanation for smaller shade tolerant trees tending to have higher ρ w (e.g. Falster and Westoby 2005;Keeling et al., 2008), ρ w is not necessarily well correlated with H max , perhaps because of differences between species in the density of fibre tissue in the wood matrix (Poorter et al., 2009;Russo et al., 2010). Indeed, although many recent papers have suggested Introduction Seed size may also relate to the above plant functional traits. For example, one of "Corner's rules" describes a tendency for species with thick twigs to have large appendages (leaves and fruit). The range of viable seed size also tends to increase with plant height (Moles et al., 2005;Grubb et al., 2005). Forests on the more fertile soils of western Amazonia tend to have smaller average seed masses than their less fer-5 tile counterparts on the Guyana Shield and elsewhere (ter Steege et al., 2006), this perhaps being related to several advantages attributable to large seeded species under nutrient-poor conditions, viz. greater initial nutrient stores, greater initial root zone expansion, and increased mychorrizal infection, all of which would be expected to increase the probability of seedling survival (Foster, 1986). An additional factor related 10 to soil fertility may be that forests on the richer soil of the western Amazon tend to have higher tree turnover rates as opposed to the less fertile eastern Amazon forest. This may be primarily a consequence of differences in soil physical properties ) with more opportunities for gap-based regeneration in the west favouring smaller seeded species. 15 This paper presents new data on leaf and leaflet size and Φ LS for 661 species located in 52 plots across the Amazon Basin. The trees sampled form a subset of those also examined for variations in branch xylem density (Patiño et al., 2009), and for foliar nutrients, M A and δ 13 C (Fyllas et al., 2009). We investigate relationships between all these parameters as well as with genetic variations in H max ) and 20 seed mass (ter Steege and Hammond, 2001;ter Steege et al., 2006). Specifically, we were interested to assess the degree to which the observed variations in the studied structural and physiological traits were coordinated with each other, especially with their genetic versus environmental components considered separately. For the environmental effects, we were also interested to quantify the extent to which the observed 25 integrated structural and physiological responses were modulated by soil fertility and climate.
5088 2 Materials and methods

Study sites
In the analysis here, RAINFOR sample plots have been aggregated as discussed in Fyllas et al. (2009), with further plot details available in Patiño et al. (2009) and . As in Fyllas et al. (2009) plots were classified into two soil fertility 5 groups based on the measured total reserve bases . Ten plots in Fyllas et al. (2009) have not been included due to insufficient structural trait data having been collected.

Structural traits
For most trees sampled in Patiño et al. (2009) and Fyllas et al. (2009), and from the 10 same terminal branches for which data has already been presented in those studies, all leaves from the branch had also been counted. From that branch, a sub-sample of 10-20 leaves was randomly chosen to estimate individual leaf area, L A , and leaflet area, A (when a species had compound leaves), and to estimate the total leaf area of the branch. All age and size leaves or leaflets were selected for this analysis except for 15 very young leaves or those which were obviously senescent. The chosen leaves were usually scanned fresh on the same day of collection. When this was not possible the same day, they were stored for a maximum of two days in sealed plastic bags to avoid desiccation and any consequent reduction of the leaf area. Scans were analysed using "Win Folia Basic 2001a" (Regent Instruments Inc., 4040 rue Blain Quebec, QC., G2B 20 5C3 Canada) to obtain L A and A . The distal (sapwood + pith) and pith diameters for each branch were also measured with a digital caliper (Mitutoyo Corporation, Japan) with sapwood area, A S , then estimated by subtracting pith area from the total branch area with Φ LS =nL A /A S where n is the number of leaves distal to the piece of branch sampled andL A is the average area Foliar traits used here are as described/measured in Fyllas et al. (2009) and  and include leaf mass per unit area (M A ) and foliar [N], [C], [P], [Ca], [K] and [Mg] expressed on dry-weight basis. Foliar 13 C/ 12 C discrimination, ∆, was estimated from measurements of foliar δ 13 C (Fyllas et al., 2009) using an assumed value for the isotopic composition of source air equal to −8.0 ‰ (Farquhar et al., 1989) and 10 subsequently transformed to a diffusional limitation index, , according to which utilises the well known relationship between ∆ and the ratio of internal to ambient CO 2 concentrations, c i /c a (Farquhar et al., 1989). Equation (1) assumes that at current day c a , photosynthesis can be considered a roughly linear function of c i and with a relation analysis on selected soil and environmental predictors, with the soil variables reduced to three principal axes to avoid multicollinearity (Fyllas et al., 2009). The climatic variables of mean annual temperature, total annual precipitation, dry season precipitation and mean annual radiation were also examined. As extensively discussed in Fyllas et al. (2009) we dealt with spatial autocorrelation issues by fitting appropriate si-10 multaneous autoregressive models (SAR) which include a spatial error term (Lichstein et al., 2002) to help interpret the significance of full and partial Kendall's τ coefficients as a measure of association between plot-level trait effects and environmental predictors. 15 Inferred genetic effects were analysed jointly for species found on fertile versus infertile soils (excluding those found on both soil types) by calculating separate variancecovariance matrices for the two species groups and then using the common principal components (CPC) model of Flury (1988) as implemented by Phillips and Arnold (1999). Within this model, it is assumed that the two populations of species 20 have the same eigenvectors (principal components; denoted here as U) but that the relative loading of the various U as expressed through their eigenvalues (λ) may potentially vary between the two populations. Flury's model provides a hierarchy of tests corresponding to a range of possible relationships between matrices including equality, proportionality, common principal components, partial common principal components 25 or unrelated (Flury, 1988;Phillips and Arnold, 1999). CPC can thus be seen as a method for summarizing the variation in two or more matrices. Nevertheless, caution needs to be applied when using CPC to address the more complex goal of diagnosing 5093 Introduction and understanding the nature of the changes that underlie the difference between the matrices. This is because CPC tends to spread any differences over many of the vectors it extracts and often over all of them (Houle et al., 2002). As the CPC model does not strictly apply to correlation matrices (Flury, 1988), we standardised each variable before calculating the input variance-covariance matrix by 5 dividing each variable by its observed range (across both high and low fertility soils) as first proposed by Gower (1966) but, due to the presence of the occasional outlier, taking the effective range as the 0.1 to 0.9 quantiles. Standard errors of the U and λ for the CPC models were estimated assuming asymptotic normality as described in Flury (1988). 10 All other multivariate analyses (e.g. PCA of the derived environmental effects) were implemented with the ade4 package available within the R statistical platform with the environmental effect PCA undertaken on the correlation matrix. 15 The structural traits distributions along with those for M A and for the complete dataset divided to low and high fertility groups are shown in Fig. 1 with overall mean values, range and variances for each plot for all traits also provided in the Supplement (Table S1). The three leaf related traits introduced here (L A , A and Φ LS ) did not differ significantly between low and high fertility sites (Fig. 1). On the other hand, ρ x and S 20 showed significant differences between the two fertility groups, with their distributions shifted to the left for fertile sites, i.e. higher ρ x and S were found for species found on infertile soils. This is similar to the shifted distributions identified for most leaf mineral concentrations across fertility gradients (Fyllas et al., 2009) but in the opposite direction, i.e. with higher structural carbon and lower mineral investment in more fertile 25 environments.

Partitioning of the variance
The variation apportioned to different taxonomic levels varies for each of the traits examined (Fig. 2). When leaf size was expressed per leaflet, most of the variation was attributed at the species level (0.31) with the overall genetic component (i.e. family ± genus ± species) adding up to a very high (0.62) proportion. When leaf size was ex- 5 pressed at the leaf level, most of the variation was attributed at the family level (0.29) with a very high overall genetic component (0.71). In contrast to L A and A , plot level contributions to the total variance were substantial for the other structural traits: being around 0.30 for ρ x and 0.27 for Φ LS . These are not necessarily higher than their respective genetic components, but underline the importance of the site growing con- 10 ditions in influencing structural traits such as ρ x and Φ LS . This must have direct implications for different physiological processes, as for for which the environmental component was the dominant source of variation.

Bivariate relationships: raw data
These are not considered in any detail here, but for the interested reader data are 15 summarised in the Supplement, Table S2A.

Bivariate relationships: genetic components
Considering data from both low and high fertility sites together, Table 1 lists correlations and SMA slopes for the derived genetic components with this same information shown in more detail (including confidence intervals) in the Supplement (Table S2A) and with 20 low and high fertility species separated for OLS and SMA regression analyses in Table S2B. Within Table 1, the SMA slopes reflect the relationship y ↔ x, with the x as the column headers and the y being the row labels. Figures 3 through 6  Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | though, where interesting and/or informative, statistically less significant relationships are also considered.

Maximum tree height
Generally only poor correlations were observed for H max , these being significant only for log 10 (M A ) (p ≤ 0.001) and log 10 (S) (p ≤ 0.01). The M A ↔ H max and S ↔ H max re-5 lationships are shown in Fig. 3. Here, due to differences in the SMA slope and/or intercept between the species associated with the two soil fertility classes (see Supplement, Table S2B) we have fitted separate lines for species found on low and high fertility soils. This shows that for species associated with low fertility soils, both M A and S tend to be slightly higher at a given H max than their lower fertility counterparts. 10 Especially for S ↔ H max the variation is considerable, particularly at low H max , with S varying three orders of magnitude for H max between 10 and 30 m.

Branch xylem density
As detailed in Table 1, the derived genetic component of ρ x was negatively correlated with log 10 [P], log 10 [Ca], log 10 , [K] log 10 , ( A ) and positively associated with 15 log 10 (S) (p ≤ 0.001), with weaker but significant positive correlation being observed with log 10 (M A ) and with a negative correlation with log 10 [Mg] (p ≤ 0.01) and even less significant, negatively with log 10 (L A ) (p ≤ 0.05). Some of these relationships are illustrated in Fig. 4 which shows the relationships between ρ x and both [P] and [K] to be particularly compelling and, as is also the case for M A and S, there was no difference 20 between species associated with low versus high fertility soils. Note that all four panels in Fig. 4 are log-linear plots meaning that a linear change in ρ x causes a proportional change in the other variable. For example, an increase in ρ x from 500 to 600 kgm −3 is associated with a reduction in [P] of about 0.3 mgg −1 but a further increase from 600 to 700 kgm −3 is associated with an additional reduction in [P] of only just over 0.

Leaf area: sapwood area ratio
Reasonably strong correlations were found between log 10 (Φ LS ) and log 10 (M A ), log 10 [N] and log 10 (L A ) (p ≤ 0.001) with the relationship between log 10 (Φ LS ) and log 10 [P] also significant (p ≤ 0.01). The relevant biplots are shown in Fig. 5. Because of the loglog nature of the bivariate relationships the slopes can in this case be interpreted as 5 scaling coefficients, the most notable being a value of 1/0.19 = 5.3 for L A ↔ Φ LS . This suggests that for each doubling of Φ LS there is a 2 5.3 or 40-fold increase in L A . Thus, tropical tree species with a higher L A tend to have relatively fewer leaves per unit stem cross-sectional area (A L ) meaning that Φ LS is a mostly conserved (but still significantly variable) plant trait. The slope for the genetic component M A ↔ Φ LS 10 relationship is 1/−1.27 = −0.79. Thus, as Φ LS increases across species, then M A declines proportionally less. That is to say, species with a higher Φ LS also tend to carry a greater weight of (generally larger) leaves per unit A L with those leaves also tending to have higher foliar [N] and [P]. 15 Strong positive correlations (p ≤ 0.001) were also observed for log 10 (L A ) with log 10 [N] and log 10 [P] as well as between log 10 [Ca] and S. These relationships are shown in Fig. 6 for which, from the slopes of

Common Principal Component modelling (genetic components)
Results from the CPC modelling are shown in Table 2, with the full model output, details of the rationale for eigenvector inclusion and assessments of the overall model fit  as given in the Supplement Tables S3, S4 and S5 and their accompanying captions. These considerations gave rise to five eigenvectors being selected, listed in Table 2 in 5 order of max [λ low,j ,λ high,j ] where λ low,j and λ high,j are the values for the j th characteristic root for the low and high fertility species respectively. The first eigenvector, U 1 , had somewhat higher λ for high versus low fertility associated species (accounting for 0.24 and 0.27 of the dataset variance respectively) and with high positive coefficients for all three foliar cations and to a lesser extent foliar [P], 10 and negative coefficients for foliar [C] with smaller but still significant coefficients for M A and S. In terms of cations, carbon and M A , this first component seems similar to that first described by Poorter and de Jong (1999) and thus we dub it the Poorter-De Jong (PDJ) dimension, relative to M A and Φ LS for d RW cf. d FW .

9
The fourth component axis is dominated by S and Φ LS with these coefficients of different 0 sign. Associated with the higher S are also lower [Ca] but higher foliar [P] and L A . With 1 lower values for their coefficients and higher standard errors, also being of different sign, are 2 the M A and [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the 3 population variance for low and high fertility species respectively) seems to be dominated by the 4 presence of large seeded members of the Leguminaceae whose importance in the phytogeography 5 of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote 6 this dimension as d TS .

7
The last eigenvector included in our analysis,, U 5 , differs from the others in having a sub-8 stantially greater importance for low fertility versus high fertility species (accounting for 0.09 9 and 0.04 of the population variances respectively). This component is characterised by H max 0 and M A having opposite signs (in contrast to d FW ) and with higher S and also being asso-1 ciated with a lower H max along with a less substantial but significant coefficient for ρ x . Also 2 of influence in characterising U 5 are greater foliar [C] associated with the higher M A and .

3
Although, U 5 presents some traits combinations as reported previously in the literature, this 4 component, mostly related with species found at low fertility soils, it does not seem to have 5 been recognised before. It is thus here denoted as d PFL .
6 PDJ . The second component, U 2 , accounts for an additional 0.18 and 0.19 of the dataset 15 variances for low and high fertility species respectively, and is characterised by high positive coefficients for foliar [N] and [P] as well as L A and, to a lesser extent, Φ LS . Also characterised by modestly negative coefficients for M A and foliar [Mg]. In terms of [N], [P] and M A , U 2 , seems to reflect some components of what is considered the classic leaf economic spectrum (Reich et al., 1997;Wright et al., 2004 RW it emerges as the dominant term for U 3 along with M A and, of opposite sign, Φ LS . Also of note here is the relatively high value for the coefficient of the diffusion limitation index, which is positively associated with both H max and M A . Interestingly, for this component L A varies 25 in the opposite direction to Φ LS (albeit with a large standard error) suggesting that there is a tendency towards considerably fewer but also significantly larger leaves in taller statured species. There also being a modest but significant negative contribution of ρ x to this dimension. We consider U 3 , which on its own accounts for 0.08 and 0.10 of the variation in the dataset respectively, to contain several features similar to those described by Falster and Westoby (2005) for climax tropical forest in Australia, and it is thus denoted as presence of large seeded members of the Leguminaceae whose importance in the phytogeography of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote this dimension as d TS .
The last eigenvector included in our analysis,, U 5 , differs from the others in having a substantially greater importance for low fertility versus high fertility species (accounting for 0.09 and 0.04 of the population variances respectively). This component is characterised by H max and M A having opposite signs (in contrast to d FW ) and with higher S and also being associated with a lower H max along with a less substantial but significant coefficient for ρ x . Also of influence in characterising U 5 are greater foliar [C] associated with the higher M A and .
Although, U 5 presents some traits combinations as reported previously in the literature, this component, mostly related with species found at low fertility soils, it does not seem to have been recognised before. It is thus here denoted as d PFL .
Overall the five eigenvectors selected, all of which we believe to be physiologically relevant (see Supplementary Information), accounted for 0.68 of the total variance for both low and high fertility soil species.
3.6 Bivariate relationships: Environmental components FW . The first three axes species scores (normalised to ± 100) are plotted against each 5 other in Fig. 8a-c. This shows the required lack of any systematic correlations between the species scores as expected for the output from any good fit of a principle components model. Clearly a wide range of combinations of these three trait dimensions can occur. But with Fig. 8a Figure 8d shows the major components of the three major CPCs and their overlap of traits in diagrammatic form. This illustrates that many traits seem to be "shared", especially M A which is an important factor for all three of and L A . With lower values for their coefficients and higher standard errors, also being of different sign, are the M A and [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the population variance for low and high fertility species respectively) seems to be dominated by the presence of large seeded members of the 25 Leguminaceae whose importance in the phytogeography of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote this dimension as The last eigenvector included in our analysis, U 5 , differs from the others in having a substantially greater importance for low fertility versus high fertility species (accounting for 0.09 and 0.04 of the population variances respectively). This component is characterised by H max and M A having opposite signs (in contrast to presence of large seeded members of the Leguminaceae whose importa  3.6 Bivariate relationships: Environmental com 410 FW ) and with higher S and also being associated with a lower H max along with a less substantial but sig-5 nificant coefficient for ρ x . Also of influence in characterising U 5 are greater foliar [C] associated with the higher M A and . Although, U 5 presents some traits combinations as reported previously in the literature, this component, mostly related with species found at low fertility soils, it does not seem to have been recognised before. It is thus here denoted as lower values for their coefficients and higher standard errors, also being of different sign, are the M A and [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the population variance for low and high fertility species respectively) seems to be dominated by the presence of large seeded members of the Leguminaceae whose importance in the phytogeography of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote this dimension as d TS .
The last eigenvector included in our analysis,, U 5 , differs from the others in having a sub- Overall the five eigenvectors selected, all of which we believe to be physiologically relevant (see Supplementary Information), accounted for 0.68 of the total variance for both low and PFL .

10
Overall the five eigenvectors selected, all of which we believe to be physiologically relevant (see Supplement), accounted for 0.68 of the total variance for both low and high fertility soil species.

Bivariate relationships: environmental components
Considering data from both low and high fertility sites together, Table 3 lists correlations 15 and SMA slopes for the environmental effects with this information provided in more detail (including confidence intervals) in the Supplement (Table S2A). As for Table 1, the SMA slopes reflect the relationship y ↔ x, with the x as the column headers and the y being the row labels. For the structural traits, the most significant relationships are all negative and appear between ρ x and log 10 [P], log 10 [Ca], log 10 [K] and, to a lesser 20 extent log 10 ( A ). The slopes observed (−0.26 to −0.41) are, however, much less than for the associated slopes for the genetic components as listed in Table 1 (−0.37 to −0.72). This means that a per unit change in ρ x is accompanied by a proportionally lesser change in these elements when environment (as opposed to genotype) was the inferred source of variation. Two of these relationships are illustrated in Fig. 7 which 25 shows the relationships of both [P] and [K] with ρ x to be quite strong and consistent across soil fertility types.

Principal component analysis of environmental effects
Especially given the strong relationships between ρ x and the foliar cation environmental components shown above, it was of additional interest to see if coordinated structural/leaf biochemical responses to the environment exist for Amazon forest. We therefore undertook a PCA analysis of the full plot effects correlation matrix (excluding H max 5 and S both of which were considered to be environmentally invariant for any given species) with the results shown in Table 4. This shows that 0.33 of the total variation in the 11 traits examined could be explained by the first PCA axis (ů 1 ) with ρ x an important contributor and this also relating positively to foliar [C] and M A , but negatively with all foliar nutrients examined and also with . The second axis of the PCA on the plot 10 effects correlation matrix (ů 2 ) is also significant, accounting for 0.25 of the variance, with substantial negative weightings for M A , foliar [C] and (and to a lesser extent foliar [P]) being balanced by positive weightings for foliar [Mg] in particular, but also with contributions from Φ LS and ρ x . Taken together, the first two axes of the environmental effects PCA account for an 15 impressive 0.58 of the total variance observed. This suggests that the modulation of the trait characteristics of tropical trees as a result of variations in their environment occurs in a highly coordinated manner.

Relationship between plot effect PCAs and soil/climate
Given the strong coherence in plot effect responses for the various traits as indicated 20 by the PCA analysis of Table 4 we were interested to see if any of the corresponding plot axes scores correlated with previously derived soil and/or climate characteristics of the same sample plots. The most significant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů 1 as a function of the first soil PCA axis of Fyllas et al. (2009), the latter considered a strong integrated measure of soil fertility and  Table 4 also suggesting that for any given species, both and ρ w also decline with increasing precipitation and, somewhat counter intuitively, with ncreasing.
Finally, as in Fyllas et al. (2009) we show values for Kendall's partial τ (denoted τ p ) for traits of interest as well as ů 1 and ů 2 as functions of f f , f t , T a , P a and Q a in Table 5.
e the calculated value of τ p and associated probability giving an indication of the effect ach soil/environmental parameter after accounting for the effect of the other four. Taking account to the potential confounding effects of spatial autocorrelation (Fyllas et al., 2009) only consider relationships with p ≤ 0.01 or better. As for the (full) Kendall's τ shown ig. 9, Table 5 suggests the f to be superior predictors than the individual variables, the 19 F increases. Interestingly, the Kendall's τ for this plot of ů 1 versus elationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů 1 rst soil PCA axis of Fyllas et al. (2009) Table 5.
alue of τ p and associated probability giving an indication of the effect ental parameter after accounting for the effect of the other four. Taking tential confounding effects of spatial autocorrelation (Fyllas et al., 2009) tionships with p ≤ 0.01 or better. As for the (full) Kendall's τ shown gests the f to be superior predictors than the individual variables, the 19 F of 0.63 is greater than for any of the original variables examined by Fyllas et al. (2009), the highest of which was 0.56 for foliar [P]. Comparison with Fyllas et al. (2009) also shows that the ů 2 contains significant weightings of leaf-level variables that, individually, were all strongly correlated with mean annual precipitation 5 (P A ) viz. positive correlations with foliar [C] and M A and a negative correlation with foliar [Mg]. It is therefore not surprising, as is shown in the middle panel of Fig. 8, that ů 2 and P A also show strong association, but with examination of Table 4 also suggesting that for any given species, both Φ LS and ρ w also decline with increasing precipitation and, somewhat counter intuitively, with increasing.   Table 4 we were interested to see if any of the corresponding plot axes scores 442 correlated with previously derived soil and/or climate characteristics of the same sample plots.

443
The most significant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů   Table 5.  Table 4 we were interested to see if any of the corresponding plot axes scores 442 correlated with previously derived soil and/or climate characteristics of the same sample plots.

443
The most significant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů   Table 5.
459 t , T a , P a and Q a in Table 5. Here the calculated value of τ p and associated probability giving an indication of the effect of each soil/environmental parameter after accounting for the effect of the other four. Taking into account to the potential confounding effects of spatial autocorrelation 15 (Fyllas et al., 2009) we only consider relationships with p ≤ 0.01 or better. As for the (full) Kendall's τ shown in Fig. 9, Table 5 suggests the manner.  Table 4 we were interested to see if any of the corresponding plot axes scores 442 correlated with previously derived soil and/or climate characteristics of the same sample plots.

443
The most significant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů

Discussion
The extent to which variations in key plant functional traits are coordinated, especially with species considered the prime source of variability, has been a key focus of plant ecophysiological research over the recent decade with several important papers investigating relationships between leaf physiology, nutrients and structure and/or wood 25 traits at the global scale (Wright et al., 2004;Chave et al., 2009;Zanne et al., 2010). Although tropical forests tree communities have been one area of focus for such plant functional trait studies, work to date has generally been limited to only one or two sample sites per study (Thomas and Bazzaz, 1999;Santiago et al., 2004a;Kenzo et al., 2006;Sterck et al., 2006;Santiago and Wright, 2007;Poorter, 2008;Zhang and Cao, 2009;Baraloto et al., 2010). Moreover, where geographically or edaphically diverse plots have been examined they have often been analysed as if environment and/or 5 soil has had no important effect on the absolute values of the various traits examined or their interrelationships (e.g., Wright et al., 2006;Baraloto et al., 2010) or limited in terms of the number of study sites and/or traits examined (e.g., . Our study contrasts in that it has involved the integrated measurement of 11 physiological and structural traits for 1021 forest trees sampled from 53 sites of widely varying 10 soil fertility across the Amazon Basin, with a statistical analysis specifically designed to disentangle the effects of environment/soil versus genotype on traits observed and their associations. Some of the data used here have been presented previously (Fyllas et al., 2009;Patiño et al., 2009), with the current analysis integrating those data with structural 15 traits introduced as part of this study (viz. L A , A , Φ LS , S and H max ) as well as with foliar 13 C/ 12 C ratios as reinterpreted through the diffusional limitation index, , as defined by Eq.
(1). We first consider the bivariate relationships between the structural components introduced as part of this study as well as relationships between these structural traits and the others already presented (Fyllas et al., 2009;Patiño et al., 2009) and then 20 the extent to which variations of these traits integrate and coordinate in response to variations in genotype and/or environment,

Maximum tree height, branch xylem density and leaf mass per unit area
These three structural traits have often been associated with each other with significant 25 positive ρ x ↔M A correlations such as for our genetic component in Fig. 4 also reported by Bucci et al. (2004), Ishida et al. (2008) and Meinzer et al. (2008). These studies have interpreted this relationship in terms of higher density wood species having lower hydraulic conductances leading to a requirement for more robust leaves capable of sustaining more severe soil water deficits. This notion is supported by more negative osmotic potentials being reported for the leaves of higher M A and ρ x species (Bucci et al., 2004;Ishida et al., 2008;Meinzer et al., 2008). On the other hand, it is also the 5 case that M A tends to increase with actual or potential (maximum) tree height (Falster and Westoby, 2005;Kenzo et al., 2008;Lloyd et al., 2010) but that ρ w and H max are sometimes negatively (as opposed to positively) correlated (Falster and Westoby, 2005;. One reason for this apparent contradiction may be that wood density and xylem 10 vessel traits do not necessarily represent the same axis of ecophysiological variation (Preston et al., 2006;Martinéz-Cabrera et al., 2009;Poorter et al., 2010;Baraloto et al., 2010). This is because, in contrast to conifers, angiosperm vessel traits can vary significantly (especially in terms of diameter distributions and density) allowing for large changes in stem hydraulic conductance (K S ) for only small changes in ρ w 15 (Roderick and Berry, 2001). Moreover, wood density can also be affected by many other characteristics such as fibre density, extent and structure of ray parenchyma and (not unrelated to the above) wood air, solid and water fractions (Preston et al., 2006;Chave et al., 2009;Poorter et al., 2010;Zanne et al., 2010). Thus, although ρ x and K S may be well correlated in some cases (e.g., Santiago et al., 2004a) this may not be 20 a strict functional linkage. Noting that for individual species ρ w and ρ x are likely to be closely correlated, at least for individual sites Sarmiento et al., 2011) and that high-density wood is associated with disturbance related advantages other than resistance to high wind (Anten and Schieving, 2010) -for example pathogen attacks (Augspurger, 1984) or a greater resistance to damage caused by falling canopy 25 debris (Putz et al., 1983;King, 1986;Chao et al., 2008), we can thus simply interpret Fig. 4a-c as indicating that tropical tree species with traits associated with a higher photosynthetic productivity such as a high foliar [P] (Domingues et al., 2010), also tend to invest less towards wood defensive strategies. The absence of a direct causal link underlying the relationships of Fig. 4a-c is also suggested by the CPC analysis of Table 3 where there is no contribution of ρ x for the second component identified, presence of large seeded members of the Leguminaceae whose importance in the phytogeography 395 of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote 396 this dimension as d TS .

397
The last eigenvector included in our analysis,, U 5 , differs from the others in having a sub-398 stantially greater importance for low fertility versus high fertility species (accounting for 0.09 399 and 0.04 of the population variances respectively). This component is characterised by H max 400 and M A having opposite signs (in contrast to d FW ) and with higher S and also being asso-401 ciated with a lower H max along with a less substantial but significant coefficient for ρ x . Also  Considering data from both low and high fertility sites together, Table 3 lists correlations 411 RW . On the other hand, a decline in ρ x does constitute one component of integrated tropical tree responses to improved soil fertility (Tables 4  and 5, Fig. 9). Similar reductions in ρ w with improved soil nutrient status, especially in 5 terms of phosphorus have been reported before (Omolodun et al., 1991;Raymond and Muneri, 2000) and for Eucalyptus at least, seems to be associated with less secondary thickening of stem fibre cell walls with some evidence of an increased xylem conduit area (Thomas et al., 2005). By similar reasoning then, those results would argue that a lower ρ w accompanying improved foliar nutrient status on more fertile soils could simply 10 represent a growth response to the most likely greater rate of carbon supply on such soils. This is as opposed to the lower ρ w observed being necessarily associated with a higher K S as might be postulated to be required to sustain the higher photosynthetic rates likely accompanying higher values of ů 1 (Santiago et al., 2004a; Table 4).
Our observation of significant within-species variation in ρ x as illustrated in 15 more detail by Patiño et al. (2009) and also observed for ρ w by Omolodun et al. (1991), Hernández and Restrepo (1995), Gonzalez and Fisher (1998), Weber and Montes (2008) and Sungpalee et al. (2009), shows important intraspecific variation in xylem and/or wood density even within the one plot (as also evidenced by the "residual" term for ρ x in Fig. 1) as well as being affected by soil fertility as mentioned above. 20 Thus, although we do not dispute that xylem traits and ρ w /ρ x may not necessarily be closely or mechanistically linked (as discussed above), studies which simply compare xylem trait "species values" as measured in one study or studies with values of ρ w /ρ x for the same species but gathered from a completely independent source (Russo et al., 2010;Zanne et al., 2010)  So, is the presence of large diameter xylem vessels with an associated high K S associated with a greater H max (e.g., Poorter et al., 2010;Zach et al., 2010) functionally 5 related to the tendency of mature forests species of a greater H max to also have a lower ρ w (Falster and Westoby, 2005;Poorter et al., 2009)? Or does it more simply reflect that the fast-growing and light-demanding species characteristic of "dynamic" tropical forests also tend to have a lower ρ w -this presumably allowing a faster height and diameter growth rate? On the basis of the discussion above, we 10 suggest the latter, also noting that ρ w is actually generally better correlated with juvenile light-exposure than H max Poorter et al., 2009).
If that argument is correct, then any associations between ρ x and H max are likely to be site specific: depending for example on the disturbance regime and the relative frequencies of regeneration in gaps versus shade (Sheil and Burslem, 2003). Indeed, 15 although the current study found no significant relationship between H max and the genetic component of ρ x when simply pooling all species together (p ≥ 0.1; Table 1), an implementation of a multilevel modelling approach similar to that applied by  found a significant (random) plot component modulating the ρ x and H max relationship (χ 2 test, p ≤ 0.001) associated with a negative relationship between the 20 within-plot ρ x vs. H max slope and the associated stand level turnover rate (Fig. 10). For forests with high tree turnover rates (k), such as are typically found on eutric soils and/or limitations to root growth and function (Phillips et al., 2004;, ρ x tends to decrease with H max , but with the opposite observed for less dynamic forests (lower k) on deeper and less fertile soils. This points to important differences 25 in species partitioning in respect to within-canopy light climate between the two forest types. We suggest that although for more dynamic forests with a high k, the classic paradigm of fast growing "pioneer" type species dominating the upper-canopy may apply (with higher wood density more shade-tolerant species lower down), for slower growing forests with a characteristic low k, that it does not apply. Rather, it is trees of a high ρ w (and by virtue of this presumably slower height and volume growth rates) that eventually achieve canopy dominance on deep but infertile soils where stand-level growth rates and k are similarly low. We suspect that that this observation may be due, at least in part, to the different dis-5 tributions and roles of members of the dominant Amazon tree family, Fabaceae, across Amazonia. Members of this family tend to have a higher wood density than average (Malhado et al., 2010;Powers and Tifin, 2010) but different members of this family play different roles in the more fertile dynamic forests of western Amazonia (where they are often relatively small statured genera such as is the case for most Inga species) as op-10 posed to the less dynamic forests of (north) eastern and central Amazon where they are more generally common as emergents such as Hymenaea and Swartzia  and, overall, represented disproportionately by the very high ρ w Caesaplinoid sub-family. This difference can also be seen in the data presented by Malhado et al. (2010) and this perhaps being attributable to the the two different forest regions 15 having been exposed to different geomorphological and edaphic conditions over extended periods of time (Hammond, 2005); this leading to (other things being equal) the forests of the western region of the Amazon Basin to be relatively short (Feldpausch et al., 2011) and also relatively dynamic (Phillips et al., 2004;. The positive relationship between M A and H max of Fig. 3a and as also evident in the 20 data of Falster and Westoby (2005) can also be inferred from the positive M A vs. tree height relationships as reported by Thomas and Bazzaz (1999), Kenzo et al. (2006) and Lloyd et al. (2010). This is also seen within  Table 3, with the leaves of (potentially) taller trees being thicker (Kenzo et al., 2006;Rozendaal et al., 2006) with a greater mesophyll thickness associated with a higher photosyn-25 thetic capacity per unit area (Kenzo et al., 2006). This increase in M A with tree height being mostly associated with a greater mesophyll thickness should allow for a more efficient use of the higher rates of insolation towards the canopy top through higher photosynthetic capacities per unit leaf area (Rijkers et al., 2000). Along with more negative osmotic potentials, the greater leaf densities associated with a higher M A and H max should also help sustain leaves of such taller trees in the face of the more severe water deficits expected for sun exposed leaves higher up in the canopy (Cavaleri et al., 2010;Lloyd et al., 2010).

5
Species with intrinsically higher foliar nutrient concentrations also tend to be found on more fertile soils (Fyllas et al., 2009), and so the positive correlation between the genetic components of leaf size variation, foliar [N] and foliar [P] observed here ( Fig. 6) is consistent with the observation that Australian tropical forest tree species associated with poorer soils tend to have smaller leaves than those associated with 10 more eutric conditions (Webb, 1968), as was also found for south-eastern Australian woodland species once precipitation effects were also taken into account (McDonald et al., 2003). Such a relationships has also been observed for pre-montane subtropical forest species in Argentina (Easdale and Healey, 2009) and has been suggested to be a widespread phenomenon (Givnish, 1987) perhaps being explainable by low 15 N and/or P leaves typically having lower gas exchange rates than those of a higher fertility status (Domingues et al., 2010); with associated lower latent heat loss rates due to lower stomatal conductances. This would give rise to a greater rate of sensible heat loss being required to avoid over-heating during times of high insolation being achieved through the higher boundary layer conductance of smaller leaf sizes (Yates et 20 al., 2010). Alternatively and consistent with the general notion of plants growing on less fertile soils having more conservative growth strategies , smaller leaves may be favoured on low nutrient soils despite their relatively higher construction costs. This is because they also have shorter expansion times with an associated reduction in herbivory losses during this susceptible phase of foliar development (Moles 25 and Westoby, 2000). If the "heat budget" explanation were to be correct, then an even better correlation with A would be expected for both foliar [N] and [P]. But this was not the case (Table 1)  On the other hand, the relationships between leaf size and expansion time does not appear to differ strongly between simple vs. compound leaves (Moles and Westoby, 2000) suggesting that the herbivory hypothesis may be the more correct.
In apparent contrast to the above result, Malhado et al. (2009) found no strong relationship between soil fertility and average leaf size across a range of Amazonian 5 forests. Their study was, however, based on an ordinal analysis of leaf-sizes based on herbarium samples which, as well as being inherently less accurate, may also have been confounded by the possibility of many of their herbarium samples being collected lower down in the forest canopy (including saplings) where for any given species individual leaf areas may be greater than for sun-exposed leaves as sampled 10 here (Bongers and Popma, 1988;Rozendaal et al., 2006;Markesteijn et al., 2007;Poorter and Rozendaal, 2008). A further consideration may be that many herbarium samples are deliberately taken with flowers and/or fruits present to aid species identification and, as is well known in the horticultural literature for example (e.g., Syvertsen et al., 2003) leaves proximal to developing fruits may be appreciably smaller than those 15 on non-fruiting branches, this also being broadly consistent with the evidence of reproductive/vegetative competition we discuss later in Sect. 4.2.4.
Although not significant across the dataset as a whole, there was a significant negative correlation between L A and ρ x for species characteristic of low fertility sites (r 2 = −0.17,p ≤ 0.05: Supplement, Table S2B) as has also been reported for Australian 20 tree/shrub species by Pickup et al. (2005) and  and for neotropical forest tree species by Swenson and Enquist (2008), Malhado et al. (2009), Baraloto et al. (2010) and, with a much lower correlation (r 2 = −0.02) by Wright et al. (2006).
Exactly as to why this should be the case is currently unclear. Earlier arguments have revolved around not only ρ x and K S being closely linked, but also with the assumption 25 that variations in L A should to a large extent reflect variations in Φ LS . But, as discussed in Sect. 4.1.2, wood density and plant hydraulics may not be as closely linked as once thought and, although Φ LS is indeed correlated with L A (Fig. 6) Table S2B). This suggests that for tropical trees at least, this correlation may be more "casual" than mechanistic. Indeed, both for the dataset as a whole and for the individual fertility groupings, ρ x was better (negatively) correlated with A than L A (Table 1, Supplement, Table S2B). Given that compound leaves are generally associated with faster diameter increment species (Givnish, 1978;Malhado 5 et al., 2010) as is a generally lower ρ x (Keeling et al., 2008) this then suggests that the negative correlation between laminar size and wood density may just reflect both traits being associated with faster growth rates. As well as tending to have lower ρ w (Sect. 4.1.2) such species also tend to exhibit less branching than more shade tolerant species Poorter and Rozendaal, 2006;Takahashi and Mikami, 10 2008). Presumably (along with wider spacings) this allows for larger leafed uppercanopy species to have greater rates of direct light interception . The statistically significant decline in M A with Φ LS as on our Fig. 6 (Meinzer et al., 2008;Zhang and Cao, 2009), perhaps due to sampling a smaller range of genetic variability, although it is notable that working with 20 a range of emergent or upper-canopy dipterocarp species, Zhang and Chao (2009) did find a significant negative relationship between Φ LS and leaf thickness, the latter often being well associated with variations in M A with tree height for dipterocarp species (Kenzo et al., 2006). Sampling across a range of sites in south-eastern Australia, Pickup et al. (2005) also found a negative relationship between M A with Φ LS but 25 this relationship was, overall, not significant for species sampled within individual sites. Our own data suggest a stronger linkage of M A with Φ LS than either L A or (indeed even of different sign) A . This suggests (as is discussed further in Sect. 4.2) that this linkage may be mostly related to plant hydraulics considerations. The positive relationship between A and M A may reflect constraints on the range of possible combinations of leaf(let) size and M A , with larger laminar areas necessarily requiring a greater (minimum) M A due to structural constraints (Grubb, 1998). Not surprisingly, L A and Φ LS were related, but with a scaling coefficient of only 0.17, meaning that a greater leaf size was to a substantial degree compensated for by re-5 duced numbers of leaves per unit sapwood area A S . This points to Φ LS being a rather conserved entity as has also been reported by others (e.g., Westoby and Wright, 2003). Φ LS was also correlated with foliar [P] and [N] (Fig. 6), although this correlation was weaker for L A especially in the case of foliar phosphorus. But for both nitrogen and phosphorus, the slope was still positive and close to 1.0. Thus tropical tree species 10 with larger leaves tend to have not only higher [P] and [N] (and by implication higher gas exchange rates) but also a higher Φ LS . As there is little evidence of greater diffusional limitations on gas exchange for such leaves (as shown by the lack of any significant relationship between Φ LS , [N], [P] or L A with ), this implies that accompanying a higher Φ LS are also increased K S as also observed by Vander Willigen et 15 al. (2000) for subtropical trees and also by Cavender- Bares and Holbrook (2001) for a range of Quercus species.
As would be anticipated on the basis of the very low proportion of the total variability within the dataset attributable to environment (Fig. 2), neither L A , A showed strong variation with environment and/or soil characteristics (Table 5), although Φ LS does 20 contribute slightly to the second environmental response trait PCA, (ů 2 ; Table 4), with it's positive coefficient then suggesting a decline in Φ LS with increasing precipitation (Fig. 9). Although this seems counter-intuitive with Φ LS generally declining with reduced soil water availability such as to the average (whole) plant root-to-leaf hydraulic conductivity (Magnani et al., 2002;Addington et al., 2006;Carter and White, 2009), 25 as is explained in Sect. 4.3 it is consistent with intra-specific variation giving rise to population of more "evergreen-like" phenotypes as dry-season lengths decrease.

Seed mass
As has also been reported by others, seed mass showed significant positive correlations with both H max (Fig. 3; Foster and Janson, 1985;Hammond and Brown, 1995;Kelly, 1995;Metcalfe and Grubb, 1995;Grubb and Coomes, 1997), and ρ w ( Fig. 4; ter Steege and Hammond, 2001), although the latter relationship was not detected 5 by Wright et al. (2006), perhaps because of methodological issues (Williamson and Weimann, 2010). Generally speaking, a greater seed size should confer a greater ability for survival and thus tend to be favoured under less favorable environmental conditions such as deep shade or nutrient poor soils ter Steege et al., 2006). This readily provides a basis for indirect correlations between S and 10 wood/stem density to exist as high values of ρ x or ρ w are similarly associated with shade and/or dystrophic soil conditions (Sect. 4.1; Kitajima, 1994). More controversial is the basis of the relationship between S and H max . For example, the suggestion of Moles et al. (2005) that, by analogy with Charnov's life history theory for mammals, larger statured species may have larger seeds because they require a longer juve- 15 nile period has been contested by Grubb et al. (2005) who maintain that it is simply the range of feasible seed sizes that a species can have that increases with H max . Moreover, for tropical trees at least, there is probably little correlation between juvenile period and H max , with faster-growing low-wood density pioneer type trees attaining greater heights than their smaller statured shade counterparts and in a shorter time 20 . Indeed, by applying a general scaling model Falster et al. (2008) showed that longer juvenile periods alone are not sufficient to generate a correlation between height and seed size. They suggested that size-asymmetric competition among recruits (i.e. competition for light) may be the main factor having caused evolution towards larger offspring size. In this scheme of things, correlations with adult height 25 comes about because larger adults have a greater total reproductive output, thus generating more intense competition among recruits. That model tested dynamics only with a single species at a time, but it is likely to still apply in more complex species systems such as tropical forests, even though relative size at the onset of maturity is much more variable for tropical trees species than for animal systems (Thomas, 1996;Wright et al., 2005). We also consider it unlikely that simple physical constraints can account for much of the relationships (also seen in Fig. 3) as even small statured species can have reasonably large seeds and/or fruits (for example Theobromba, or 5 many members of the genus Licania: Prance, 1972). Likewise, wind dispersed species have both small seeds and a tendency to occur in the upper canopy strata where higher wind velocities aiding dispersal are greater (Hughes et al., 1994), one obvious example being the widespread neotropical species Jacaranda copaia (Jones et al., 2005). As was also found by Wright et al. (2006), the study gives little support for one of 10 "Corner's rules", viz. that due to their mutual dependence on the available supporting twig mass that leaf size and seed size should be positively correlated (Corner, 1949). There may be two reasons for this. First, as pointed out by Grubb et al. (2005) such biomechanical explanations would only be expected to apply where there is little flexibility in the number of fruits per inflorescence. Second, as for Φ LS (Fig. 5) the ratio of total 15 leaf area to the supporting stem mass is to a large degree independent of L A . Indeed, if anything, what our data suggest is that reproductive structures compete with leaves for available space as there is a nearly significant correlation between Φ LS and S (r 2 = −0.09,p = 0.07) with this negative relationship significant for the low fertility species (Supplement , Table 2). Thus, in contrast to vegetation types from 20 more xeric habits where leaf areas may be substantially constrained by hydraulic considerations, leaf area per unit available stem area or mass may actually be constrained by the requirements for simultaneous allocation of available carbohydrate to reproductive structures for most tropical forest trees. That being consistent with their tropical forest productivity being carbon limited as argued by Lloyd and Farquhar (2008). 25 Competition between foliage and developing fruit may also be the reason for the negative relationship between seed size and foliar [Ca] shown in Fig. 7, an observation also made for sub-tropical montane tree species by Easdale and Healey (2009). It has long been known that calcium is relatively immobile in plants (e.g., Kirby and , 1984) with high rates of calcium supply to developing fruit essential for cell wall development and for longer term maintenance of membrane integrity. Sufficient levels of calcium are also required to maintain the integrity of the fruit flesh including resistance to fungal attack even after abscissed from the plant (Bangerth, 1979). Due to its immobility, this calcium accumulation in fruit tissues must occur at the expense of 5 the leaves, and thus Fig. 7 does not necessarily imply that Ca itself may be limiting for either reproductive tissue development or leaf physiological function. Indeed, the SMA slope fit of −8.3 suggests that for each doubling of S foliar [Ca] declines by only about 10 %, a value roughly consistent with the similar [Ca] in both seed and leaf tissue (as evidenced from the seed data of Grubb and Coomes (1997)) and with about 0.1 of total 10 South American tropical forest "soft" litterfall occurring as reproductive organs (Chave et al., 2010). Even though such a result does not, therefore, necessarily imply an direct effect of Ca availability on tree function, it is interesting to note that species growing on extremely cation poor spodosols are characterised by relatively small seed masses as compared to more fertile nearby forests (Grubb and Coomes, 1997) as well as with leaf 15 photosynthetic rates showing an apparent dependence of leaf calcium concentrations (Reich et al., 1995). Moreover, for forests on such nutrient poor soils, carbon allocation to photosynthetic organs is apparently prioritised over that to reproduction (Chave et al., 2010), this being consistent with neotropical forest reproductive structure frequency being highly sensitive to soil fertility as inferred (apparently) from soil nitrogen status 20 (Gentry and Emmons, 1987), being markedly lower for forests growing on less fertile soils. Overall, these observations suggest, as also discussed in Sect. 4.1.2, that foliar and reproductive tissue development may be in direct competition for either carbon or available nutrients where soil fertility is low. 25 Although an examination of the various bivariate relationships, as discussed in Sect. 4.1 has hopefully proved informative, it is also of additional interest to quantify the extent to which all the various traits examined coordinate in their variability as a 5114 Introduction whole. In this respect, PCA was considered the most appropriate approach, as the first dimension of a PCA analysis can also be considered (with data normalisations prior to analysis as undertaken here) as the multivariate equivalent of an SMA model fit (Warton et al. 2006). Moreover, as long as it is rigorously established that subsequent axes remain orthogonal to each other (as we have done here for example 5 through the sphericity and residuals matrix analyses of Tables S4 and S6 in the Supplement as well as can be seen from the species scores for the first three axes in Fig. 8) then these the M A and [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the 3 population variance for low and high fertility species respectively) seems to be dominated by the 4 presence of large seeded members of the Leguminaceae whose importance in the phytogeography 5 of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote 6 this dimension as d TS .

7
The last eigenvector included in our analysis,, U 5 , differs from the others in having a sub- can be considered to represent distinct integrated plant functional dimensions, although sometimes having individual traits in common (Fig. 8d, Table 2). In our analysis here we have also taken the additional step of allowing for the covari-10 ance matrices to vary between low and high fertility sites through the CPC approach of Flury (1988), this effectively being a multivariate extension of the allowance in an SMA analysis for different groups to have differences in slope and/or elevation (Warton et al., 2006) with analysis of the "Flury hierarchy" in Table S3 (Supplement) strongly suggesting that the CPC model presents a significant improvement over models assuming 15 equality or strict proportionality between the two species groups. We thus interpret Table 2 as indicating five discrete integrated trait dimensions of tropical tree function and with the relative importance of these effects varying between high and low fertility species. This interpretation is made even though some of the measured properties such as M A and ρ x are modelled as having significant contributions to several dimen- 20 sions. This is argued as reasonable on two counts. First, variations in some of the traits measured may have different underlying causes. For example, changes in M A may be a consequence of variations in leaf thickness, tissue density or both (Witkowski and Lamont, 1991;Niinemets, 1999;Poorter et al., 2009) and likewise, variations in ρ x could reflect differences in the proportions of gas, air and dry matter content (for 25 hydrated tissue) in a wide range of combinations (Poorter, 2008 Bivariate relationships: Environmental components ering data from both low and high fertility sites together, Table 3 lists correlations

PDJ : Leaf structural costs and lifespan
Although it is often considered that the primary dimension of the leaf economic spectrum is that proposed by Wright et al. (2004) viz. systematic variations in rates of photosynthetic carbon acquisitions (dry weight basis) being linked with foliar dry-weight concentrations of nitrogen, phosphorus, M A and leaf longetivity, our analysis found that 5 U 1 (accounting for the greatest component of the total variation in the dataset) did not involve nitrogen at all, and was actually dominated by leaf cation concentrations and (of opposite sign) a low carbon content. We suggest that this dimension, 409 PDJ , reflects different plant strategies in terms of leaf construction costs, with the tendency for low M A in these leaves of high mineral content presumably attributable to a low tis-10 sue density associated with thinner, less lignified cell walls and with the higher cations content presumably also balanced by higher levels of organic acids (Poorter and de Jong, 1999) and with lower overall construction costs and less investment of phenols and other carbon rich compounds in defense (Poorter and Villar, 1997 reasonable to expect that, associated with lower levels of lignification and reduced tissue densities, would be relatively more flexible cell walls and a low bulk modulus of elasticity (Niinemets, 2001), also with the high cation concentrations, especially potassium making a substantial contribution (in association with organic acids) to leaf tissue osmotic potentials (Olivares and Medina, 1992

397
The last eigenvector included in our analysis,, U 5 , differs from the others in having a sub-398 RW is that usually considered to be the principal dimension of the leaf economic spectrum (Wright et al., 2004) Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | also been presented for tropical forest tree species (Sandquist and Cordell, 2007;Santiago and Wright, 2007;Fyllas et al., 2009;Baltzer and Thomas, 2010;Baraloto et al., 2010;Domingues et al., 2010). Although we did not measure the photosynthetic or respiratory components, our analysis does suggest that for tropical forest species, L A should also be included as part of population variance for low and high fertility species respectively) seems to be dominated by 394 presence of large seeded members of the Leguminaceae whose importance in the phytogeogra 395 of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore de 396 this dimension as d TS .

397
The last eigenvector included in our analysis,, U 5 , differs from the others in having a 398 stantially greater importance for low fertility versus high fertility species ( (Fig. 7) within this dimension. Such an involvement of L A in the classic resource acquisition/utilisation spectrum has also been suggested from a data analysis involving 29 sub-tropical montane tree species across 12 ha of permanent sample plots in Tucumán, Argentina (Easdale and Healey, 2009 Overall the five eigenvectors selected, all of which we believe to be physiologically relevant 407 RW ) were also reported for tropical forest leaves sampled across a range of soil substrates in French Guiana (Baraloto et al., 2010). They concluded, however, that L A was not closely linked with either [N] or [P]. This could be for several reasons. First, their sampling strategy covered a range of (undefined) soil 15 types and as discussed in Sect. 4.3, these are likely to have modulated foliar nutrient levels but not L A . Second, our sampling has covered a much wider range of environments and soils, presumably bringing wider species-level variation into the dataset as whole. Thirdly, our analysis shows the L A is also an important component of and M A having opposite signs (in contrast to d FW ) and with higher S and also being asso-401 ciated with a lower H max along with a less substantial but significant coefficient for ρ x . Also 402 RW was Φ LS , this being consistent with the general trend of Φ LS to increases with increasing L A (Fig. 6d). Especially as there was little contribution of to this dimension ( sapling stem data for a Bolivian forest presented by Poorter wt al. (2010) does indeed suggest an increase in the proportion of non-wood stem components occupied by the gas, (as oppossed to water fractions) as presence of large seeded members of the Leguminaceae whose importance in the phytogeo 395 of Amazon forest has already been recognised by ter Steege et al. (2006). We therefore 396 this dimension as d TS .

397
The last eigenvector included in our analysis,, Considering data from both low and high fertility sites together, Table 3 lists corre 411 RW increases (using leaf data from  with CO 2 assimilation rates (leaf mass basis) taken as a proxy for variance for low and high fertility species respectively) seems to be dominated by the large seeded members of the Leguminaceae whose importance in the phytogeography forest has already been recognised by ter Steege et al. (2006 Overall the five eigenvectors selected, all of which we believe to be physiologically relevant 407 (see Supplementary Information), accounted for 0.68 of the total variance for both low and 408 high fertility soil species.

409
RW . The former is, of course, well documented and, for woody plants at least, seems to be associated with an increased foliar tissue density 10 rather than changes in leaf thickness (Niinemets, 1999;Poorter et al., 2009) and with a concurrent reduction in photosynthetic nutrient efficiency when expressed on a dry weight basis (Niinemets, 1999;Domingues et al., 2010). One possibility to account for this is low internal conductances to CO 2 transfer for higher M A species (Lloyd et al., 1992;Syvertsen et al., 1995;Warren and Adams, 2006), as perhaps evidenced by a 15 small but significant positive contribution in to this dimension (0.014 ± 0.05: Table 2). Alternatively, relatively more nitrogen being allocated to cells walls of low also included L A in (d  (Onoda et al., 2004;Takashima et al., 2004), much of which would be expected to be in the form of defense related proteins (Feng et al., 2009) Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | a lesser extent ρ x . This linkage is most likely through the hydraulics/plant height considerations already discussed as part of Sect. 4.1.1 and 4.1.2. That is to say, as H max increases, a suite of trait adjustments occur; these including a reduction in Φ LS with estimates of also suggesting that leaves with a high H max also tend to operate at a lower c i /c a . As it seems likely that the higher M A with increasing H max is mostly at-5 tributable to increased leaf/mesophyll thickness and hence increases in photosynthetic capacity per unit leaf area, A max (Sect. 4.1.1), this reduction in c i /c a may be attributable to stomatal capacity increasing less with H max than should A max . Such a tendency to operate at a lower c i /c a would also help to conserve water for species more likely to be higher-up in the canopy and hence exposed to higher levels of insolation and an 10 associated greater evaporative demand .
Although H max was not determined in their study, many of the above measured and/or inferred traits, viz. Φ LS and A max , were found to co-vary in a similar manner as for The fourth component axis is dominated by S an Overall the five eigenvectors selected, all of which 407 FW across a range of tropical forest trees in Panama by Meinzer et al. (2008). Though in that case, variations in ρ x were considered of key importance in terms of trait co-15 ordination, especially through linkages to plant hydraulic parameters such as K S . Our observed contribution of ρ x is likewise significant (−0.22 ± 0.10), though as discussed in Sect. 4.1.1 taken across a wide range of species and sites the strong relationship between ρ x and/or ρ w and K S as observed by Meinzer et al. (2008) and also in some other studies (e.g., Santiago et al., 2004a) may not necessarily always apply. 20 Interestingly, in contrast to relative to M A is Φ LS . Although with a high estimated standard error as part of d The fourth component axis is dominated by S and Φ LS with these coefficients of different FW the tendency of potentially taller trees to have fewer but larger leaves than their more vertically challenged counterparts. But with a lower Φ LS overall. This lower Φ LS pre- 25 sumably serves to help maintain favourable water relations by counteracting greater resistances in the hydraulic pathway for potentially taller trees. Nevertheless, along with a higher , this lower Φ LS must also serve to reduce overall rates of whole tree carbon gain such as otherwise might be expected on the basis of higher A max and a greater probability of high levels of incoming radiation. This trade-off associated with a greater H max may be one reason for the observation that light demanding species with a low ρ w do not necessarily show higher above-ground growth rates than their more shade tolerant counterparts as reported by Keeling et al. (2008).

4.2.4
tion variance for low and high fertility species respectively) seems to be dominated by the e of large seeded members of the Leguminaceae whose importance in the phytogeography zon forest has already been recognised by ter Steege et al. (2006). We therefore denote ension as d TS .
last eigenvector included in our analysis,, U 5 , differs from the others in having a sublly greater importance for low fertility versus high fertility species (accounting for 0.09

of the population variances respectively). This component is characterised by H max
having opposite signs (in contrast to d FW ) and with higher S and also being assowith a lower H max along with a less substantial but significant coefficient for ρ x . Also ence in characterising U 5 are greater foliar [C] associated with the higher M A and .
gh, U 5 presents some traits combinations as reported previously in the literature, this ent, mostly related with species found at low fertility soils, it does not seem to have cognised before. It is thus here denoted as d PFL .
rall the five eigenvectors selected, all of which we believe to be physiologically relevant pplementary Information), accounted for 0.68 of the total variance for both low and rtility soil species.
TS : large seeds at the expense of leaf area 5 As mentioned in Sect. 4.1.3, a major factor in accounting for this trait dimension is the presence of many large seeded Fabaceae, especially on nutrient poor soils, for whom it turns out do not have as large a Φ LS as they would otherwise be expected to have on the basis of their other integrated trait values. Thus species with a high ted with the higher S are also lower [Ca] but higher foliar [P] and L A . With for their coefficients and higher standard errors, also being of different sign, are N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the riance for low and high fertility species respectively) seems to be dominated by the rge seeded members of the Leguminaceae whose importance in the phytogeography rest has already been recognised by ter Steege et al. (2006). We therefore denote n as d TS .
igenvector included in our analysis,, U 5 , differs from the others in having a subater importance for low fertility versus high fertility species (accounting for 0.09 he population variances respectively). This component is characterised by H max g opposite signs (in contrast to d FW ) and with higher S and also being assolower H max along with a less substantial but significant coefficient for ρ x . Also n characterising U 5 are greater foliar [C] associated with the higher M A and .
presents some traits combinations as reported previously in the literature, this ostly related with species found at low fertility soils, it does not seem to have ed before. It is thus here denoted as d PFL .
e five eigenvectors selected, all of which we believe to be physiologically relevant TS should best be regarded as those having a larger than average seed size with that 10 being associated with a lower than average Φ LS as compared to trees of an equivalent h component axis is dominated by S and Φ LS with these coefficients of different ted with the higher S are also lower [Ca] but higher foliar [P] and L A . With for their coefficients and higher standard errors, also being of different sign, are N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the riance for low and high fertility species respectively) seems to be dominated by the rge seeded members of the Leguminaceae whose importance in the phytogeography rest has already been recognised by ter Steege et al. (2006). We therefore denote n as d TS .
igenvector included in our analysis,, U 5 , differs from the others in having a subater importance for low fertility versus high fertility species (accounting for 0.09 he population variances respectively). This component is characterised by H max g opposite signs (in contrast to d FW ) and with higher S and also being assolower H max along with a less substantial but significant coefficient for ρ x . Also n characterising U 5 are greater foliar [C] associated with the higher M A and .
presents some traits combinations as reported previously in the literature, this ostly related with species found at low fertility soils, it does not seem to have ed before. It is thus here denoted as d PFL .

FW and/or
The fourth component axis is dominated by S and Φ LS with these coefficients of different n. Associated with the higher S are also lower [Ca] but higher foliar [P] and L A . With er values for their coefficients and higher standard errors, also being of different sign, are M A and [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the pulation variance for low and high fertility species respectively) seems to be dominated by the sence of large seeded members of the Leguminaceae whose importance in the phytogeography Amazon forest has already been recognised by ter Steege et al. (2006). We therefore denote s dimension as d TS .
The last eigenvector included in our analysis,, U 5 , differs from the others in having a subntially greater importance for low fertility versus high fertility species (accounting for 0.09 d 0.04 of the population variances respectively). This component is characterised by H max d M A having opposite signs (in contrast to d FW ) and with higher S and also being assoted with a lower H max along with a less substantial but significant coefficient for ρ x . Also influence in characterising U 5 are greater foliar [C] associated with the higher M A and .
though, U 5 presents some traits combinations as reported previously in the literature, this ponent, mostly related with species found at low fertility soils, it does not seem to have en recognised before. It is thus here denoted as d PFL .
RW . This lower Φ LS is also accompanied by reduction in L A suggesting that it is not so much competition for lateral meristems (Kleiman and Aarssen, 2007) that gives rise to the negative association between Φ LS and S within this dimension. But rather some sort of mechanical constraint such as the total mass capable of being 15 borne per unit stem weight (Westoby and Wright, 2003)  Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | leaf thickness) and associated increased leaf toughness (Kitajima and Poorter, 2010) and with high [C] linked through higher than average levels of more reduced structural compounds such as lignin as well as the typically high C-content defense related compounds such as tannins and phenols (Fine et al., 2006;Read and Stokes, 2006;Read et al., 2009). Also associated with this is a higher , which may be suggestive 5 of a greater internal resistance to CO 2 diffusion within the leaves of high tissue density woody species (Lloyd et al., 1992;Syvertsen et al., 1995;Warren and Adams, 2006). Interestingly, as well as these correlated leaf traits in this dimension there is the coordinated involvement of a lower H max . Species with strong weightings along this trait dimension are also characterised by larger seeds as would be expected for 10 shade adapted trees (Sect. 4.1.3). Along with a small but significant contribution of ρ x , ociated with the higher S are also lower [Ca] but higher foliar [P] and L A . With es for their coefficients and higher standard errors, also being of different sign, are d [N] terms. As mentioned in the Discussion, U 4 (accounting for 0.09 and 0.07 of the variance for low and high fertility species respectively) seems to be dominated by the f large seeded members of the Leguminaceae whose importance in the phytogeography forest has already been recognised by ter Steege et al. (2006). We therefore denote sion as d TS .
st eigenvector included in our analysis,, U 5 , differs from the others in having a subgreater importance for low fertility versus high fertility species (accounting for 0.09 f the population variances respectively). This component is characterised by H max aving opposite signs (in contrast to d FW ) and with higher S and also being assoh a lower H max along with a less substantial but significant coefficient for ρ x . Also e in characterising U 5 are greater foliar [C] associated with the higher M A and .
U 5 presents some traits combinations as reported previously in the literature, this t, mostly related with species found at low fertility soils, it does not seem to have nised before. It is thus here denoted as d PFL .
l the five eigenvectors selected, all of which we believe to be physiologically relevant PFL is thus strongly suggestive of a coordinated trait dimension associated with shade tolerance and longevity. Not surprisingly then, it seems to play a much greater role in accounting for the trait variations of species associated with low fertility as opposed to high fertility soils as indicated by the different values for the characteristic 15 roots (λ low = 698, λ high = 318, Table 2).

Significance of integrated trait dimensions and their components
Although it is axiomatic that to be included in any of the above integrated dimensions that a trait would have had to be actually measured, what is perhaps more subtle, is that the mix of suites of traits coming together on any one PCA (or CPC) axis is also 20 dependent on what is not measured. For example, our differentiation of the first two components of the CPC analysis of RW it was also found to be part of lation variance for low and high fertility species respectively) seems to be dominated by the ence of large seeded members of the Leguminaceae whose importance in the phytogeography mazon forest has already been recognised by ter Steege et al. (2006). We therefore denote dimension as d TS .
he last eigenvector included in our analysis,, U 5 , differs from the others in having a subtially greater importance for low fertility versus high fertility species ( 409 RW would also have appeared more equivocal (and potentially missed) as confounding species potential height effects on L A would have muddied it's role in this classic leaf resource acquisition dimension (see for example Baraloto et al., 2010). 10 It is thus clear, that in the presence of additional parameter measurements (for example direct determination of K S ) our derived dimensions may well have been different. Nevertheless, as discussed above, all five Considering data from both low and high fertility sites together, T 411 RW , then a negative relationship between M A and L A would be expected to be observed as, for example, was found to be the case for a range of herbaceous angiosperms by Shipley (1995). Or, as 5 found in some cases by Pickup et al. (2005)  PDJ is be the primary source of variation in the latter (as L A is effectively absent from this dimension). Indeed, although much touted as a fundamental plant trait (e.g., Poorter et al., 2009;Asner et al., 2011;Kattge et al., 2011) M A seems to us to be too confounded a measurement to be practically useful 10 in differentiating different plant growth strategies as evidenced by its contribution to the five ociated with the higher S are also lower [Ca]  l the five eigenvectors selected, all of which we believe to be physiologically relevant above and future work would be better directed towards separate measurements of foliar tissue density and thickness as well as leaf dry matter content (Witkowski and Lamont, 1991;Wilson et al., 1999). It is probably because of its ambiguous nature that M A does not seem to be as good a predictor of demographic rates as first thought, 15 especially when comparisons are done across different sites .
Our results give no support for the supposed "second dimension" of the leaf economics spectrum proposed by Baltzer and Thomas (2010). That study, primarily based on data from Bornean forest trees did, however, fail to differentiate between genetic versus soil effects on foliar properties as has been done here. And with their "second 20 dimension" (hardly likely to be orthogonal to the first dimension in any case) most likely simply reflecting soil fertility effects on foliar [P] as already well documented by Fyllas et al. (2009) and considered further below.

Coordinated trait responses to environmental variability
As evidenced by the 0.3-0.4 portion of the total variance associated with the Φ LS , 25 ρ x and "plot effect" terms ( Fig. 4)  with a negative weighting) in the first environmental PCA axis, ů 1 (Table 4), which was itself closely correlated with a PCA of soil chemical and physical properties (Fyllas et al., 2009;. This suggests a coordinated plant physiological and structural response to changes in soil fertility, with the concurrent increases in foliar 5 cations and phosphorus not that surprising as their availability within soils tends to decline more or less in concert as part of the soil weathering process . Such a decline is not necessarily the case for N, however  discernible effect on ρ x . As mentioned in Sect. 4.1.1, such a fertility effect on ρ x has been seen before as mediated by soil phosphorus availability for eucalypt and mangrove (Thomas et al., 2005;Lovelock et al., 2006). Although working with Brazilian savanna trees, Bucci et al. (2006) found it was nitrogen (as opposed to phosphorus) fertitlisation that induced changes in ρ x and K S with in their case with N-fertilisation 20 causing attendant increases in Φ LS not detected here (Table 4). It seems likely that higher foliar [P], especially in combination with the lower M A also associated with ů 1 would give rise to higher photosynthetic rates on an area basis (Domingues et al., 2010;Mercado et al., 2011). Thus, with tropical forest tree hydraulics and photosynthetic capacity being closely linked (Brodribb and Field, 2000;25 Brodribb et al., 2002;Santiago et al., 2004a) the likely increase in K S accompanying a decrease in ρ x with improved nutrient status may serve to help maintain some homeostasis in leaf water relations, this offsetting the higher rates of water-use per leaf area that would be expected to accompany any increase in ů 1 . This suggestion supported by the only modest contribution of to this dimension (Table 4). As to how such a coordination could occur is currently not clear, although the greater rates of cambial activity in the wood of higher P status trees giving rise to a lower ρ x might be attributable through sugar signalling mechanisms (Rolland et al., 2006;Hölttä et al., 2010), this resulting in less secondary thickening of vessels walls and a higher conduit 5 area ( Thomas et al., 2005). Other elements may also be involved though, for example effects of calcium and/or potassium on sapwood cambial activity (Fromm, 2010). The second integrated environmental response dimension identified, ů 2 , essentially represents an integration of previous observed foliar trait responses to precipitation, viz. increased M A , [C] and and decreased [Mg] as mean annual precipitation in-10 creases as detailed in Fyllas et al. (2009). Although this response to P A seems at odds with the general observation from inter-species analyses that leaves of more arid environments should have a higher M A and often with a higher (Miller et al., 2001;Santiago et al., 2004b) as discussed by Fyllas et al. (2009) this tendency towards more structurally rigid leaves at higher P A may be an intra-specific response to increased dis-15 ease pressure, possibly even as a result of the development of phenoptypically distinct provenances under different precipitation regimes -as seems to be the case for Costa Rican populations of Cordia alliodora in terms of susceptibilities to soil water deficit induced cavitation (but not other hydraulic properties: Choat et al., 2007). An aligned interpretation is that as severe dry season water deficits become increasingly less of 20 a driving force in determining leaf lifetimes, leaves of any given species become more "evergreen" in their structural characteristics. And indeed it is worth noting that the distinction between evergreen and deciduous phenologies for tropical forest trees is a somewhat arbitrary one (Brodribb and Holbrook, 2005;Williams et al., 2008). In such an interpretation, an increase in with P A could be interpreted as either a tendency 25 towards more conservative stomatal behavior in evergreen species where the precipitation regime is not strongly seasonal (Lloyd and Farquhar, 1994) or, alternatively to an increased resistance to CO 2 diffusion within higher M A leaves due to a higher cell wall resistance (Syvertsen et al., 1995). Although not emerging as any sort of integrated response through the PCA analysis of the derived environmental effects, the temperature responses of [N], [K] and ρ x are all also of note; these have already been considered separately by Fyllas et al. (2009) and Patiño et al. (2009).

5
Extending beyond a simple bivariate analysis approach, this study has separated environmental from genetic effects for a range of structural and physiological traits for Amazon forest trees then using Common Principal Component Analysis to reveal as many as five discrete integrated axes of genetic variation. The relative weightings of the axes varies between low and high fertility soil associated species. The first com-10 ponent (accounting for the highest proportion of the total variance in the dataset) was not the classic "leaf economic spectrum", but rather to relate mostly to variations in leaf construction costs. The leaf economic spectrum was the second most important dimension identified in terms of variance accounted with our results suggesting that it also involves differences in leaf size as well as in leaf area: sapwood area ratios. Our 15 third dimension brings together several structural traits, including species specific maximum height, individual leaf areas, leaf mass per unit area and xylem density and leaf magnesium concentrations. The fourth and fifth dimensions were interpreted as relating to a seed size/leaf area trade-off and shade tolerance characteristics respectively.
Several traits, in particular leaf mass per unit area, foliar carbon content and xylem 20 density had significant weighting on many axes of variation, this being attributed to their somewhat ambiguous "proxy" nature for a range of underlying and more fundamental plant physiological properties. In particular, variations in twig xylem density may arise as a consequence of differences in a range of different underlying phenomena and with its generally poor correlation with other plant traits suggesting that it may not be 25 as good a proxy for plant hydraulic conductivity as once thought. Significant effects of environment on many plant traits were also identified. Some of these integrated into discrete dimensions of variation and with discrete but different changes being associated with variations in soil fertility versus differences in mean annual precipitation. Whether these differences relate to strict "environmental effects" or reflect systematic patterns in intra-specific trait variation with soils and/or climate 5 remains to be established.
Acknowledgements. We thank our many South American collaborators for help with logistics 10 and practical help in the field. Much of the data collection phase of this work was supported through the EU FP6 "LBA-Carbonsink" project with subsequent data analyses supported by funding through the UK National Environment Research Council "QUERCC" and "TROBIT" consortia. We also thank David Warton (University of New South Wales) for pointing out to us the possibilities of CPC analysis and Patrick Phillips (University of Oregon) for making his 15 CPC model estimation and evaluation programs freely available. Shiela Lloyd assisted with manuscript preparation. Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Table 2. Common principal component analysis of derived genetic effects for species associated with low and high fertility soils. Values in brackets represent standard errors for each component. Coefficients given in bold are either those whose absolute values are 0.50 or more, or 0.30 or more with a standard error or less than 0.1. M A = leaf mass per unit area; elemental concentrations are on a dry weight basis, L A = leaf area; Φ LS = leaf area:sapwood area ratio, ρ x = branch xylem density, = diffusion limitation index (see Eq. 1), S = seed mass, H max = species maximum height.   Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Table 4. Summary of the Principal Components Analysis of the correlation matrix for the derived environmental/soil effects on observed structural and physiological traits. Coefficients given in bold are those whose values are 0.3 or more. M A = leaf mass per unit area; elemental concentrations are on a dry weight basis, L A = leaf area; Φ LS = leaf area: sapwood area ratio, ρ x = branch xylem density, = diffusion limitation index (see Eq. 1). Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Table 5. Kendalls partial correlation coefficient, τ P , for the environmental contribution (plot effect estimate) of each foliar property with the set of environmental predictors with their significance is computed based on Maghsoodloo and Laszlo Pallos (1981). Bold values indicate a very strong correlation (p < 0.001) and italics indicate significant correlations at p < 0.01; see text for details. M A = leaf mass per unit area; elemental concentrations are on a dry weight basis, L A = leaf area; A = leaflet area, Φ LS = leaf area/sapwood area ratio, ρ x = branch xylem density, = diffusion limitation index (see Eq. 1) and ů 1 and ů 2 are the first two principal components of the PCA analysis on the environmental effects correlation matrix (See Table 4).  Table 4 we were interested to see if any of the corresponding plot axes scores with previously derived soil and/or climate characteristics of the same sample plots.
ignificant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů1  Table 4 also suggesting that for any given species, both also decline with increasing precipitation and, somewhat counter intuitively, with g.
as in Fyllas et al. (2009) we show values for Kendall's partial τ (denoted τp) for f interest as well as ů1 and ů2 as functions of ff, ft, Ta, Pa and Qa in Table 5.  Table 4 we were interested to see if any of the corresponding plot axes scores with previously derived soil and/or climate characteristics of the same sample plots. significant relationships are shown in Fig. 9. First, the top panel of Fig. 9 shows ů1 on of the first soil PCA axis of Fyllas et al. (2009), the latter considered a strong measure of soil fertility and denoted fF. The strong relationship observed suggests tegrated response of Amazon tropical forest trees to soil fertility, with most nutrients and with foliar [C] and ρx decreasing as fF increases. Interestingly, the Kendall's τ t of ů1 versus fF of 0.63 is greater than for any of the original variables examined et al. (2009), the highest of which was 0.56 for foliar [P]. Comparison with Fyllas 9) also shows that the ů2 contains significant weightings of leaf-level variables that, y, were all strongly correlated with mean annual precipitation (PA) viz. positive s with foliar [C] and MA and a negative correlation with foliar [Mg]. It is therefore sing, as is shown in the middle panel of Fig 8, that ů2 and PA also show strong , but with examination of Table 4 also suggesting that for any given species, both w also decline with increasing precipitation and, somewhat counter intuitively, with g.
, as in Fyllas et al. (2009) we show values for Kendall's partial τ (denoted τp) for f interest as well as ů1 and ů2 as functions of ff, ft, Ta, Pa and Qa in Table 5. , (m 2 ), leaf area:sapwood area ratio (Φ LS ; cm 2 m −2 ), branch xylem density (ρ x ; kg m −3 ), = stomatal limitation index (dimensionless; see Eq. 1), species maximum height (H max ; m) and seed mass (S; g). Open red bars represent low and blue dashed bars high soil fertility plots, as defined by the quantitative determinations of the level of total reserve bases from 0.0-0.3 m depth (Fyllas et al., 2009;. Also given for each histogram are the mean and the variance for each trait. Significant differences in mean values and/or variances between the two fertility groups were identified with Fligner-Killeen test respectively. Significance codes: *** < 0.001, ** < 0.01,* < 0.05. 8,2011 Tropical tree trait dimensions S. Patiño et al.  2. Partitioning of the total variance for each structural property into genetic (family/genus/species), environmental (plot) and an error (residual) components. Foliar properties are sorted from less to more phylogenetically constrained. Significance of each variance component was tested with a likelihood ratio test (Galwey, 2006). Significance codes: *** < 0.001, ** < 0.01,* < 0.05. Introduction and leaf mass per unit area (L A ) for the top two panels and between species estimated foliar [Ca] associated average seed mass (S) for the associated genus. Open circles indicate species found on low fertility sites and the closed circles indicate species found on high fertility sites. Species found on both soil fertility groups are designated by a "+" (see text for details