Study site

The study was conducted in the 25 km long Guisane valley (c. 260 km²; 44.9° N, 6.6° E) in the central French Alps (between Briançon at 1200 m and Combe Roche Noire at 3000 m a.s.l.). The valley is characterized by strongly contrasted climatic conditions, with monthly mean temperatures ranging from −6.3 °C to −1 °C for the coldest and from 10 °C to 17 °C for the warmest months. As in other valleys of the central Alps, the landscape is a mosaic of coniferous and deciduous forests, shrub heaths, subalpine grasslands and alpine meadows, which were all included in the sampling strategy.

Species selection

We sampled, during summer 2007 and 2008, 16 species (see Table 1) (i) from different life forms (graminoïd, forbs, legume, shrub, tree) to obtain the widest possible scope for the robustness and generality of detected patterns; (ii) that were sufficiently common, i.e. widely distributed, to be found in contrasting conditions; (iii) with contrasting climatic preferences, i.e. species for which optimal conditions are lowland, subalpine or alpine, to be able to detect potential functional changes between marginal and optimal populations (Hampe & Petit 2005). We used two sets of congeners (Juniperus: sibirica and communis and Pinus: sylvestris and uncinata) because of their apparent hybridization in the study area (P. Salomez, personal communication). We checked post hoc that these congeners were relevant functional continua, i.e. had unimodal trait distributions. We calculated for each species the theta index to estimate its degree of specialization (Fridley et al. 2007). This species co-occurrence index assumes that generalist species co-occur with a larger number of species than specialists (see Appendix S1).

Table 1.   Characteristics of the sixteen species sampled. For all sampled species are given: name (species were sampled in 2007, except species with an asterisk, which were sampled in 2008), label, family, life form (G: grass, H: herb, S: shrub, T: tree), the number of populations where they were sampled, the number of measured individuals, the number of measured leaves per individual plant, the way the Leaf Nitrogen Content (LNC) was sampled (SP: samples pooled by subpopulations, I: measured individually). ‘Niche1′ represents a species’ altitudinal range [considered as the 10–90% inter-quantile length from the presence points observed by the Conservatoire National Botanique des Alpes (CBNA)], ‘Niche2′ represents the optimal altitude of the species (mode of the distribution curve of presence along altitude, data from CBNA), ‘Sampling’ represents the altitudinal range sampled in the study and ‘Spe’ represents species’ Fridley’s specialization index (see Appendix S1). ‘Environmental requirements and sampling’ describes the altitudinal range of the species between 0 and 4000 m a.s.l.: (i) white rectangles give the full altitudinal range from the presence points observed by the CBNA; (ii) hatched rectangles give the inter-quantiles (10–90%) of the altitudinal range and (3) dark rectangles give the altitudinal range we sampled

Sampling strategy

Based on the results of a principal component analysis (data not shown) on a set of interpolated topoclimatic variables from the meteorological model Aurelhy at a 50-m resolution (Benichou & Le Breton 1987), we selected two uncorrelated direct gradients: mean minimal temperature in winter (temperature) and solar radiation in August (radiation). We hypothesized that functional traits would respond to these orthogonal gradients either linearly (in case of a strong limiting gradient) or following a bell-shaped curve (with an optimum trait expression along the gradient). These traits could then be expressed using the following equation with ‘Temp’ for the Temperature and ‘Rad’ for the Radiation:

Then, following a hierarchical stratified sampling design based on a surface response methodology (Box & Draper 1959), at least nine samples were required to parameterize the model and account for environmental heterogeneity. To select populations for each species throughout the whole valley, we applied the environmental stratification presented above to the known occurrences of this species in the valley (data set from the French National Alpine Botanical Conservatory).

To determine the spatial structure of the intraspecific FV, we further stratified the sampling design. For each species, we selected three subpopulations within each population (50 cm× 50 cm or 1 m× 1 m for herbaceous species, and 10 m× 10 m for shrubs and trees), and in each subpopulation three random individuals (tussocks or ramets) were measured (Table 1). Subpopulations were scaled to the individual size in order to make them (i) contain at least three distinct individuals and (ii) be sufficiently homogeneous, so that differences between subpopulations integrate environmental micro-heterogeneity. For leaf trait measurement, we collected one leaf per individual for herbaceous species and 10 for shrubs and trees (see Table 1); trait values were averaged at the individual level for the analyses. Only sexually mature plants and unshaded individuals were sampled (except in woodlands). We collected non-senescent, non-grazed, non-frozen, well-developed and sun-exposed leaves. We tracked the growing season according to altitude, aspect and field observations to sample all individuals of a species at a similar phenological stage.

Trait selection and measurement

We measured three functional traits to characterize the ecological strategies of the studied species. (i) Maximum vegetative height (H_max, not for trees), i.e. the distance between the top of the photosynthetic tissue and the ground, is associated with plant competition vigour and tends to be allometrically correlated with above-ground biomass (Cornelissen et al. 2003). (ii) Leaf dry matter content (LDMC), the dry mass of a leaf divided by its water-saturated fresh mass (Cornelissen et al. 2003) expressed in mg g⁻¹, reflects plant growth rate and carbon assimilation. Leaf dry matter content is considered to be a robust trait (Roche, Diaz-Burlinson & Gachet 2004) and is usually negatively correlated with relative growth rate (Weiher et al. 1999). Leaf dry matter content was measured with the partial-rehydration method, which has been demonstrated to not produce significant biases as compared to the more intensive full-rehydration method (Vaieretti et al. 2007). (iii) Leaf nitrogen content (LNC), the total nitrogen per unit of dry leaf mass (in mg g⁻¹), is closely linked to the mass-based maximum photosynthetic rate (Cornelissen et al. 2003). Dried and marble-ground leaf samples of 3–5 mg were analysed with FlashEA 1112 elemental analyzer (Thermo Fisher Scientific Inc., Milan, Italy) at the individual or at the subpopulation-level for species with small leaves (Table 1).

To validate our data and compare trait values with existing literature, we extracted H_max, LDMC and LNC from several available data bases (Fig. 1: own data and see Table S1 in Appendix S1).

Statistical analysis

We used linear mixed models to represent the hierarchical sampling design (Pinheiro & Bates 2000) and decompose the trait variation of each species. We started with intercept random models (m₀) with only random effects (population, subpopulation nested into population) to decompose the variability of trait values (H_max, LNC and LDMC) at the different levels of sampling. We then built more complex mixed models with linear and quadratic fixed effects of the gradients (m₁). We checked the normality of the predicted random effects and residuals, as well as for stability of the estimates obtained from different models. To determine what part of the variability at the population-level was due to the fixed effects, we calculated a measure of explained variation based on the variances at the population-level (σ²) for the different models (Xu 2003):

where σ_m0 (resp. σ_m1) is the estimated error standard deviation at population-level estimated under model m₀ (resp. m₁). Variances were estimated by maximizing the restricted log-likelihood (REML).

We further tested if some emerging patterns in intraspecific FV could be related to species’ characteristics such as growth form (‘Life Form’), niche breadth (‘Niche1′), optimal altitude (‘Niche2′), sampling range (‘Sampling’) or specialisation (‘Spe’) (Table 1). We tested the effects of these characteristics on response variables (Y) with linear models built as follows: Y∼‘Characteristics’. To test the existence between species of general patterns in: (i) trait variability, we used for Y the coefficient of variation (CV) of the trait for each species (Lande 1977); (ii) decomposition of trait variance between the different sampling levels, we used for Y the percentage of variance at the population-level; (iii) the percentage of trait variation explained by gradients, we used for Y the maximal value of R² obtained for the species.

Statistical analyses were carried out with R 2.7.0 (R Development Core Team 2008) using the packages nlme, stats, ade4 and vioplot.

Quantification of intraspecific functional variability

A comparison of the altitudinal range observed for the selected species in the French Alps and our sampling showed that the sampling strategy was rather efficient and that a large part, or at least the upper part, of the species’ ranges was sampled (Table 1). We found large interspecific FV across the measured traits (Fig. 1). Species mean values ranged from 2.4 to 44 cm for H_max (Fig. 1a, without Juniperus), from 142 to 518 mg g⁻¹ for LDMC (Fig. 1b) and from 9.3 to 33.1 mg g⁻¹ for LNC (Fig. 1c). Each species showed large intraspecific FV. For example, the mean LDMC of Sesleria caerulea was 363 mg g⁻¹ but it ranged between 200 and 720 mg g⁻¹. H_max was particularly variable with coefficients of variation (CV) for each species running from 0.19 to 0.49 while LDMC and LNC tended to be less variable (CV in 0.08–0.25 and 0.09–0.29, respectively, Fig. 1). This large variability within each species resulted from both between- and within-population-levels. The case of Dactylis glomerata is given as an example in the left corner of each sub-figure in Fig. 1. Study species did not have similar magnitudes of variability and no clear trend emerged across species types (Life Forms, Niche1, Niche2, Sampling, Specialization index, see Table 1). Moreover, a given species could be highly variable for one trait and not for another (Trifolium alpinum had a CV of 0.46 for H_max and of 0.09 for LNC). Comparison of the measured values with the literature and existing data bases (Fig. 1 and Table S1 in Appendix S1) showed that they are consistent with former measurements.

Attribution of the sources of variation

Variance decomposition

The sources of variation varied notably across traits and species (Fig. 2). For most species, these sources varied between the three traits; for example 50% of the variance in H_max of T. alpinum was due to differences between populations, but only 12% for LNC and LDMC. On average, variability between populations represented a large part of the variance (with individual data, on average 47% for H_max and 35% for LDMC). There was also high individual variability, reaching as much as 60–80% (LDMC of Geum montanum, Leucanthemum vulgare, T. alpinum), and 53% on average for LDMC with individual data. In contrast, only a small part of the variance was due to subpopulations (with individual data, on average 10% for H_max and 12% for LDMC).

Analysing the structure of variability within data from literature and data bases led to results comparable to ours; however, existing data were not precise enough to achieve the same complete quantification. In particular, very few data were available for LNC and at the within-population-level (Tables S1 and S2 in Appendix S1).

Effects of environmental variables

All the calibrated models converged and most of them showed significant effects of the fixed factors. Most trait response shapes were quadratic with strong interactions between both temperature and radiation gradients. The models’ results remain, however, a complex patchwork with response curves and variation explained by models differing between species and between traits (Fig. 3 & Table S3 in Appendix S1). Overall, variations in H_max were better explained by gradients (on average 30% of the variance at the population-level explained) than variations in LNC (on average 21%), with LDMC (26%) being intermediate. Again, there were however strong differences between traits and between species (Table S3 in Appendix S1).

A large variability and a complex hierarchical structure but no general pattern

The highly structured sampling design allowed us to answer our four key questions about the structure of intraspecific FV. (i) Trait values were highly variable within species, but this variability depended on both traits and species, suggesting that the behaviour (response to and effect on the environment) of individuals or populations of a species is not fixed and might differ depending on the conditions. (ii) A large part of the variability came from differences between populations. Moreover, the variability observed within populations was not spatially structured. In selecting random individuals in small (scaled to the individual size), homogeneous subpopulations, we assumed first, that the differences between subpopulations within a given population would integrate environmental micro-heterogeneity, and second, that individuals within a subpopulation may be genetically related, although it was beyond the scope of this study to check this assumption. We thus expected large differences between subpopulations and smaller differences within them, but we actually found the opposite with a strong individual variability within subpopulations. This indicates that the subpopulations probably did not integrate either environmental heterogeneity or genetic structure, leaving a large part of the variability unexplained. As our field measurements were not designed to tease apart local adaptation (the observed local variability would be due to the presence of several genotypes within populations) and phenotypic plasticity (the observed local variability would be due to the plasticity of given genotypes), which are both integrated by measures of phenotypic variability (Byars, Papst & Hoffmann 2007), more studies are needed to research causes for the idiosyncratic variability that we highlighted. Similar results have been obtained on other measured traits (e.g. specific leaf area, leaf carbon content, number of inflorescences, green biomass); however, as these traits were not available for all investigated species, we focused this paper on three important and available traits. (iii) We showed that the shapes and strengths of trait responses to environmental gradients are idiosyncratic. Depending on species and traits, climatic gradients explained more or less of the variability between populations. Moreover, including more local factors (at population-level) like soil (pH, organic matter content, texture), disturbance (grazing and mowing) or competition (vegetation cover) significantly increased the variability explained in models (the explained part of variance increased on average from 31% to 59% for H_max, from 22% to 51% for LDMC and from 17% to 33% for LNC), but did not increase generality in patterns. (iv) Although we conducted this study in a very environmentally heterogeneous area and with contrasting species (different life forms, environmental requirements and degrees of specialization), we did not detect any pattern in terms of variability, sources of variation or response to gradients. This absence of patterns thus suggests that species are affected by the environment in idiosyncratic ways (Hultine & Marshall 2000) and corroborates the results of Gerdol (2005), who showed that two related species from the same life form (deciduous dwarf shrubs, Vaccinium myrtillus and V. uliginosum) can have very different growth performance (net primary production) and nutrient concentrations (N and P) along environmental gradients.

Traits and gradients: a long story, no happy end?

Although we specifically attempted to sample trait values to identify their response shapes along environmental gradients, we did not obtain clearer patterns than previous studies. The apparent complexity and inconsistency of the curves obtained is then not due to a problem of gradient selection and comparability, or of sampling methodology. Nevertheless, traits are expected to respond to environmental gradients (Cornelissen et al. 2003) and patterns have been found at the interspecific level at broad spatial scales along resource, disturbance or environmental gradients (Reich et al. 1999; Wright et al. 2004).

We propose a conceptual model (Figs 4 and 5) to disentangle the apparent complexity of our results and of data sourced from the literature and data bases. We assume that within species, traits could respond in a generic way, e.g. following a bell-shaped curve along an environmental gradient, but with parameters differing between species, e.g. maximum trait value at the species’ environmental optimum on the gradient (Fig. 4). Then at the intraspecific level, when studying the lower end of the relevant gradient, one should find a positive relationship between the trait value and the gradient, while at the upper end a negative relationship is expected, and at an intermediate gradient position either a bell-shaped or a flat relationship is expected (Fig. 4a). Therefore, depending on the part of the gradient sampled and the environmental requirements (range and optimum) of the different studied species, the observed relationships can be very different and inconsistent along the gradient, appearing as idiosyncratic (Fig. 4b). However, when the observed curves are combined in terms of species range (in the virtual example: lower part for species 3, higher part for species 1 and intermediate part for species 2), they lead to a bell-shaped curve, i.e. a generic intraspecific response shape along the gradient (as in Fig. 4a). Whereas, when the observed curves are combined along the whole environmental gradient, they lead to an interspecific pattern that can be very different from the generic intraspecific one (as in Fig. 4b).

Following this conceptual model, we can further interpret our results. We obtained a unimodal response for H_max along the temperature gradient, reaching a maximum for intermediate temperature values corresponding to more suitable conditions for the species (Figs 3 and 5ai). For some species (D. glomerata, S. caerulea), H_max might continue to increase with temperature because of strong light competition (Woodward 1986). At the interspecific level, H_max increased with temperature (R²=17%) and there was no relationship with radiation (Fig. 5bi). For LDMC, the conceptual reconstruction leads to an inverse bell-shaped response, showing lower LDMC indicating a higher growth rate for herbaceous species (Weiher et al. 1999) in more suitable places at the core of their distribution (Figs 3 and 5aii). We found, for example, a decreasing LDMC with increasing temperature for S. caerulea which was measured at the upper part of its distribution. LDMC could also increase or follow a bell-shaped response to radiation, consistent with Shipley (2003). At the interspecific level, LDMC has been shown to be either decreasing (Ryser & Aeschlimann 1999) or increasing (Roche, Diaz-Burlinson & Gachet 2004) with altitude, whereas we found no correlation between LDMC and the two studied gradients (Fig. 5bii). At the intraspecific level, LNC decreased along the temperature gradient with a clearer pattern for nitrogen content per area (mmol m⁻²) than for nitrogen content per dry mass (mg g⁻¹, Figs 3 and 5aiii), corroborating the ecophysiological literature at the species or life form level (Woodward 1986; Körner 1989; Cordell et al. 1998). Nitrogen has been extensively investigated along gradients because it is a key element for a variety of metabolic functions in plants, linked to their photosynthetic capacity. At the interspecific level, LNC (mg g⁻¹) decreased (R²=10%) along the temperature gradient (Figs 5biii and S1 in Appendix S1), contrary to what has been found at larger scales (Craine & Lee 2003).

Thus, pooling species together masks more complex patterns that exist at the intraspecific level and can even change the direction of the relationship, as conceptualized by Shea & Chesson (2002) for the relationship between invasion success and species richness. By using species range and environmental requirements as key parameters in trait-environmental gradient relationships, our conceptual model manages to reconcile the different results from our study and the literature. As a way (i) to limit the shortcomings resulting from gradient selection and sampling range and (ii) to reconcile the general and the species-specific aspects of trait–gradient linkages, we suggest rethinking them as trait–niche relationships; with the niche defined as a synthetic index of species environmental requirements (see Fig. 4). However, methods for evaluating this niche are beyond the scope of this study (Wright et al. 2006; Elmendorf & Moore 2008; Thuiller et al. 2010).

Consequences for the future use of functional traits

This study offers a better understanding of intraspecific FV in establishing an original framework for quantifying its extent, structure and sources. Such a framework is a preliminary requisite to future analyses aiming to test whether taking this variability into account is compulsory or not and what are its implications in diverse fields of ecological research. First, we needed to advance our knowledge on the extent of intraspecific FV. Showing a large variability of functional traits within species, this study challenges the use of mean values of functional traits to describe species in single species studies, like the response of a species to environmental changes or to herbivory, or the effect of a species on community dynamics or on nutrient cycling. The behaviour of individuals and populations will likely depend on the location of the study along environmental gradients and on other unknown factors (the unexplained part of the variability). Species mean traits should thus be replaced by populations mean traits (in each environmental conditions) or by a distribution of traits for more realism. This requires, however, that descriptions of trait variances as well as means are given in the trait data bases (Weiher et al. 1999; Garnier et al. 2001). Second, concerning multi-species studies, our results suggest that further analyses are required to test the relative influence of intraspecific versus interspecific variability.