Impacts of trait variation through observed trait-climate relationships on performance of a representative Earth System model: a conceptual analysis

Introduction Conclusions References


Introduction
Terrestrial vegetation plays a pivotal role in land-atmosphere interactions, modifying carbon, water and heat fluxes via biochemical processes such as photosynthesis and respiration or via biophysical vegetation properties such as stomatal conductance and albedo. Therefore, a correct representation of terrestrial vegetation and its dynamics in Earth System Models (ESMs) is essential, especially for future climate projections. For One of these issues is the way in which Plant Functional Types (PFTs) are used to represent vegetation. PFTs are classes of plant species with presumably similar roles in an ecosystem, responding in a comparable manner to environmental conditions like water and nutrient availability (Harrison et al., 2010;Lavorel et al., 1997). They are defined by a combination of attributes such as plant growth form (e.g. trees, shrubs, grasses, herbs), phenology (evergreen, raingreen, summergreen) and bioclimatic tolerances (e.g. minimum temperature requirements). Notably, most current PFT classifications use constant parameter values for some key plant traits, i.e. plant properties that reflect the way plants cope with their environment (McGill et al., 2006;Violle et al., 2007). Using constant plant traits in PFTs has serious limitations (Harrison et al., 2010;15 Ordoñez et al., 2009;Van Bodegom et al., 2012), as it does not reflect the trait variation observed within and between communities (Ackerly and Cornwell, 2007;Freschet et al., 2011;Kooyman et al., 2010;Westoby et al., 2002), therefore not accounting for local environmental constraints. Furthermore, global trait database analyses have shown that the variation in plant traits is large within PFTs and often even greater than parameter values per PFT for each individual grid cell, based on a systematic analysis of observed trait-climate relationships.
These relationships were implemented in the JSBACH DGVM, which is part of the Max Planck Institute Earth System Model (MPI-ESM) (Brovkin et al., 2009;Raddatz et al., 2007;Roeckner et al., 2003). It is representative of most DGVMs currently used 5 in the context of ESMs and used in carbon cycle model intercomparisons (Friedlingstein et al., 2006). We simulate variation in three originally constant and PFT-specific key leaf traits in JSBACH: specific leaf area (SLA), maximum carboxylation rate at a reference temperature of 25 • C (Vcmax 25 ) and maximum electron transport rate at 25 • C (Jmax 25 ). To determine relationships between these three traits and climatic drivers for 10 each PFT, data from the TRY global plant trait database (Kattge et al., 2011) were related to global climatic data. Based on these relationships between traits and climate, SLA, Vcmax 25 and Jmax 25 were re-parameterized for each grid cell on a yearly basis, depending on the (local) climatic conditions in a grid cell. This enabled feedbacks between plants and environment, as traits within natural PFTs could vary dynamically in 15 space and time. A simulation with variable traits is compared for trait distribution, productivity and vegetation distribution to a default simulation with the original trait values of the model, and to an additional simulation with constant, but observation-based, trait values.
As DGVMs are already parameterized to produce approximate realistic results, our 20 variable traits simulation will not necessarily approach reality better than the default simulation. Therefore, in this paper the focus lies on simulation intercomparisons to evaluate the impact of incorporating climate-driven trait variation. A formal validation with observational data is not the primary aim of this paper, as the simulations are not meant to represent current vegetation or climate states, but instead are run into equilib- 25 rium states with their own internal climate and vegetation composition in order to study the impacts of trait variation within PFTs. Nevertheless, we include comparisons with global biomass and gross primary productivity (GPP) estimates and potential vegetation distribution to evaluate the realism of the various simulation results.
18912 mosphere model ECHAM5 (Roeckner et al., 2003) to allow for vegetation-atmosphere feedbacks. The model version on spatial resolution of T63 (approx.

Selected trait and climate data
In JSBACH, SLA (m 2 kg −1 carbon) is related to the amount of carbon that can be stored in the green and reserve carbon pools. Vcmax 25 (µmol m −2 s −1 ) and Jmax 25 (µmol m −2 s −1 ) both control carbon assimilation. Vcmax 25 also determines the reference respiration at 25 • C. Observational data for these traits (see Table 1 for references) Most Vcmax and Jmax values had already been standardized to this temperature via the formulation of the photosynthesis model by Farquhar et al. (1980) used in the con-5 text of JSBACH . For other records for which the temperature during measurement was recorded, standardization was done according to this formulation. For C4-grasses, variation in PEPcase CO 2 -specificity (PEP, mmol m −2 s −1 ) instead of Jmax 25 was determined, following Collatz et al. (1992). Since for C4-grasses insufficient observational PEP and Vcmax 25 data were available, these traits were estimated more indirectly applying insights from Simioni et al. (2004), who determined PEP and Vcmax 25 based on leaf nitrogen (N) content (g m −2 ) (regressions based on two C4-species). Therefore, for C4 grasses, additional information on leaf N was obtained from the TRY database.
To link climate and trait data, only geo-referenced trait observations from field sam- 15 pling and field experiments were used. PFT assignment of species was based on available data in the above mentioned databases on growth form (shrub, grass, tree), leaf habit (deciduous/evergreen) and photosynthetic pathway (C3/C4). The climatic domain of a species (tropical, boreal etc.) was determined based on the Köppen-Geiger climate classification (Kottek et al., 2006) and applied to the geo-referenced observations. This 20 resulted in 12 394 observations for SLA distributed over 2869 species and 1052 (PFTspecific) sites, 761 observations over 129 species and 70 sites for Vcmax 25 and 402 observations over 108 species and 56 sites for Jmax 25 (see supplementary material S1 for global map with locations of trait data). Global climate data on mean annual precipitation (MAP, mm yr −1 ), mean annual rela- 25 tive humidity (Reh, %), mean annual temperature (MAT, • C), mean temperature of coldest and warmest month (T min and T max , • C) were collected from a global 10 min gridded data set of mean monthly climate data based on weather stations from the Climatic Research Unit (CRU) (New et al., 2002). Net shortwave radiation (NSWR, (W m  5 since soil N was not modeled in this version of JSBACH.

Trait-climate relationships
Given the lack of global quantitative deterministic predictions, the observed natural variation of SLA, Vcmax 25 and Jmax 25 was described for each PFT by the simplest models possible, i.e. linear regressions. For this purpose, we used PFT-specific site 10 means to match the scale of modeled fluxes most closely. Observed site means were related to the multiple environmental conditions at these sites in a linear regression, weighting sites by the square root of the number of observations (data points) per site. Such relationships likely reflect habitat filtering which commonly leads to trait convergence at community scales. Regressions with the highest explained variance (highest 15 R 2 adjusted ) were selected, after checking for significance of climatic drivers and distribution of residuals. In case of co-linearity between environmental drivers (Pearson's correlation higher than 0.7 or lower than −0.7), the driver with lowest significance was omitted. Due to the low number of sites for Vcmax 25 and Jmax 25 , data for the two tropical tree PFTs were combined, as well as data for the two shrub PFTs. This resulted in 20 eight relationships for SLA and six for both Vcmax 25 and Jmax 25 . These relationships were subsequently implemented in JSBACH. In addition, to allow trait calculations to go beyond climatic regions used in the regressions, but to still maintain traits within ranges as observed in nature, simulated trait ranges were constrained to the 2.5-97.5% quantiles of all individual observations within a PFT (instead of the site means). For Jmax 25 25 and Vcmax 25, an additional constraint was applied to maintain the strong physiological correlation between them (Medlyn et al., 2002;Wullschleger, 1993), keeping Jmax 25 and Vcmax 25 values within the 95 % confidence interval of the linear regression among 18915 these two traits (as determined per PFT). Predicted values outside this range were adjusted to the confidence interval border, based on the shortest distance needed to reach the border. Since this version of JSBACH did not simulate N-cycling, soil N could not be used to parameterize Vcmax 25 , even though N-availability is a strong determinant (see references in Kattge et al., 2009). However, the above two constraints on Vcmax 25 were based on observational data, indirectly incorporating N-limitations.

Simulation setups
Three different scenarios were performed: 1. a simulation containing the default parameterization of JSBACH with constant parameter values per PFT, based on Raddatz et al. (2007) and Kattge et al. (2009), and adapted to approximate realistic vegeta-10 tion functioning within the vegetation dynamics module (Brovkin et al., 2009); hereafter called "default simulation", 2. an "observed traits simulation" again with constant values for SLA, Vcmax 25 and Jmax 25 per PFT, but based on observational data only (the weighted site means for each PFT, from here on called "observed global means"), and 3. a "variable traits simulation" in which traits were allowed to vary depending on local climatic conditions. In the latter simulation, at the beginning of every year each of the three key traits was re-parameterized for each PFT in every terrestrial grid cell of the world, depending on the local simulated climatic conditions in each grid cell. The observed traits simulation was performed in order to separate the effects of replacing trait constants by observations from the effects of trait variation in the model. 20 To get vegetation into quasi-equilibrium with the simulated climate (but with CO 2 concentration fixed), the coupled model (JSBACH/ECHAM5) was run for 150 yr with vegetation dynamics in an accelerated mode (i.e. vegetation was simulated 3 times for each year of simulated climate). Next, to get the slow soil carbon pools into equilibrium, the uncoupled carbon model of JSBACH ("CBALANCE") was run for 1500 yr. The coupled 25 model was continued for 160 yr with updated carbon pools (with the vegetation module in an accelerated mode for the first 50 yr), until carbon pools and vegetation cover had reached equilibrium. The last 10 yr were averaged and used for further analysis, 18916 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | which in this setup (with prescribed seasonal sea surface temperatures) was sufficient to account for the inter-annual climate variation.

Data comparison
Model performance was investigated by comparing vegetation distribution, GPP and biomass of the simulations with observations. Vegetation distribution of the dominant 5 PFT (Fig. 5) was compared to the potential (natural) vegetation map of Ramankutty and Foley (1999), which in part is used to initialize simulations in JSBACH. The mismatch in spatial resolution (the potential vegetation maps is at a global 30 min spatial resolution) was solved by counting the number of PFTs according to Ramankutty and Foley (1999) that were present in each JSBACH-grid cell. The PFT with the highest occurrence was compared with the simulated dominant PFT of JSBACH.
To match the PFT-classifications, aggregated PFTs were constructed (Table S4.1, Fig. S4.1), as the PFTs of Ramankutty and Foley (1999) could not be directly related to those of JSBACH. Temperate and boreal evergreen trees were merged to match extratropical evergreen trees in JSBACH. The same was done for deciduous trees. Shrubs 15 (marginal in JSBACH) were merged into a single PFT ("shrubs") in both JSBACH and the vegetation map. As savannas consist of grasslands and woodlands, matches between savannas and C4-grasses or tropical broadleaved deciduous trees were both classified as correct. Tundra and mixed forests were omitted from the comparison as no equivalent PFT was available in JSBACH. 20 Model performance was determined with Cohen's kappa (κ) (Cohen, 1960), which is the proportion of agreement between vegetation maps, while accounting for chance agreement. Grid cells with sea or ice as the dominant cover were not taken into account, and neither were cases in which more than 1 PFT shared the highest occurrence (i.e. equal number of grid cells). This resulted in a comparison of 2819 grid cells. 25 Simulated biomass was compared to estimates from Robinson (2007), which include adjusted estimates of belowground (root) carbon in plant biomass. As we simulated only natural vegetation and the estimates of Robinson (2007)  including land use change and crops, comparing global total biomass would result in an overestimation by our simulations. Therefore, comparisons were made per m 2 per (aggregated) PFT. JSBACH does not have a separate savanna-like PFT. Therefore, tropical forests and C4-grasslands were averaged and compared with averages of tropical forests and sa- 5 vanna. Furthermore, extra-tropical trees were averaged and compared with averaged temperate and boreal trees of Robinson (2007). Mediterranean shrubs and tundra were omitted as there were no comparable PFTs in JSBACH.
Latitudinal patterns of median GPP were compared with data taken from Beer et al. (2010), who combined observational data (eddy covariance fluxes) with diagnostic 10 models to approximate GPP.

Mismatch between observed trait data and default trait settings
Within each PFT, a large observed range in the three trait values was apparent, as well as large overlap in trait ranges among PFTs (Fig. 1). This was most manifest for 15 SLA, where all PFTs overlapped with each other. For Vcmax 25 , the only PFTs of which the 95 % confidence interval did not overlap were both tropical evergreen trees and C4-grasses with C3-grasses and for Jmax 25 both tropical trees and deciduous shrubs with C3-grasses.
For all three traits, the default parameter values deviated strongly from the PFT-20 specific observed global means (grey diamonds and black circles in Fig. 1 tropical Evergreen Trees, Raingreen shrubs and C4-grasses), the other half higher. These differences point to a strong mismatch between default trait-parameters in JS-BACH with observed natural trait means.

Simulated trait variation based on climatic drivers
The observed variation in PFT-specific site means was related to variation in environ-15 mental drivers (Table 2). All regressions were significant, except for C3-grasses for Vmax 25 and for Jmax 25 . Although the selected environmental drivers (Table 2) differed among PFTs, net shortwave radiation (NSWR) and mean annual precipitation (MAP) were most frequently selected as drivers for SLA, as were MAP, mean annual temperature (MAT) and relative humidity (Reh) for Vcmax 25  PFTs had one constant value per trait, a large climate-driven trait range was predicted for the variable traits simulation. The higher SLA of deciduous versus evergreen trees observed in natural vegetation was reflected in the simulated SLA ranges. As for the observations, there was large overlap in trait ranges among PFTs for the three simulated traits as well, particularly for Vcmax 25 and Jmax 25 . Simulated SLA-ranges were narrower and overlapped less among PFTs compared to observed ranges (Fig. 1). For all three traits, a strong mismatch of default traits with the simulated trait variation was apparent again: for SLA default values fell (at least) outside the 25-75 % inter-quartile range in four PFTs (vs. one in the observed data), and for Vcmax 25 and Jmax 25 for every PFT.

Inclusion of trait variation alters predicted global patterns of productivity
The observed traits and variable traits simulation showed a similar pattern of median GPP along the latitudinal gradient compared to the default simulation (Fig. 3 different ways. Again, at higher latitudes, the observed traits simulation predicted highest NPP. For the variable traits simulation, the reduction from gross to net primary productivity around the tropics is much stronger than for either other simulation, resulting in smaller NPP differences among simulations compared to GPP differences (even though differences may still go up to 75 % between 25-30 • N for both observed 25 and variable traits simulations). The large reduction in differences between NPP and GPP among simulations cannot solely be attributed to a higher respiration of PFTs due to higher Vcmax 25 (see Fig. 4), and is mainly a consequence of net productivity 18920 Introduction exceeding the maximum sizes of the different carbon pools (e.g. of leaves, wood, reserves) as defined in JSBACH, resulting in respiration of excess carbon. This loss of excess carbon is visible in the tropical regions for all simulations, but for the variable traits simulation this resulted in a proportionally larger amount of productivity removed. This was enhanced by the higher SLA in these regions, which partly determines the 5 amount of carbon stored in the living parts of the plants, consisting of leaves, fine roots and sapwood. These results show that, while reproducing global patterns of productivity, incorporating global trait variation in the model leads to strong changes in predicted productivity. Even though GPP is not only determined by the photosynthetic parameters (Vcmax 25 and Jmax 25 ), as it is affected by e.g. water availability as well, on average, shifts in GPP were accompanied by shifts in similar directions by mean Vcmax 25 and Jmax 25 (weighted by fractional coverage of PFTs), compared to the default simulation (Fig. 4). However, this was not always the case, as e.g. the drop in Vcmax 25 and Jmax 25 above 40 • N did not lead to a coinciding drop in GPP. Productivity could not be related to 15 similar shifts in SLA. In the observed traits simulation, tropical evergreen trees, dominant in the default simulation, were taken over by deciduous trees in Africa, South America and Australia 18921 the areas where trees were dominant. The shifts from tropical evergreen to tropical deciduous trees cannot be explained by Vcmax 25 and Jmax 25 , since these tropical PFTs were parameterized with the same values for these traits. The most profound difference between these tropical PFTs seems to be their leaf turnover rate, which is higher for the deciduous than for the evergreen trees. As a consequence, tropical deciduous trees had somewhat lower leaf area index (LAI), which meant lower productivity in favorable periods, but also less carbon loss in more stressful circumstances (e.g. drier periods). In some areas, this could have resulted in a higher total yearly NPP for tropical deciduous trees, thereby outcompeting evergreen trees. Another shift in predicted dominant vegetation in the variable traits simulation was 15 an increase in C4-grasses (from 4.8 % in the default simulation to 8.4 %). This occurred mostly in Africa and Australia at the expense of tropical trees and raingreen shrubs. This expansion of C4-grasses below the Sahara coincided with higher fractions of burned area, which promoted the expansion of grasses at the cost of trees.

Major shifts in vegetation distribution
In the variable traits simulation bare ground increased as dominant cover type (from 20 15.4 in the default simulation to 19.5 %) in the southwest of the United States (and Mexico), northern Canada and northeast of Siberia at the expense of C3-grasses and deciduous trees, resulting in a shift of the boreal tree-line toward lower latitudes. These shifts often coincided with a decrease in Vcmax 25 (see S3), suggesting lower productivity and consequently less expansion of these PFTs.

Modulation of climate by traits
Predicted climate differed among simulations due to trait-dependent climate-vegetation feedbacks. Precipitation differences were local and showed no clear spatial pattern 18922 the US and Australia, and lower (up to 1 • C) in large parts of Europe and Russia, South-Africa and South-America (Fig. 6a), meaning that (at least) temperature is very sensitive to parameterization of traits. Temperature differences were less profound between default and the variable traits simulation (Fig. 6b), but still went up to around 1 • C. Changes in temperature did not correlate with clear changes in traits or vegetation shifts (e.g. tree-grass shifts), but in the Southern Hemisphere (Australia, Africa, South-America) corresponded to differences in transpiration, where cooler areas coincided with higher transpiration. This could be related to the higher Vcmax 25 in these areas, resulting in higher GPP and consequently an increase in transpiration. 15 The differences between the observed traits simulation and variable traits simulation indicate that by allowing traits to vary and respond to environmental conditions (as in the variable traits simulation) feedbacks between climate and traits result in more moderate temperature shifts, showing the significant magnitude of adaptive traits and on climate.

Comparison of model output with observational data
Cohen's κ, indicating the correspondence of the global map of potential (natural) vegetation of Ramankutty and Foley (1999) with simulated vegetation distribution, was 0.289, 0.282 and 0.334 for the default, observed traits and variable traits simulation, respectively. These values are somewhat lower than the performance of other DGVMs 25 (κ = 0.42 for LPJ (Hickler et al., 2006) or κ = 0.42 for a consensus map of multiple DGVMs, Cramer et al., 2001). This means that for either simulation, a large proportion of the grid cells did not match the potential vegetation map (Fig. S4.2) there is a substantial increase in similarity to observed vegetation from the default and observed traits simulation toward the variable traits simulation. Mismatches occurred in large parts of the US and Canada, where the simulations predicted mostly C3-grasslands, while according to the potential vegetation map also forests should be present. The same holds for large parts of Europe. The potential 5 vegetation map shows less bare ground than any simulation, resulting in mismatches in the US, but also other parts of the world. Furthermore, almost the whole continent of Australia did not correspond to this map; shrubs and savanna are dominant according to the map, and even though the models did predict C4-grasses there, it was in different areas. Where the default and observed traits simulation had low correspondence 10 with the vegetation map in Africa and South-America, the variable traits simulation performed better, mainly with respect to the tropical trees. Even though differences in performance are small, the variable traits simulation matched the potential vegetation map most closely.
Comparisons of biomass estimates per m 2 per (aggregated) PFT with current 15 biomass estimates by Robinson (2007), show that for the combined tropical trees and savannas the variable traits simulation (13.12 kg C m −2 ) most closely matched biomass estimates (12.82 kg C m −2 ) (  Fig. 3), each simulation produces substantial higher GPP at most latitudes (on average leading to 2, 2.4 and 2.6 times higher GPP for the default, observed traits and variable traits simulation, respectively), with the default simulation in general actual differences may be less. Also many other DGVMs show higher GPP in the tropical areas than the observed median GPP by Beer et al. (2010) and the GPP estimates of the variable traits simulation still do fall within the upper range of GPP predictions by other DGVMs.

10
The aim of this study was to improve modeling of vegetation responses and allow vegetation-atmosphere feedbacks in DGVMs by incorporating trait variation. This trait variation was based on relationships between measured traits and climate and soil moisture, representing major assembly processes by the abiotic environment. As model intercomparisons have shown that large uncertainties exist in projections of land 15 carbon uptake by DGVMs (Cramer et al., 2001;Friedlingstein et al., 2006), incorporation of variation in vegetation responses is important to allow feedbacks between vegetation and climate and to increase plausibility of model predictions, especially under strong climate change. Here, we incorporated variation in plant responses based on observed trait and climate data into the JSBACH DGVM, which revealed profound 20 effects on carbon fluxes and vegetation distribution.

Advantages of modeling trait variation based on trait-climate relationships
We used trait-climate relationships to describe the observed natural trait variation and implemented these in JSBACH. The use of such relationships is broadly accepted and applied in ecology (Ordoñez et al., 2010;Wright et al., 2005b)  captures multiple (approximate) drivers of natural trait variation. The relationships are thought to reflect abiotic assembly processes and integrate multiple vegetation responses at different temporal and spatial scales, including acclimation, adaptation of species and species replacement into a spatially and temporally varying trait mean. Recently, other DGVMs implemented some trait variation necessary to allow feedbacks 5 when modeling N-cycles (Gerber et al., 2010;Zaehle and Friend, 2010). However, in these models effects of trait variation on vegetation cannot be separated from nutrient effects, whereas our study focuses on direct effects of trait variation. Furthermore, the drivers of trait variation in those studies are limited and mainly based on soil N (indirectly included in our study through variation in Vcmax 25 ), whereas here multiple drivers 10 of variation have been identified, including water availability (Misson et al., 2006). Even though our proposed method does not explain mechanistically the processes behind trait variation and does not take into account effects of biotic interactions or dispersal on trait values (similar to most DGVMs), it is in our opinion an important and necessary step as it reflects the observed causality between traits and climatic drivers (Niinemets, 15 2001;St. Paul et al., 2012;Wright et al., 2005b). It has the advantage that it does identify and quantify multiple abiotic drivers of trait means and in this way captures a large part of observed trait variation, as shown by the substantial R 2 adjusted of most regressions.
We updated trait values once per year. This time-step is a balance between com-20 putational efficiency and ecological realism. By this approach we avoid to evaluate ontogenic impacts on trait values and whether environmental impacts differ for different parts of the growing season, for which currently insufficient information is available. Our approach assumes that the yearly updated leaf trait values are in equilibrium with their environments, which is consistent with ecological observations. Compared to e.g. wood 25 traits, the selected leaf traits are rather plastic and can vary within a year (Dubey et al., 2011;Misson et al., 2006). SLA may be somewhat less plastic than the two photosynthetic traits, but a yearly re-parameterization is reasonable, as leaves of deciduous species are shed every year. The maximum lifespan for leaves of evergreen PFTs is Introduction three years in the model, so shifts occur for at least a third of the leaves for these trees. Moreover, the yearly shift in leaf trait values may not only reflect acclimation, but genetic adaptation and species replacements may also contribute. In our approach, in contrast to a mechanistic approach, the impacts and (unknown) time-scales of those processes leading to trait shifts do not have to be differentiated.

Implications of incorporating observation-based trait variation
The observed global mean trait values of natural vegetation as used in the observed traits simulation strongly deviated from trait values in the default simulation, indicating a mismatch between PFT trait means of modeled and natural vegetation. Moreover, either set of constant values contrasts strongly with the large range of trait values observed in natural vegetation (Fig. 1). While we applied the most comprehensive database available today, we are aware that estimates of observed trait variation are still uncertain (Table 2) and need to be improved in future applications. Nevertheless, the wide range of observed trait values illustrates how simulations with constant traits do not reflect natural trait variation. In contrast, this variation was reflected by the 15 variable traits simulation where trait variation represented abiotic assembly processes (Fig. 2).
To investigate the effects of trait variation on vegetation-climate feedbacks it was essential to incorporate vegetation dynamics, to allow trait shifts to alter vegetation distribution and in this way modulate productivity and climate. In contrast to the simulations 20 with constant traits, the variable traits simulation enabled such interactions between vegetation and climate via traits to occur. These interactions resulted in more spatial variation in dominant vegetation compared to the other two simulations. As such, predicted vegetation distribution is more a result of temporal dynamics in vegetation properties than is the case in the other simulations, where these vegetation proper- Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | directly modulated model output. This problem is well known (Quillet et al., 2010) and stresses the importance of simulations like the ones presented. Provided the strong effect exerted by climate on traits, major differences among the simulations in predicted vegetation distribution and productivity were expected, the latter especially when parameters that affect assimilation rate are concerned, as sensitiv-5 ity analyses of DGVMs have shown (White et al., 2000;Zaehle et al., 2005). Indeed, for the observed traits simulation and variable traits simulation this resulted in large differences in the new equilibrium state, both compared to the default and to each other. Not only were vegetation properties affected, but also climate changed and mean temperatures were altered by up to more than 1 • C.
As the simulations in this study provide equilibrium states, and as such do not necessarily correspond to current climate or vegetation composition, model comparison to observational data must be interpreted with care, although it does provide insights in the realism of the simulations. DGVMs are parameterized to produce approximately realistic results, and therefore our simulations were not expected to approach obser-15 vations better than the default simulation. Even though the variable traits simulations produced high GPP for tropical areas, its biomass estimates and vegetation distribution more closely resembled observational data than the default simulation.
The large differences in model output, the mismatch between default trait values in the model and observed trait variation in nature, and the improved performance of the 20 variable traits simulation, demonstrate that -besides a correct representation of plant physiology (e.g. photosynthesis, transpiration) -integration of ecological theory will help to improve vegetation representation in DGVMs and ESMs. Allowing for variable vegetation responses to climate and enabling plant-atmosphere feedbacks will have important consequences for predictions by vegetation models, for vegetation distribution 25 and productivity as well as for global current and future climate. A model intercomparison with this approach under elevated CO 2 projections, for which large uncertainties in predictions by current DGVMs exist (Friedlingstein et al., 2006), should therefore be one of the next steps.

Discussion
Paper | Discussion Paper | Discussion Paper | Discussion Paper |  For C4-grasses, PEPcase CO 2 -specificity (PEP, in mmol m −2 s −1 ) instead of Jmax 25 is modeled. Boxplots and symbols as in Fig. 1. PFTs as in Fig. 1 but without cold/deciduous shrubs (DSh), because this PFT was never dominant. Observed global means are added again for illustrative purposes. Note that direct comparison with simulated data is not appropriate here as the range of climatic conditions do not overlap for observed and simulated trait data.
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |