Symbiotic nitrogen fixation (SNF) is the main natural source of nitrogen (N) in terrestrial ecosystems worldwide. Previous studies have shown that fixation of N by plants can be limited by the availability of phosphorus (P) in soils. We used global meta-analysis to investigate how P availability controls SNF. In experiments in which plants were grown in an artificial medium, severe P deficiencies in the nutritive solution ([PO4] < 5–42 μM) depressed SNF flux through both a direct decrease in the plant fixation rate (i.e., decreased N fixed per unit of plant biomass) and an indirect effect (i.e., through plant biomass). In most experiments with plants grown in soils, SNF was proportional to plant biomass and was consequently only indirectly limited by P. Some cases using unfertilized and weathered soils (ultisols or oxisols), where plants were particularly P stressed, were an exception with both direct and indirect P limitations. Our global analysis of the P-SNF relationship indicated that P bioavailability commonly limited SNF flux. We conclude that the main driver of in situ P limitation is indirect via limitation of plant growth, except in certain cases where both indirect and direct constraints may play a role. These cases of severe P deficiency may be mainly found in weathered tropical soils of Africa and South America, probably in unfertilized croplands which are depleted in P due to repeated biomass harvests.
 Nitrogen (N) is an essential element for all living organisms as it is a component of vital molecules such as proteins and nucleic acids [Marschner, 1995]. Because N is not present in the original Earth's crust, one major challenge for organisms is to access a sufficient supply of N [e.g., Hedin et al., 2009]. As a consequence, N limits the productivity of terrestrial ecosystems in large areas in the world [Elser et al., 2007; Vitousek and Howarth, 1991]. Nitrogen is not present in most soil parent materials, but N2 is the major component of the atmosphere. However, only a relatively small number of species of microorganisms have the ability to use atmospheric N2 for their own metabolism [Mylona et al., 1995]. These microorganisms use specific enzymes (i.e., nitrogenase) to break the triple chemical bond of atmospheric N2 and transform this molecule into bioavailable NH3 [Hartwig, 1998; Mylona et al., 1995; Reed et al., 2011; Vitousek et al., 2002]. When performed by free-living microorganisms, this reaction is called asymbiotic N2fixation [Reed et al., 2011]. Due to the very high energetic cost of the reaction [Hartwig, 1998; Rastetter et al., 2001; Schulze, 2004; Vitousek et al., 2002], fluxes of asymbiotic N2 fixation are generally low [e.g., Barkmann and Schwintzer, 1998; Jurgensen et al., 1992; Son, 2001] (see Reed et al.,  for exceptions) because microorganisms first need to find enough energy (i.e., carbohydrates and phosphorus (P) for ATP) [Reed et al., 2011]. Evolution provided a much more efficient way to break down atmospheric N2 molecules by associating specialized microorganisms with plants in a symbiotic relationship. These microorganisms are located inside nodules on roots of vascular plants [Hartwig, 1998], or in the tissues of other plants like mosses [DeLuca et al., 2002], and supply available N in exchange for energy, phosphorus, and small quantities of other nutrients necessary for nitrogenase activity, like iron (Fe) or molybdenum (Mo) that is provided by the plants [Finzi and Rodgers, 2009; Hartwig, 1998; Vitousek et al., 2002]. Symbiotic N2fixation (SNF) represents a flux of available N to terrestrial ecosystems that is of the same order of magnitude as fertilization in agriculture (i.e., about a hundred Tg N yr−1 depending on the method of estimation) [Galloway et al., 2004; Herridge et al., 2008]. SNF is a process of major importance at the global scale because the global carbon cycle is tightly coupled with the N cycle [Finzi et al., 2011b; Galloway et al., 2004; Gruber and Galloway, 2008; Hungate et al., 2003; Thornton et al., 2007; van Groenigen et al., 2006]. How SNF is limited (or not) is consequently a major question to be addressed in the context of global change.
 In aquatic ecosystems, SNF is probably limited, among other factors like temperature or sunlight, by the concentrations of P and Fe in the water [Dutkiewicz et al., 2012; Hutchins et al., 2009; Mills et al., 2004; Sanudo-Wilhelmy et al., 2001; Vitousek et al., 2002]. In terrestrial ecosystems, soil content in Fe appears to be a much less limiting factor of SNF [Finzi and Rodgers, 2009; Vitousek et al., 2002] than P availability which has long been recognized to be a limiting condition [e.g., Graham and Rosas, 1979]. Although the interactions between P and SNF have been identified as one of the few important research areas in N2 fixation [Finzi et al., 2011a; Graham and Vance, 2000], the way P controls SNF is still a source of controversy [e.g., Israel, 1987] and is not yet clear [Hartwig, 1998; Vitousek et al., 2002]: Do plants respond to P by increasing their growth (with a constant specific fixation rate: Nfixed gplant−1) and hence indirectly fix more N, or do they directly adjust the relative amount of N2 fixation per unit biomass? Some experiments in laboratory conditions indicated that nutritive solutions highly deprived of P can have direct negative effects on plant fixation [Reddell et al., 1997; Ribet and Drevon, 1996] mainly by disturbing nodulation or nodule functioning [O'Hara, 2001]. Conversely, some field experiments showed only indirect effects of P supply on SNF via plant growth [e.g., Robson et al., 1981]. Considering the importance of SNF for most terrestrial ecosystems, our aim was to elucidate the dependence of the process on P availability.
 In practice, we studied the relative importance of direct and indirect effects of P supply on SNF by meta-analysis. To this end, we compiled published studies that measured the SNF of plants under different levels of P supply. Direct effects were assessed through the specific plant fixation rates. Following the hypothesis of McLaughlin et al. , we assumed that direct effects of P availability on SNF would also be visible through altered N content of plants (μg N gplant-1), because in this case plants would not be able to satisfy their requirement for N (through SNF or uptake of available N from the soil) and consequently N content would be low.
2 Material and Methods
2.1 Compilation of Symbiotic Nitrogen Fixation Studies With Different Levels of P Supply
2.1.1 Collection of Published Studies and Definition of a Case Study
 We used the ISI Web of Science database to locate published studies on the effect of P supply on SNF. Following the protocol of Pullin and Stewart , we entered holistic nonspecific queries to cover a large proportion of the literature related to our subject. Most of the published studies were (1) “nutritive solution experiments,” i.e., plants grown in an artificial substrate in which the concentration of P (in phosphates [PO4] forms) in solution was maintained constant (unit used: μM P), or (2) “soil experiments,” i.e., plants grown in soil (in pots or in the field) with an initial P supply (unit used: µg P gsoil−1). In all cases, at least two different levels of P supply were used in the experimental design of each study. Obviously duplicate studies (i.e., studies by the same team of researchers and using a similar experimental design) were excluded (study with the largest data set was retained). Other experimental approaches (e.g., starvation experiments or experiments using both soils and constant P supply) were rare and were consequently not included in our meta-analysis.
 In the studies we retained, we selected treatments that enabled us to quantify the effect of P supply. In practice, when factors other than the level of P supply were tested (e.g., atmospheric CO2 concentration or N supply), we chose the control and ignored the other factors; i.e., we selected the measurements made when only P supply varied. If there was no control treatment for factors other than P, we selected the minimum constant treatment related to these factors. The experimental designs of the selected studies are summarized in Table S1 (supporting information A).
 A crucial point when performing a meta-analysis is the way a case study is defined [Gurevitch and Hedges, 1999]. Indeed, the quantity of data varies considerably among studies, and assuming that each value is an independent case study leads to different statistical weights. However, the values reported in a study may be repeated measurements (e.g., several sampling dates or several similar sites) and consequently bias the meta-analysis. Here, one case study reporting results on the effects of P supply on SNF was considered as the unique combination of one team of researchers, one species, one growth medium (artificial substrate or one type of soil), and one sampling date. When several results corresponded to only one case study, we selected a single mean value to avoid pseudoreplications (additional details on the method are available in the supporting information A).
2.1.2 Description of the Database
 The final database on SNF experiments contained data from 68 references (Table S1). Among them, 25 used an artificial growing medium and a nutritive solution and 43 used at least one type of soil (14 pot experiments; 29 field experiments).
 We identified 88 case studies to build the database. The compilation included results on 36 plant species grouped in three different types (annual plant species, nonwoody perennial plant species, and woody perennial plant species) from most biomes, climates (snowy to tropical), land uses (croplands, forests, grasslands, and savannas), and soil types (see below for a description of the soil classification). Case studies using nutritive solution (N = 34; [PO4] = 0–3229 μM) or soil (N = 54; P dose = 0–1852 µg P gsoil−1) were both well represented. Soil experiments were well distributed worldwide (Figure S1).
 The database had the following responding variables, when available: plant biomass (aboveground, belowground, total biomass), nodules (number, specific biomass (mg nodule−1), biomass (mg plant−1)), plant N or P content (foliage or aboveground, belowground, nodules), specific N2 fixation rate of plants (Nfixed gplant−1) or nodules (Nfixed gnodule−1), and N2 fixation flux. In addition to the dose of P supplied, other potential explanatory variables were recorded including soil order (see below), climate (sensu Kottek et al. ), biome (adapted from Melillo et al. ), plant type (annual, nonwoody perennial, or woody perennial), type of ecosystem (forest, cropland, grassland, or savanna), or symbiont genus (Rhizobium, Bradyrhizobium, or Frankia). We also recorded data on soil available P content as determined by chemical extraction, when this variable was measured in the case study. The pool of soil available P depends on soil type [Batjes, 2011b] and more precisely on soil total content in P [Messiga et al., 2010] and soil physical-chemical properties which control the P buffering capacity [Stroia et al., 2007]. Therefore, we needed to group case studies based on a soil classification and a buffering capacity classification.
 We used the USDA soil classification which has the advantage of being usable for all soils in the world and has a relatively low number of macroclasses of soils (12 orders of classification): alfisols, andisols, aridisols, entisols, gelisols, histosols, inceptisols, mollisols, oxisols, spodosols, ultisols, and vertisols (see http://soils.usda.gov/technical/classification). When not provided in the reference, the soil order was defined based on the available soil description, soil geographical location, and correspondences among systems of soil classification [Batjes, 2011a; Esu, 2010; Gerasimov, 1962]. The P buffering capacity of a soil refers to its ability to not only fix P on its solid phase but also control and maintain the PO4 concentration in the soil solution: the higher the P buffering capacity, the lower and more stable the PO4 concentration. We based our classification of the P buffering capacity of soils on physical-chemical properties and specific references [Achat et al., 2009; Batjes, 2011b; Sanchez et al., 2003; Smeck, 1985]: very high buffering capacity (andisols (soils rich in allophanes) and oxisols (soils rich in oxides)), high (ultisols (soils rich in secondary minerals)), moderate (alfisols, aridisols, gelisols, inceptisols, mollisols, and vertisols), low (entisols (young soils) and spodosols (topsoil horizons with low P affinity)), and very low (histosols and typical holorganic soil horizons like the forest floor or the organic top layer of tundra soils).
2.1.3 Methods of Quantification of SNF Flux and Processes
 SNF is generally quantified using isotopic methods (with 15N) or the acetylene reduction assay for nitrogenase activity (ARA method). Less common methods, based on measuring ureides in sap, natural abundance of 15N or total nitrogen pools, are also found in the literature. We ranked results in three level of confidence (low, moderate, or high; all details are provided in the supporting information A in Table S2 and related text) because all these methods have some drawbacks, depending on the experimental design and context [e.g., Shearer and Kohl, 1986]. Briefly, 15N methods are integrative as they measure N2 fixed in the biomass produced throughout the course of the experiment and are generally considered as reliable, provided that the 15N signature of the N2 fixing plant is different enough from that of a nonfixing plant (high level of confidence). Conversely, the ARA method measures nodule activity over a short period of time, but estimation of nodule weight may be subject to serious bias. Therefore, we considered the results at the plant scale (Nfixed gplant−1) obtained by ARA method with variable level of confidence, depending on the context (Tables S2). In all cases, the ARA method was assumed to be adequate (high level) to assess physiological aspects of SNF at the nodule scale (e.g., specific activity at the nodule scale in Nfixed gnodule−1 or the nutrient content of nodules).
2.2 Data Analysis and Statistics
2.2.1 Characterization of Plant P Availability
 In nutritive solution experiments, P supply values were grouped into logarithmic classes (e.g., 0–10, 11–100, and >100 μM) to be able to compare plant responses under P supplies that vary by several orders of magnitude between studies. This classification enabled us to emphasize plant responses under a very low P supply where we suspected that strong P limitation could result in nodule dysfunction. P supply was also used as a continuous variable during analyzes.
 In soil experiments, it was not appropriate to quantify the P supply using the dose of P-fertilizer applied because the dose effect strongly interacts with the soil P buffering capacity and initial soil P content [Batjes, 2011b]. Because the latter variables were not provided in the majority of soil experiments, instead we used the response of the N2 fixing plant to P fertilization. To do so, we used the combination of the biomass relative response of N2 fixing plants and the plant foliar P index to assess the local P limitation of N2 fixing plants (see below).
2.2.2 Characterization of Plant Response to P Supply
 To ensure comparability among studies which were not originally designed to be compared, we used three statistics:
 In all soil experiments that included an unfertilized control treatment, we used the concept of the effect size metric (hereafter referred to as relative response) calculated as follows:
Values close to 0 are associated with a negligible effect of the treatment, while negative and positive values indicate negative and positive effects of the treatment, respectively. Confidence intervals were calculated with the bootstrap method in MetaWin-2.1 software (Rosenberg et al.: www.metawinsoft.com). The relative response is a reliable approach to quantify the effects of a binary experimental design, like effects on plant biomass of fertilization versus control [Elser et al., 2007].
2.We estimated the foliar P index of plants:
The critical P content of foliage is the value below which a defined plant species cannot grow at optimal rate due to P stress. The values of critical P were recorded from a plant analysis manual which compiled data on a wide range of plants species from all types of ecosystems in the world [Commonwealth Scientific and Industrial Research Organisation, 1997]. When several values of critical P were indicated, we selected the value obtained with experimental designs comparable to those of our SNF experiments or used a mean value. Values of foliar P index below 1 indicate P-stressed plants, whereas values ≥ 1 indicate an adequate or high P foliage content.
3.We normalized data to the maximum value of each case study (e.g., plant fixation rate):
Each value is then a proportion of the maximum response, ranging from 1 (highest response in the experiment) to values closer to 0 (low responses). This approach is widely used to compile experiments, even without a true control treatment [Houlton et al., 2008].
 Relative response values were calculated only in soil experiments because nutritive solution experiments were lacking the necessary control treatment. Foliar P index values and normalized values were calculated for all experiments.
2.2.3 Testing the Significance of Plant Response to P Availability and Other Explanatory Variables
 Meta-analysis is a useful tool because it enables aggregation of data [Ainsworth et al., 2007; Ioannidis, 2010], but the aggregation is subject to several potential biases and hence requires a careful a priori protocol for data compilation [Gurevitch and Hedges, 1999; Ioannidis, 2010]. More methodological information on the meta-analysis is provided in the supporting information B.
 Differences among classes of explanatory variables (e.g., P supply, type of vegetation, or climate) were tested with a Bonferroni t test (or paired t test). Beforehand, the variables were transformed if they did not fit normality and variance homogeneity assumptions (Shapiro-Wilk test and Bartlett test). The normalized values were transformed with the logit function suitable for proportions comprised between 0 and 1 [Warton and Hui, 2011]:
 Following Warton and Hui, ϵ was set as the smallest difference between 1 and values (not equal to 1) of the data set (0.005 and 0.002 for nutritive solution experiments and soil experiments, respectively). The relative response was already log transformed and fitted the normality and variance homogeneity assumptions. For other variables, values were log transformed, if necessary.
 Finally, regression analyses were used to study the relationship between PO4 concentration and plant response (normalized specific fixation rate and plant biomass) in nutritive solution experiments. The equations used in our study were linear-plateau, bilinear, logarithmic, or of the Mitscherlich form [cf. Alivelu et al., 2003]. We only kept the fitted regressions which displayed realistic values (e.g., normalized values ≤1). The best model was selected using the modeling efficiency index [Mayer and Butler, 1993].
 The N2 fixing activity of the studied plants was generally high, as shown by the plant percentage of nitrogen derived from the atmosphere (first quartile, median, and third quartile values = 59%, 71%, and 81%, respectively; mean value = 68% ± 2%).
 Increasing P supply resulted in increased plant biomass, plant P content, and SNF flux (Table 1). These results were more pronounced in experiments using artificial growth medium (with nutritive solution) than in soil experiments (see Table 1 and below).
Table 1. Effect of Improved P Supply on Biomass, N-P Contents, and SNF Flux of Plants
aValues = mean ± 1 standard error. Significant differences were tested with paired t test between high and low P supply; low supply was 0–10 μM (artificial growth medium) or 0 µg P gsoil−1 (growth medium = soil). Calculations were not performed when there were less than five replicates. More results about soil studies are presented in Table S3 (supporting information C).
Artificial medium (nutritive solutions)
+286 ± 74%
P < 0.001
+42 ± 12%
P = 0.003
+58 ± 19%
P = 0.005
+277 ± 64%
P < 0.001
+354 ± 140%
P = 0.022
+113 ± 41%
P = 0.004
+980 ± 419%
P < 0.001
+106 ± 40%
P < 0.001
+6 ± 2%
P = 0.020
+41 ± 10%
P < 0.001
+52 ± 28%
P = 0.018
+176 ± 54%
P < 0.001
3.1 Symbiotic Nitrogen Fixation Studies Using a Nutritive Solution as Growing Media
 Plant biomass increased with an increase in P concentration in the nutritive solution up to an approximate value of 200–350 μM (75%, 85%, and 95% of optimal growth at 113, 175, and 368 μM, respectively; Figure S4 in the supporting information C).
 Values of foliar P index indicated that plants were highly P stressed, slightly P deprived, and well P fed at PO4 concentration in the nutritive solution of ≤10, 11–100, and >100 μM, respectively (Figure 1a). Increasing P supply often resulted in a high increase in plant N content (Table 1).
 The specific fixation rate of plants (Nfixed gplant−1) was highly dependent on the P concentration (in PO4) in the nutritive solution. Most of the positive effect of increasing P concentration in the solution on the plant fixation rate occurred between 0 and 100 μM (Figure 2a). The slope of the relationship was very steep in this range of values (Figure 2b), and it was difficult to fit a satisfactory regression because the beginning of the plateau section depended much on the case study concerned, with no clear influence of any of our other explanatory variables (data not shown). Nevertheless, it appeared that a close to optimal fixation rate (i.e., ≥ 75% of N2 fixing efficiency) was possible for P values higher than 5–42 μM, depending of the curve equation used (see methodological comments about this problem in Alivelu et al. ). The Mitscherlich regression (see Figure 2b) had the highest modeling efficiency (ME = 0.33 versus 0.28–0.32 for other equations).
 The increase in the specific fixation rate of plants via an increase in P supply was the consequence of a concomitant enhancement of the proportion of biomass allocated to nodules and their N2 fixing activity (Figure S5a in the supporting information C). A further increase in the P concentration did affect the specific fixation rate of plants, but the effect was of much lower amplitude as most cases ranged between 80% and 100% of N2 fixing efficiency above 100 μM (Figure 2a; see also Figure S5b for results on nodules). This result was consistent among all plant types tested (i.e., annual versus perennial species or nonwoody versus woody species; Figure 2b).
3.2 Symbiotic Nitrogen Fixation Studies Using a Soil as Growing Media
 In more complex and realistic conditions, i.e., in soil experiments, the response of plants to P supply displayed a somewhat different pattern compared to nutritive solutions experiments.
 P fertilization enhanced plant growth as shown by biomass values (Table 1 and Figure 3). There was a more or less pronounced effect depending on climate, ecosystem, soil type, or P dose (Figures 3a–3d). Conversely, the specific fixation rate of plants was generally not significantly modified by P fertilization. Only a moderate, though significant, increase (+14 to +31%) in plant specific fixation rate was observed when plant growth was strongly P limited (Figure 3e), or in soils with high—or very high—P buffering capacity (Figure 3c). These cases appeared to occur more frequently at low latitudes (tropical and dry climates) than at high latitudes (temperate and snow climates; Figure 3a), but this was not statistically confirmed. In our data set, the fixation rate of plants growing in soils did not depend on the type of ecosystem, P dose, or symbiont genus, even if plant growth was often severely P limited (Figures 3b, 3d, and 3f).
 The response of plant to P fertilization depends on the local soil P supply, which itself depends on the initial soil P availability and the soil P buffering capacity. We wanted to understand the effect of soil P supply on the specific fixation rate of plants. Data on soil available P, as estimated by chemical extractions, were very scarce: whereas 88% of soil experiments had at least one value of soil available P, only 16% (i.e., N = 7 experiments) had values for their P fertilization treatments. These seven experiments used all different methods of analysis (e.g., Bray, resins, and bicarbonate). Moreover, we recorded at least 13 analytical methods among the 72% of studies which analyzed only the control treatment (the method was not even indicated in nine references). Consequently, data on soil available P was hardly exploitable to interpret results on plant response (see below, however).
 Since we did not find enough data on soil P availability in the references compiled in our database, we used the combination of the biomass relative response and the foliar P index as a proxy to assess the degree to which specific N2 fixation rate of plants was limited by soil P availability. In soil experiments, for almost all the categories of biomass responses to P fertilization, plants showed an unchanged specific fixation rate. Only plants which had a strong response to P fertilization in terms of biomass had a significant increase of their plant fixation rate (Figure 4a). But within the group of plants having a high biomass response to fertilization, not all plants showed an increase of fixation rate. As a result, we recorded three different types of soil P limitation:
Plants which were extremely P limited: these plants were severely P deficient, as shown by their foliar P index which remained below 1, even after fertilization (high response studies in Figure 1b). They responded strongly to P fertilization, with a concomitant increase in biomass and specific fixation rate (Figure 4b). This significant effect on the specific fixation rate was related to a few experiments with soils mainly characterized by a high—or very high—soil P buffering capacity (i.e., ultisol or oxisols). This was confirmed by some of the few data on soil available P: control values of references using the same soil extraction method suggested that specific fixation rate of plants was suboptimal for soils known to be highly deficient in P, like ultisols (Figure S6).
Plants which were highly P limited: as for the former type of plant response, these plants were P deficient, as shown by their foliar P index which remained lower than 1, even after fertilization (high response studies in Figure 1b) and responded to fertilization by a high increase in biomass. Conversely, no change in specific plant fixation rate was recorded (Figure 4c). These cases corresponded to soils with only moderate—or lower—P buffering capacity, or andisols (i.e., soils known to have high P buffering capacity but also very high P content).
Plants which were only slightly P limited: these plants had their foliar P index slightly below 1 without fertilization and adequate (close to 1) after fertilization (no or low response studies in Figure 1b). Fertilization of these plants resulted in no—or low—biomass response and an unchanged specific fixation rate (Figure 4d).
 For most of the experiments where there was no effect of P supply on specific fixation rate of plants (types 2 and 3), both the nodule:plant biomass ratio and the specific fixation rate of nodules (Nfixed gnodule−1) remained stable (Figure S5b). In a few other experiments showing a stable plant fixation rate, the pattern seemed to be different as there was a decrease in the nodule fixation rate, but this was offset by an increase in the nodule:plant biomass ratio (Figure S5c).
 All in all, it appeared that plant fixation rate was not, or only moderately, modified by P fertilization (Figure 4a). In most of the cases, plants showing a positive biomass response to P fertilization had their N2 fixation rate unchanged, especially for plants which were only slightly P limited. The few cases where plants had their specific fixation rate improved by P fertilization were all initially extremely P limited, as shown by their biomass gain, foliage nutritive status, and soil P availability. Because plant fixation rate was at most moderately modified by P fertilization (Figures 3 and 4), the SNF flux of plants growing in soil was proportional to the increase in plant biomass (Figure 5). In soil experiments, improving P supply mainly resulted in an enhancement of plant biomass, which in turn, increased SNF flux. This result did not depend on plant type (e.g., annual versus woody) or ecosystem (Figure 5).
 Because the majority of experiments were carried out at the plant scale (e.g., plant biomass expressed as g plant−1), we finally investigated to which extent our results were applicable at the ecosystem scale (e.g., plant biomass expressed as Mg ha−1). We tested this by comparing experiments designed at the plant scale to experiments designed at the ecosystem scale. We found no significant difference of relative response for plant growth (P = 0.36) or plant fixation rate (P = 0.86) with no influence of any explanatory variables (like type of ecosystem, climate, or continent; data not shown). We interpreted this lack of difference as a negligible effect of P fertilization on the relative abundance of N2 fixing plants in the ecosystem over the course of the experiments. Consequently, we assumed that the entire data set was representative of short term responses of ecosystems to P supply. Medium- and long-term effects of P supply on N2 fixers abundance remained untested. In the same way, there was no clear difference between field studies and pot studies (Table S3 in the supporting information C).
 Our main aim was to understand if and how P availability limits global SNF. We achieved this goal by compiling published results on SNF under different P supply. To our knowledge, the present study is the first attempt to link at the global scale SNF of terrestrial plants with P supply.
4.1 Direct and Indirect Effects of P Bioavailability on SNF
 The analysis of plant response to P bioavailability in experiments based on a nutritive solution confirmed that it is physiologically possible for a wide range of symbiotic N2 fixers (from diverse ecosystems) to demonstrate both direct and indirect limitation of SNF when the PO4 concentration in the nutritive solution was extremely low [Sa and Israel, 1991]. High P deficiencies, e.g., at [PO4] lower than 5–42 μM, had direct effects on SNF. Concretely, both the nodule:plant biomass ratio and the specific fixation rate of nodules were reduced in conditions of severe PO4 scarcity. With increasing PO4 concentrations in the nutritive solution, the plant fixation rate increased considerably up to an optimum plateau. The shape of the response curve of the specific fixation rate of plants to the concentration of PO4 in the nutritive solution is common to the terrestrial plant species we studied but also appears to be applicable to aquatic higher plants [Kitoh and Shiomi, 1991; Sah et al., 1989; Watanabe and Cholitkul, 1990] and free-living cyanobacteria [Liengen, 1999].
 It could thus be concluded from data based on experiments using a nutritive solution that P availability can constrain SNF through both direct (e.g., plant fixation rate) and indirect effects (e.g., plant biomass). McLaughlin et al.  argued that possible direct effects of P bioavailability on SNF should be also visible through the N content of plants. This hypothesis is in line with our results, which showed a +42% increase in shoot N content. Two interpretations may explain direct P limitation of SNF. First, extreme P scarcity may disturb nodulation or nodule functioning [O'Hara, 2001] which in turn decreases plant fixation rate. From this point of view, direct limitation must be related to a disturbance of plant physiology [Sa and Israel, 1991]. Alternatively, it could be stated that high P deficiency does not physiologically disturb nodules but that plants adopt a particular strategy to face very unbalanced N-P limitations. In this frame, it would be more profitable for plants to allocate their internal resources for the acquisition of the most lacking nutrient (here P) through root growth rather than spending these valuable resources in a process of N acquisition (here SNF) which is more expansive than N uptake by roots [Bloom et al., 1985; Wang et al., 2007]. Our data set was not suitable to determine what really happens in N2 fixing plants under extreme P limitation, and this question requires specific research. Indirect P limitation of SNF through the limitation of plant biomass appeared much easier to interpret. In this case, plants are both moderately N and P limited. Plants are P limited as demonstrated by the fertilization effect of increasing PO4 concentration. At the same time, plants are N limited; otherwise, the plant fixation rate would have been low [Vitousek and Field, 1999; Voisin et al., 2002; Wang et al., 2007], at least for facultative N2 fixing species [Menge et al., 2009], which was not the case for most of the studies.
 The analysis of plant response to P bioavailability in soil experiments revealed similarities but also discrepancies with nutritive solution experiments. Like in experiments using a nutritive solution, plants grown in soil increased their biomass and N2 fixation flux under improved P supply. But, contrary to the plants used in the former experimental approach, their fixation rate remained constant in the majority of cases. P availability mostly influenced the SNF flux only by modifying plant biomass (Figure 5). Our results were consistent with the hypothesis of McLaughlin et al.  as plant N content remained almost unchanged in the soil experiments we compiled (i.e., +6%). Our results were both valid at the plant scale and at the ecosystem scale, suggesting that plant biomass and fixation flux are linearly linked when comparing ecosystems of the same type (e.g., croplands) from different sites. This hypothesis is supported by large-scale studies which already showed that SNF flux was strongly and linearly influenced by plant biomass [Herridge et al., 2008; Unkovich et al., 2010]. What our results showed in addition is that the variability of plant fixation rates observed in these large-scale studies was probably not caused by different levels of P availability in soils. Variations in the specific fixation rate of plants in the field was probably the result of other factors, like plant N status [Hartwig, 1998; Rastetter et al., 2001; Schulze, 2004; Soussana and Tallec, 2010; Voisin et al., 2002] rather than the level of P supply. There were, however, some noticeable exceptions to this linear relationship. In some soil experiments, the plant fixation rate was sometimes also directly constrained by P availability. The soils used in these field trials were weathered soils (mainly ultisols and oxisols).
4.2 Symbiotic Nitrogen Fixation Linked With P Availability in Soils
 The discrepancy between the results presented here (i.e., experiments based on nutritive solutions versus experiments using soils) is in line with the occasional controversy reflected in the scientific literature [e.g., Israel, 1987; Robson et al., 1981]. The discrepancy would be explained if soil available P was most of the time high enough to enable a high fixation rate [Cavard et al., 2007]. We could not test this hypothesis because soil P availability data were very scarce, and the few studies in which soil P availability was measured used different soil P extraction methods. Nevertheless, this hypothesis is consistent with current knowledge on global soil chemistry and plant strategies for nutrient acquisition:
 Our global analysis of SNF soil experiments showed that only plants growing on weathered soils, like ultisols or oxisols, might display a direct effect of P fertilization on plant fixation rate. These soil classes are well known to be severely P deficient. More precisely, ultisols and oxisols are likely to be in average the soil classes with the lowest P content [Yang et al., 2013] and among the soils with the highest P buffering capacities. It means that soil available P content might be low enough to induce a decrease of the plant fixation rate in certain soils, like some oxisols, but that most soils would be in general above this theoretical threshold.
 Strategies of plants to acquire nutrients may also help explaining the apparent discrepancy between nutritive solution experiments and soil experiments. Most plants are adapted to limiting resources [Bloom et al., 1985]. More specifically, N2 fixing species can develop compensatory strategies [Vance, 2001], for example, P mining strategies. The production of phosphatases is higher in N2 fixing species than in other plants [Houlton et al., 2008; Venterink, 2011] and is enhanced when the PO4 concentration in the soil solution is low [Kouas et al., 2009; Venterink, 2011], for instance, below 40 μM [Araujo et al., 2008]. In addition, N2 fixing species can benefit from soil heterogeneity. Nutrient availability in soils varies both in space and over time [Johnson et al., 2010; McClain et al., 2003]. High concentrations of soluble PO4 can be observed over short periods [van der Salm et al., 2009] or in fertile patches [Jonard et al., 2009]. N2 fixing plants can take advantage of such heterogeneities [Gentili and Huss-Danell, 2003] to store P and subsequently use their reserves to maintain a high fixation rate when P is deficient [Hogh-Jensen et al., 2002; Sah et al., 1989; Teixeira et al., 1999], particularly in nodules which are the plant compartment where P content is best maintained [Le Roux et al., 2006; McLaughlin et al., 1990] (Table 1). We assume that all these strategies developed by N2 fixing species enable them to maintain a sufficient P supply in most environments, except the very poorest soils in P.
 We conclude that our results are consistent with the literature and thus reconcile previous conflicting studies as they indicate that all the statements regarding the effects of P availability on SNF are valid but depend on the experimental conditions concerned: P scarcity in the nutritive solution can have a direct negative effect on the plant fixation rate, but soil available P may be often high enough to enable a plant fixation rate that is close to optimal. On the other hand, this pattern was not validated in the most P deficient soils of the world, e.g., certain oxisols, where the plant fixation rate may be moderately depressed. Beyond its possible, but minor, direct effects on plant fixation rates, soil P availability predominantly constrained SNF flux via the limitation of plant growth (Figure 5), mainly under warm or temperate climates, indicating that effect of P availability on plant growth was the main driver of the SNF flux at the global scale.
4.3 Toward a World Map of SNF Limitation by P: Uncertainties and Need for Further Research
4.3.1 Direct P Limitation of SNF
 Ultisols and oxisols orders are known to have intrinsic low total P contents [Smeck, 1985; Yang et al., 2013] along with a high P buffering capacity. From our results, it appeared that ecosystems with this kind of weathered and P deficient soils which are not fertilized may display such a low level of P bioavailability that N2 fixing plants cannot acquire enough PO4 for the correct functioning of their nodules, despite a wide range of compensatory strategies [Vance, 2001]. However, some points require further investigation before a reliable map of areas subjected to direct limitation can be drawn:
 First, there is a wide range of P soil contents within each USDA soil order. Larger differences within soil profiles of the same order than among soil orders are frequently reported [Smeck, 1985; Yang et al., 2013]. Indeed, the total P content of soils initially depends on the composition of the parent material and the geological context may vary considerably for the same kind of soil order [Jenny, 1941; Smeck, 1985].
 There is a huge methodological heterogeneity regarding the practical quantification of the soil pool in available P [e.g., Condron and Newman, 2011; Morel et al., 2000], making it difficult to build a global and coherent database of soil available P.
 The fertilization regime of the soil has to be taken into account. Indeed, P pools of soils are strongly dependent on their P fertilization history [Sharpley et al., 2004]. Unfortunately, fertilization regimes are rarely well documented at the subnational scale.
 Finally, the beginning of the plateau section of the relationship between the fixation rate of plants and soil available P remains unknown and may vary to a certain extent with the plant species, soil order, or with environmental conditions.
 Despite these uncertainties, some general trends can be identified based on the global distribution of soils, geology, land management, and fertilization statistics. Looking at soil orders distribution at the globe scale (see map at http://soils.usda.gov/use/worldsoils/mapindex/order.html), it appeared that soils poor in P (ultisols, oxisols, and to a lesser extent vertisols) are mostly in tropical or subtropical regions (latitudes < 30–40°). However, from a geological point of view, Asian soils are much younger, and somewhat more fertile, than soils in South America or Africa [Huston, 2012; Huston and Wolverton, 2009]. Therefore, Asian soils might be not so P deficient. In addition, most Asian soils, and some South American soils, are fairly well supplied with P fertilizers, while soils are P depleted in many regions of Africa [Bouwman et al., 2009; MacDonald et al., 2011; Sattari et al., 2012]. In terrestrial ecosystems subjected to repeated biomass harvests, like croplands, not fertilizing would cause a very negative input-output budget of P which would exacerbate the natural P deficiency of some soils [Gilbert, 2012]. Based on these patterns, we speculate that cases of direct P constraining of SNF mainly occur in certain type of soils, mostly in tropical regions of Africa—and to a lesser extent in South America—and probably under cropland land use.
4.3.2 Limitation of SNF Through Limitation of Plant Growth
 The global distribution of indirect P constraining of SNF is mainly related to the question of plant growth limitation by nutrients [Fisher et al., 2012]. It is commonly accepted that soils at low latitudes are often P limited whereas other regions are N limited or NP colimited [Wang et al., 2010]. However, many reports suggest that the real pattern is much more complex [Cleveland et al., 2011; Elser et al., 2007; Harpole et al., 2011], particularly in tropical regions where different nutrient colimitation may occur [Baribault et al., 2012; Townsend et al., 2011; Vitousek and Sanford, 1986; Wright et al., 2011; Wurzburger et al., 2012]. Moreover, the nutrient status of soils may be not stationary. In certain industrial regions of the world, N—but not P—is continuously deposited from the atmosphere at high rates [Dentener et al., 2006; Elser, 2011; Mahowald et al., 2008; Okin et al., 2004]. In regions like Europe, Eastern North America, or Eastern Asia, previous N limitation might be replaced by a P limitation [Braun et al., 2010; Gradowski and Thomas, 2008; Thayer et al., 2008]. Therefore, P limitation of vegetation growth, and hence SNF, is likely to be an important and widespread process [Elser et al., 2007], except at high to very high latitudes where P limitation seems to be of low magnitude [Elser et al., 2007; Wang et al., 2010]. Unfortunately, the spatial distribution of P limitation in terrestrial ecosystems remains, up to now, quite unclear [Fisher et al., 2012; Yang et al., 2013] and should be investigated before attempting to draw any global map of SNF limitation by P.
 The contribution of our study was to show that (1) direct limitation of SNF by P is negligible, except in certain terrestrial ecosystems characterized by particularly P deficient soils, and (2) global SNF is mostly indirectly limited by P availability through plant productivity. The influence of P availability on global SNF flux is primarily dependent on the capacity of ecosystems N2 fixing vegetation to grow in more or less P depleted environments. However, the process by which the production of plant biomass, including N2 fixing species, is limited by nutrients remains poorly defined at the global scale [Fisher et al., 2012]. Our study thus confirms that improving our knowledge of soils and understanding the role of P in the global productivity are two of the main challenges that will face the scientific community in coming decades [Korner, 2011; McNeill and Winiwarter, 2004; Tilman et al., 2002]. In particular, it remains unclear to which extent terrestrial ecosystems are P limited and what are the determinants of this limitation within a defined region. In our opinion, large-scale and multidisciplinary studies [e.g., Quesada et al., 2012] are necessary to address this issue.
 We sincerely thank Xavier Cavard, Jean-Christophe Domec, Hervé Jactel, Alain Mollier, Christian Morel, Sylvain Pellerin, Etienne Saur, and particularly Mark Bakker and André Schneider for their fruitful comments, advice, and encouragement. We also thank Daphne Goodfellow and Simeon Smaill for revising the English. Finally, we are grateful for the very pertinent suggestions made by the Editor and three anonymous reviewers.