Ecosystem effects of CO 2 concentration : evidence from past climates

Atmospheric CO2 concentration has varied from minima of 170–200 ppm in glacials to maxima of 280– 300 ppm in the recent interglacials. Photosynthesis by C 3 plants is highly sensitive to CO 2 concentration variations in this range. Physiological consequences of the CO 2 changes should therefore be discernible in palaeodata. Several lines of evidence support this expectation. Reduced terrestrial carbon storage during glacials, indicated by the shift in stable isotope composition of dissolved inorganic carbon in the ocean, cannot be explained by climate or sea-level changes. It is however consistent with predictions of current processbased models that propagate known physiological CO 2 effects into net primary production at the ecosystem scale. Restricted forest cover during glacial periods, indicated by pollen assemblages dominated by non-arboreal taxa, cannot be reproduced accurately by palaeoclimate models unless CO2 effects on C3-C4 plant competition are also modelled. It follows that methods to reconstruct climate from palaeodata should account for CO 2 concentration changes. When they do so, they yield results more consistent with palaeoclimate models. In conclusion, the palaeorecord of the Late Quaternary, interpreted with the help of climate and ecosystem models, provides evidence that CO 2 effects at the ecosystem scale are neither trivial nor transient.


Introduction
Atmospheric CO 2 concentration has varied in a quasicyclical manner from minima of 170-200 ppm in glacials to maxima of 280-300 ppm in the recent "warm" interglacials, varying predictably with Antarctic temperature variations through the past 0.8 million years (Siegenthaler et al., 2005; Correspondence to: I. C. Prentice (colin.prentice@bristol.ac.uk)Lüthi et al., 2008).Atmospheric CO 2 concentration (c a ) is a limiting factor for the photosynthesis of C 3 plants even at today's elevated values (>380 ppm), and was much more strongly limiting at glacial values (Polley et al., 1993(Polley et al., , 1995;;Beerling and Woodward, 1993;Cowling and Sage, 1998;Guiot et al., 2001).Free Air Carbon dioxide Enrichment (FACE) experiments have shown that an increase of c a by 200 ppm increases net primary production (NPP) in temperate forests by 23±2% (Norby et al., 2005).The response of photosynthesis to CO 2 in C 3 plants is a consequence of both substrate (CO 2 ) limitation and competition from O 2 at the reaction site on Rubisco, the enzyme reponsible for CO 2 fixation.Plants using the C 4 photosynthetic pathway are less strongly influenced by c a because they are anatomically and physiologically adapted to low c a , using mechanisms that concentrate CO 2 near the chloroplasts.Because CO 2 concentration affects C 3 photosynthesis, and must indirectly influence the competition between C 3 and C 4 plants (e.g. between C 3 trees and C 4 grasses in tropical savannas), it makes sense to look for CO 2 effects that might be superimposed on climate change effects in palaeoecological records.Moreover, if these variations in c a have caused changes that are detectable in compositional data, such as pollen assemblages, then conventional approaches to reconstructing past climate using statistical or analogue methods -if applied to periods with c a different from that of the late Holocene -are certain to yield incorrect results.Although this potential problem in palaeoclimate reconstruction has been known in principle for more than two decades (Solomon, 1984;Idso, 1989;Farquhar, 1997;Street-Perrott, 1994;Street-Perrott et al., 1997;Cowling andSykes, 1999, 2000;Bennett and Willis, 2000;Williams et al., 2000;Loehle, 2007), until recently few systematic attempts have been made to rectify it.
The relative neglect of CO 2 effects in Quaternary palaeoecology may have been encouraged by an influential school of thought in contemporary biogeochemistry, which questions the relevance of plant-physiological effects of CO 2 over the Published by Copernicus Publications on behalf of the European Geosciences Union.
long term and at the ecosystem scale (e.g.Körner, 2000).A much-debated hypothesis suggests, in particular, that limitations in the supply of nitrogen needed to support increased plant growth should over time reduce or eliminate any effect of c a on NPP (Luo et al., 2004).However, clear evidence in support of this "progressive nitrogen limitation" (PNL) hypothesis has not emerged to date (see e.g.Moore et al., 2006).Equivocal results from a single FACE experiment in a mature forest have been interpreted as indicating a limited or nonexistent CO 2 fertilization effect in mature forests generally (Körner et al., 2005;see Norby et al., 2005 for a critique).Interpretations of experimental data have tended to emphasize the influence of N limitation on the CO 2 effect (e.g.Nowak et al., 2004).Nevertheless, it is well established that elevated CO 2 can increase NPP, even in ecosystems where N supply is demonstrably limiting to plant growth (e.g.Lloyd andFarquhar 1996, 2000;Nowak et al., 2004).There is some evidence that plants can increase their N supply to support CO 2enhanced growth, perhaps by increased root penetration or increased labile carbon subsidy to the rhizosphere (Finzi et al., 2007).The extent to which ecosystems can respond to CO 2 enhancement over timescales longer than a decade has not been unambigously demonstrated by experiments, and is subject to our still incomplete quantitative understanding of the mechanisms of N acquisition by plants.
Controversy thrives in this field in part because the evidence base from contemporary studies is, inevitably, limited.FACE has provided a great deal of extremely valuable information, and remains the key experimental technology needed to unravel CO 2 effects in intact ecosystems.However, FACE experiments are expensive and technically challenging, especially in forests.No feasible experiment can test the multidecadal responses of ecosystems and biomes on a large spatial scale.In this paper, we show that long-term, ecosystemand biome-level effects of CO 2 effects on plant physiology can be inferred from the palaeorecord.We argue that CO 2 effects are fundamental in establishing consistency between palaeovegetation data and palaeoclimate models.

Background
The concentration of CO 2 in the substomatal cavity, or "internal" CO 2 concentration (c i ), is a key quantity for photosynthesis in C 3 plants.The internal concentration in illuminated leaves is less than the ambient concentration, because photosynthesis draws down CO 2 while the stomata present a resistance to the inward diffusion of CO 2 .The relationship between photosynthesis and the CO 2 concentration gradient across the leaf epidermis is represented by the diffusion equation, where A is the net rate of carbon assimilation and g is the stomatal conductance (the reciprocal of resistance) to CO 2 .
Stomatal conductance is regulated in a way that maintains the c i /c a ratio typically around 0.7-0.8 in C 3 plants and 0.3-0.4 in C 4 plants under conditions of moderate vapour pressure deficit (vpd) and adequate soil moisture (Wong et al., 1979).Diffusion through the stomata also controls plant water loss: where E is the rate of transpiration per unit leaf area, and D is the vpd at the leaf surface.As D increases, g declines; in consequence, under conditions of increasing D, c i /c a is reduced while E increases towards a maximum (Monteith, 1995).Equation ( 1) describes the control of c i by A. At the same time, A is controlled by c i , according to the following equation which summarizes the biochemical controls of photosynthesis (Farquhar et al., 1980: simplified here for expository purposes): where: Here A c is the Rubisco-limited photosynthetic rate, V c max is a maximum rate (dependent on the activity of Rubisco), is the CO 2 compensation point (the concentration at which photosynthesis is zero), and K is an effective Michaelis-Menten coefficient (dependent on O 2 concentration, but this has not varied significantly over the time scales considered here).A j is the light-limited photosynthetic rate, φ o is the intrinsic quantum efficiency of photosynthesis, and I is the absorbed flux of photosynthetically active radiation (PAR).R d is the respiration rate required to maintain the activity of Rubisco and other photosynthetic enzymes; b is a small constant, representing a respiratory loss of 1-2% of V c max .
Leaves typically operate with values of V c max such that typical daytime values of A c and A j are similar, i.e. there is approximate co-limitation by Rubisco and PAR (Farquhar et al., 1980).Co-limitation yields the optimum assimilation rate, because a lower V c max would result in reduced utilization of available PAR while a higher V c max would increase the loss of carbon in maintenance respiration for no gain in photosynthesis (e.g.Haxeltine and Prentice, 1996a).In practice, with light and other environmental conditions varying over the diurnal cycle, photosynthesis can be limited by Rubisco at some times and by PAR at others.However, both rates are characterized by a response to c i that increases most steeply just above c i = , and approaches an asymptote at high c i .Thus, the response of A to c a is steepest at low c a and approaches saturation at high c a .If V c max is optimal, increased c a should lead to reduced V c max (this "downregulation" of V c max has been observed widely in raised-CO 2 experiments: see e.g.Ainsworth and Long, 2005).Low c a , similarly, should lead to increased V c max .Down-regulation of V c max at high c a would tend to reduce the plant demand for N, while increasing A would tend to increase it.Allocation of carbon to fine roots is increased at high c i , suggesting an adaptive response to an overall increased N demand (e.g.Palmroth et al., 2006).Provided that N demands are met, increased c a should lead to increased NPP, and reduced c a should lead to reduced NPP, although changing allocation patterns might limit the magnitude of this response.
The asymptotic nature of the relation between A and c i , combined with the conservatism of c i /c a , implies that g should decline with increasing c a .This stomatal response to CO 2 has been observed in many species.It could allow water conservation, further increasing photosynthesis in seasons and climates where soil moisture is limiting.Similarly, g should increase with declining c a .Leaf area index (LAI) under changing CO 2 concentrations is therefore subject to several competing effects.With increasing c a , increasing A should promote increased LAI, and increasing g might be compensated by increased LAI; but increasing below-ground allocation would reduce LAI.A survey of experimental results by Cowling and Field (2003) indicated that LAI generally declines as CO 2 decreases below present values, while at higher CO 2 levels there is no consistent response of LAI to CO 2 .
C 4 plants, with their characteristic CO 2 -concentrating leaf anatomy, are believed to have evolved and spread in response to low CO 2 levels (relative to earlier geological epochs) that developed during the Cenozoic (Cerling et al., 1993;Cowling, 2001;Sage, 2004).The rate of C 4 plant photosynthesis can be roughly approximated by disregarding CO 2 effects on A in Eq. ( 2) and using a reduced value for the quantum efficiency, representing the "cost" of the CO 2 -concentrating mechanism.Using the known temperature dependencies of the various photosynthetic parameters, it can be shown that there is a crossover temperature above which C 4 plants can fix carbon at a faster rate than C 3 plants (Ehleringer et al., 1997).Below this temperature, C 4 plants fix carbon at a slower rate than C 3 plants.The crossover temperature increases with increasing CO 2 concentration.Thus, other things being equal, we would expect C 4 plants to be more competitive relative to C 3 plants at low CO 2 (Cole and Monger, 1994;Ehleringer et al., 1997;Collatz et al., 1998).In today's world, by contrast, C 3 plants should be gaining ground.A widespread trend towards an increase of (C 3 ) tree cover at the expense of (C 4 ) grasses has indeed been observed in tropical savannas, and may be an effect of increasing CO 2 concentration (Bond and Midgley, 2000;Eamus and Palmer, 2007).However, other factors including climate change and grazing intensity have been advanced as alternatives (e.g.Archer et al., 1995).This debate continues, while contemporary observations seem unable to resolve it.
Past environments offer a variety of scenarios involving large and long-lasting CO 2 changes.Although CO 2 changes are inevitably accompanied by climate changes due to the greenhouse effect of CO 2 , the involvement of other climate drivers (orbital variations and ice-sheet growth and decay) implies a degree of decoupling between CO 2 and climate which is potentially useful for attempts to attribute causes.
Here however we concentrate exclusively on the major shifts in CO 2 , climate and vegetation between the last glacial maximum (LGM) and the Holocene, and we use climate and ecosystem modelling results to separate effects of CO 2 from effects of climate on vegetation.In doing so, we briefly survey the history of attempts to depict and model glacialinterglacial variations in the terrestrial biosphere.This endeavour has included some false starts but has now led to a broad consensus, with wide-ranging implications for palaeoclimatology and for carbon cycle science.
3 Glacial-interglacial variations in carbon storage Shackleton (1977) noted that the calcium carbonate shells of subfossil benthic forminifera from the last glacial period are light in 13 C, presumably indicating a change in the stable isotope composition of dissolved inorganic carbon (DIC) in the ocean.His estimate of the δ 13 C offset was −0.7 per mille for the LGM, relative to the Holocene.His hypothesis to explain this offset relies on a transfer of carbon between the land and ocean reservoirs.Terrestrial organic carbon is depleted in 13 C, so a simple mass balance can be used to infer that terrestrial organic carbon storage was substantially reduced in glacial times.
A glacial-interglacial shift in the δ 13 C of DIC has been confirmed, although its global mean value appears to be smaller, in the range of −0.3 to −0.4 per mille (Currey et al., 1988;Duplessy et al., 1988;Sarnthein et al., 1988;Ku and Luo 1992).Using a canonical value of −0.32 per mille, Bird et al. (1994Bird et al. ( , 1996) ) estimated a terrestrial carbon storage reduction of 310-550 PgC at the LGM relative to preindustrial time.This calculation took account of changes in atmospheric CO 2 content and its δ 13 C, as measured in ice cores, and a possible shift in land carbon towards values up to 2 per mille heavier in glacial time (Crowley, 1991).Allowing for further uncertainty in the isotopic shift, Bird et al. (1996) obtained a range of 300-700 PgC.Later estimates have continued to lie within this interval, for example 430-665 PgC (Street-Perrott et al., 1998) and 550-680 PgC (Beerling, 1999), respectively constrained by observations and modelling of the 13 C content of land carbon at the LGM.Ikeda and Tajika (2003) estimated 630 PgC, by assimilating atmospheric CO 2 concentration and surface-and deepwater 13 C records into a box model of the ocean carbon cycle.Köhler and Fischer (2004)  Various sources of bias in this ocean carbon isotope constraint on terrestrial carbon storage have been suggested, including a dependence of isotopic fractionation during shell formation on carbonate ion concentration (Spero et al., 1997;Pedersen et al., 2003) and degassing of CH 4 clathrates with strongly negative isotopic signatures followed by conversion of the degassed CH 4 to CO 2 (Maslin and Thomas, 2003).We do not attempt a critical analysis of these mechanisms; however, none of them is likely to have a large enough effect to overturn Shackleton's hypothesis, or to require substantial revision of the broad range of estimates of LGM terrestrial carbon storage indicated by Bird et al. (1996).In particular, we note that hypotheses seeking to explain glacialinterglacial CO 2 changes by increased carbon storage during glacial times, e.g.buried under ice (Zeng, 2003) or in permafrost soils (Zimov et al., 2006), are manifestly incompatible with the ocean carbon isotope constraint.
Reduced carbon storage on land cannot be accounted for simply by the presence of continental ice sheets, because the area of exposed continental shelf -mainly in the tropics -roughly balanced the land area occupied by ice (Prentice et al., 1993;Montenegro et al., 2006).The explanation requires a change in the distribution of terrestrial biomes, and/or their carbon content.During the early 1990s, many independent estimates of terrestrial carbon storage change between glacial and interglacial regimes were made using "book-keeping" methods that assign fixed vegetation and soil carbon storage per unit area to each biome.Such estimates were based either on cartographic reconstructions of past vegetation (e.g.Adams et al., 1990;van Campo et al., 1993;Crowley, 1995;Maslin et al., 1995;Adams and Faure, 1998; and numerous regional studies not cited here), or on palaeoclimate and biogeography modelling using the simple models available at that time (Prentice and Fung, 1990;Friedlinsgtein, 1992;Prentice, 1993;Prentice et al., 1993;Prentice and Sykes, 1995).All of these estimates should now be rejected, because (a) they overlook variations in carbon storage within biomes, which are large today, and might well vary systematically between climate states; (b) they exclude a priori any effect of CO 2 concentration on carbon storage; and (c) the data-based estimates, in particular, offer no consistent way to deal with pollen assemblages that lack modern analogues.
The first modelling studies on this topic produced no reduction in carbon storage at the last glacial maximum (LGM) (Prentice and Fung, 2000), or too small a reduction (e.g.Friedlingstein, 1992;Prentice et al., 1993).This last study invoked peatland development as an additional cause of net postglacial carbon accumulation, but this now seems implausible: there is abundant evidence for glacialage peats in tropical lowlands, including the exposed continental shelf, which may have stored on the order of 200 PgC that has largely been removed subsequently (Faure et al., 1996).Esser and Lautenschlager (1994), Peng et al. (1995Peng et al. ( , 1998) ) and Friedlingstein et al. (1995) applied early processbased land carbon cycle models to simulate changes in carbon storage since the LGM.These studies noted the potential of physiological CO 2 effects to further reduce carbon storage, and used empirical formulations to demonstrate that the impact could be substantial.Friedlingstein et al. (1995), for example, obtained a total reduction in carbon storage of 470 PgC when the CO 2 effect was included.They also showed that the near-zero carbon change previously simulated by Prentice and Fung (1990) was an artefact, caused by an uncorrected bias in the model used to simulate the LGM climate.Thus, although much was made in older literature of a supposed discrepancy between a small or zero shift in carbon storage (based on models) and a very large shift (based on observations), neither the model-based or the observationally-based estimates can now be considered wellfounded.
More recent studies have exploited the advances in palaeoclimate and biosphere modelling made since the 1990s.The current standard approach starts with a palaeoclimate simulation made with a general circulation model (GCM), and applies anomalies (differences in mean monthly climate variables between the simulation and the GCM's control run) to correct a baseline climatology; the altered climate then drives a terrestrial biosphere model (e.g.Franc ¸ois et al., 1998).Current models calculate the CO 2 effect based on the Farquhar equations to give changes in photosynthesis, which are propagated through carbon allocation and plant competition algorithms to generate effects at the ecosystem level.Snapshot analyses (e.g.comparisons between the LGM and preindustrial time slices) can be made using equilibrium models such as CARAIB (Warnant et al., 1994); transient analyses apply dynamic global vegetation models such as LPJ (Sitch et al., 2003).A consensus finding has emerged from these analyses, namely that a carbon shift of approximately the right magnitude can be reproduced -but only if physiologically mediated effect of CO 2 are included (Franc ¸ois et al., 1998(Franc ¸ois et al., , 1999;;Otto et al., 2002;Kaplan et al., 2002).These recent analyses have shown consistently that the carbon storage reduction (LGM minus pre-industrial Holocene) with the CO 2 effect "turned off" either is too small, or in some cases even has the wrong sign, i.e. the terrestrial biosphere is modelled to have slightly (<100 PgC) greater carbon storage at the LGM than during the Holocene.The result of Franc ¸ois et al. ( 1998) is typical: producing a carbon storage reduction of 610 PgC when the physiological effect of CO 2 is included, but only 160 PgC when it is not.
There is a simple reason why climate change alone is not sufficient to account for the glacial-interglacial change in terrestrial carbon storage.Other things being equal, the (dominant) soil component of terrestrial carbon storage increases as global temperatures decline, due to slow decomposition of soil organic matter.This explains, for example, why the highest soil carbon storage today occurs in the boreal zone and not in the tropics.This response is steeper than the positive response of NPP to warming, over a wide range of temperatures.It is a key component of the feedback mechanism believed to be responsible for the small reduction in atmospheric CO 2 content during the Little Ice Age (Joos et al., 2004a;Cox and Jones, 2008).Gerber et al.'s (2003) equilibrium sensitivity analysis with the LPJ model suggested that global cooling to LGM levels would have had a minor impact on total terrestrial carbon storage, whereas the CO 2 change would have reduced terrestrial carbon storage substantially.Kaplan et al. (2002) used LPJ to perform transient simulations of land carbon storage changes since the LGM.They obtained an increase of 820 PgC after the LGM, mainly due to the CO 2 effect.Without the CO 2 effect, the simulated increase was only 210 PgC (Joos et al., 2004b).The increase implies a net extraction of carbon from the ocean to build biospheric carbon on land.This in turn triggered the carbonate compensation mechanism in the ocean, and if the larger figure for post-LGM carbon storage increase is accepted, this mechanism accounts for the greater part of the observed rise in atmospheric CO 2 concentration during the pre-industrial Holocene (Joos et al., 2004b).
The carbon isotope palaeorecord of deep ocean water as preserved in benthic foraminiferal shells, combined with process-based modelling studies to separate the climate and CO 2 effects, thus contains two important messages for contemporary biogeochemistry.(1) The 100 ppm increase in CO 2 concentration from the last glacial period to the Holocene had a major, long-lasting effect on NPP and carbon storage.(2) The approximate magnitude of this effect can be predicted (within the uncertainties of the models and the ocean carbon isotope constraint), but only by models that propagate the physiological effect of CO 2 on photosynthesis into NPP.The earlier proposal by Prentice and Sarnthein (1993) -that climate-induced biome shifts alone might be sufficient to explain glacial-interglacial changes in carbon storage -can no longer be supported.The available evidence on glacial-interglacial changes in terrestrial carbon cycling indicates an important role for physiological effects of CO 2 .

Biome shifts
The reconstruction of glacial-interglacial changes in terrestrial carbon storage is indirect and therefore subject to rather large uncertainties, as discussed above.The reconstruction of biome shifts is more directly linked to the extremely rich set of observations in the form of pollen and plant macrofossil records from terrestrial sediments.Here, too, current evidence supports a major role for changes in CO 2 .Although climate change can certainly influence biome distribution, physiological CO 2 effects modify the growth and competition of different plant functional types (PFTs) and thereby are expected to influence vegetation composition, LAI, structure and biome boundaries (Cowling, 1999(Cowling, , 2004;;Bond et al., 2003;Cowling and Shin, 2006).Jolly et al. (1997) used BIOME3, a process-based coupled biogeochemistry-biogeography model (Haxeltine and Prentice 1996b), to show that CO 2 changes could profoundly affect montane vegetation zonation in the African tropics.BIOME3 is a forerunner of LPJ that lacks only the transient vegetation dynamics which LPJ simulates; it mimics the response of LAI and NPP of different PFTs to climate and CO 2 , and competition among PFTs.Jolly et al. (1997) were the first to use a process-based model to analyse the effects of changing CO 2 concentration on ecosystem composition and structure in a palaeoecological context.Their analysis indicated that the large elevational extension of the heath belt on East African mountains was a predictable consequence of the low CO 2 concentration in glacial time.Indirect support for this finding came from 13 C measurements on leaf waxes preserved in the sediments of a high-elevation lake in the region, which indicated a marked shift towards C 4 plant dominance during the last glacial period (Street-Perrott et al., 1997;Huang et al., 1999).Boom et al. (2002) produced similar results for South America, and used BIOME3 to derive a function relating C 4 plant abundance (as indicated by leafwax 13 C measurements) to temperature and CO 2 concentration which they then inverted to yield a proxy CO 2 record that is broadly consistent with ice-core measurements -a further, indirect confirmation of the role of CO 2 in controlling C 3 /C 4 plant competition over glacial-interglacial cycles.
Some early discussions had assumed that effects of glacial CO 2 concentration would be greater at higher elevations, because of low atmospheric pressure (implying a low partial pressure of CO 2 for the same concentration).However the elevation effect also lowers the partial pressure of O 2 , which competes with CO 2 at the Rubisco reaction sites; this effect counters the hypothesized effect of a low partial pressure of CO 2 at high elevations (Terashima et al., 1995).In fact, the effects are expected a priori to be greater in warmer climates because the O 2 competition effect is stronger at high temperatures, as reflected in the temperature dependence of the photosynthetic parameters (Cowling, 1999a).Indeed there is abundant evidence in pollen and carbon isotope records from tropical lowland regions for shifts away from forest, toward C 4 dominated vegetation, during glacial times.This evidence comes principally from records of the δ 13 C changes in lake sediments (e.g.Talbot and Johanessen, 1992;Giresse et al. 1994), which can be related to changes in the relative abundance of C 3 and C 4 plants.Modelling with BIOME3 has also indicated consequences for the structure of tropical forests ecosystems which would be largely "silent" in terms of the palaeorecord, yet could be profoundly significant for biogeography (Cowling et al., 2001).Harrison and Prentice (2004) quantified CO 2 effects at a global scale, based on the BIOME 6000 synthesis, which compiled mid-Holocene and LGM pollen records worldwide and assigned a biome to each record using a standardized method (Prentice et al., 2000).The effect of LGM climate was accounted for by using all of the available GCM www.clim-past.net/5/297/2009/Clim.Past, 5, 297-307, 2009 experiments from the Palaeoclimate Modelling Intercomparison Project (PMIP) to drive BIOME3.The results were unambiguous.Whichever GCM was used to simulate the LGM climate, the extent of simulated global forest reduction fell short of that observed when CO 2 effects were neglected; but became closer to observations when these effects were included.The influence of CO 2 was largest and most consistent in the tropics, but not confined there: the same effect was seen in the northern and southern extratropics as well.Thus finding is consistent with that of Cowling (1999b), who had also used BIOME3 to show that the pattern of LGM vegetation in eastern North America can be explained satisfactorily only through consideration of reduced water use efficiency (A/g) by C 3 plants in low CO 2 , favouring more droughttolerant plant types, as well as climate change; a point also taken up by Loehle (2007).
The conclusion from these studies is that physiological CO 2 effects, as simulated by models based on the fundamentals of photosynthesis and propagated into the simulation of LAI and NPP of different PFTs, are essential in order to fully account for global shifts in forest cover, and probably also to explain a wider range of changes in the relative abundances of PFTs, between glacial and interglacial regimes.

Climate reconstruction
Until recently, methods to reconstruct past climates from palaeovegetation data were all statistical, based on the assumption that the climate controls on plant distribution are invariant.These methods broadly fall into two families: those based on some form of regression, and those based on modern analogues -either using a direct search among a set of analogues, or using response surfaces to fit the empirical relationship between the abundances of pollen taxa and climate variables (see e.g.Brewer et al., 2008).It has occasionally been noted that physiologically mediated CO 2 effects could compromise climate reconstructions made by such methods (e.g.Idso, 1989;Cowling and Sykes 1999), but this observation had no impact on research practice until Guiot et al. (2000) developed a novel approach to palaeoclimate reconstruction, based on the numerical inversion of BIOME3.
In inversion-based palaeoclimate reconstruction, a climate change (between the past time under consideration and recent times) is selected using a search algorithm that seeks to maximize goodness of fit between the palaeodata and simulated ecosystem composition (in terms of simulated abundances or productivity of PFTs).A major advantage of using inversion is that it allows the CO 2 level to be accounted for, as it can be prescribed to the model independently of the palaeovegetation data.Guiot et al. (2000) found that prescribing the correct (low) CO 2 made a substantial difference to LGM climate reconstructions.A key finding was that when low CO 2 was prescribed for the LGM, reconstructed winter con-ditions in the Mediterranean region became (a) warmer than in earlier, analogue-based reconstructions, and (b) systematically closer to the predictions made by climate models for the LGM (Ramstein et al., 2007).This result is all the more remarkable because BIOME3 does not model any direct effect of CO 2 on plants' low-temperature tolerance.However, low atmospheric CO 2 implies both reduced NPP (so that the distribution of trees towards climates offering low potential production becomes more restricted) and increased water use per unit of NPP, reinforcing this restriction.As a consequence, steppe vegetation can expand under low CO 2 at the expense of forest.The earlier, analogue-based methods has selected modern analogues for the glacial steppe in central Asia in climates with very cold winters and short growing seasons.Low CO 2 , however, permitted the occurrence of steppe vegetation under milder conditions.An expanded set of analogues produced milder reconstructed winters than the original set, but could not produce consistency with palaeoclimate model results.This case study illustrates a relatively littlediscussed problem with analogue methods.Although the "no-analogue" problem (fossil pollen assemblages for which similar modern pollen assemblages cannot be found) is well known, there is also potentially a "wrong-analogue" problem whereby the method of modern analogues selects similar pollen assemblages that actually originated in very different physical environments.The Mediterranean case study shows how this problem can be alleviated through the inversion of process-based models.The inversion method has been applied in three continents (Wu et al., 2007a, b;Guiot et al., 2008).
The use of inverse modelling to reconstruct past climates has further advantages that are beginning to explored, using BIOME3 and most recently its successor, BIOME4 (Kapan et al., 2003).Inverse modelling provides a natural way to incorporate additional observational constraints, such as palaeo 13 C information (e.g.Hatté and Guiot, 2005;Rousseau et al., 2006;Hatté et al., 2009;Guiot et al., 2009).It makes it straightforward to incorporate additional external forcing of vegetation changes, such as insolation changes caused by orbital variations, and by the differences in latitude between locations at which particular vegetation types occured during glacial versus interglacial climates; these differences may have implications for plant productivity and water use (Kaplan et al., 2003).The involvement of a process-based model should also allow the use of a data-assimilation approach to the reconstruction of ecosystem properties that palaeodata do not directly record, such as NPP and carbon storage (Wu et al., 2009).Finally, a recent development of the inversion approach builds on a version of the LPJ model to allow time-dependent climate reconstruction taking account temporal lags in the response of vegetation to climate (Guiot et al., 2009).What began as a solution to a specific problem in palaeoclimate reconstruction may turn out to be a tool with a wide field of application in palaeoclimatic analysis.

Conclusions
Despite persistent controversies about the contemporary and future effects of rising CO 2 , surprisingly few attempts have been made to use palaeorecords to help resolve them.We have summarized evidence based on a model-assisted interpretation of the palaeorecord, which supports the idea that physiological effects do scale up to ecosystem effects, through changes in primary production and through competition between plants with different photosynthetic pathways.
This palaeoperspective has practical implications for modelling of the contemporary carbon cycle.Atmospheric O 2 and 13 CO 2 concentration measurements show that the land biosphere has been taking up a part of the CO 2 emitted by human activities during recent decades (Prentice et al., 2001;Denman et al., 2007).The land biosphere, in other words, is a net sink for CO 2 due to some process which outweighs the CO 2 release due to land-use change.Process-based models consistently identify this process as a consequence of rising CO 2 concentration, acting through the physiological effect of CO 2 on photosynthesis (McGuire et al., 2001).Without this, models would predict a higher rate of atmospheric CO 2 increase than is observed.Models also predict that this uptake will increase while CO 2 continues to rise, as expected during this century, although it will eventually decline again due to the asymptotic nature of equation ( 3) and competing effects of warming on the terrestrial carbon balance (Cramer et al., 2001).The palaeorecord suggests that the assumptions underlying existing models regarding CO 2 effects are broadly correct.We do not find support for the opinion (e.g.Körner, 2000) that other constraints effectively eliminate the ecosystem-level effects of changing CO 2 concentration on carbon storage over long time scales.The palaeorecord also supports the attribution of increases in the woody component of tropical savannas to physiological effects of rising CO 2 , although this does not rule out a contribution of other factors such as land-use change.
Our findings further imply that for palaeoclimate reconstruction involving periods with substantially different CO 2 levels, inversion of process-based models is likely to yield more realistic results than statistical modelling that excludes physiological effects of CO 2 .Recent advances in inverse vegetation modelling offer a promising way forward for the integration of biophysical process understanding into palaeoclimate analysis.

Epilogue
Dominique Jolly made pioneering contributions to Quaternary palaeoecology, especially of Africa.He was closely involved in the development of the data analysis technique called "biomization" which made the global BIOME 6000 project possible, and he developed a vision of how modelling and data analysis could work together to achieve new insights about the past.He also produced the first global-scale simulations of the LGM world that took low CO 2 into account, using an early version of BIOME3.These results were showcased in Berrien Moore III's plenary presentation at the first International Geosphere-Biosphere Programme Congress in 1996, and presented as an example of how different disciplines of global change science could productively collaborate, and of the power of the new process-based approaches to terrestrial biosphere modelling.We have dealt here with a few of the themes to which Dominique contributed important results and insights.We dedicate this article to the memory of an inspiring colleague and friend.
Edited by: J. Guiot Reviewed by: N. Viovy and E. Gritti Publication of this paper was granted by EDD (Environnement, Développement Durable) and INSU (Institut des Sciences de l'Univers) at CNRS.
concentration and δ 13 C of atmospheric CO 2 to constrain a box model of the land carbon cycle.