The Cenozoic era is characterized by large climatic variations, including the fundamentally important transition from an ice-free “greenhouse” planet to the modern “icehouse” planet with polar glaciations. This climatic transition is generally thought to have been driven primarily by changes in pCO2 and/or the thermal isolation of Antarctica by the opening of Southern Ocean gateways (DeConto and Pollard, 2003; Zachos et al., 2008; Hansen et al., 2013; Hren et al., 2013; Goldner et al., 2014; Inglis et al., 2015). However, the full extent of the role of pCO2 in Cenozoic climate change remains unresolved. The most detailed Cenozoic temperature and pCO2 records are derived from marine isotope proxies (e.g. Foster et al., 2012; Pagani et al., 2011; Pearson et al., 2009; Zachos et al., 2001, 2008). Isotope records, however, may be influenced by a variety of taphonomic and diagenetic biases (see Coxall and Pearson (2007) for review; and Pagani et al., 2011), which can obscure the climatic signal, and thus need independent evaluation by separate proxy records (Beerling and Royer, 2011).

Eocene temperatures were globally much higher than today, leading to a weakened Equator-to-pole temperature gradient and muted seasonal cycle compared to today – the so-called “Eocene equable climate problem” (Sloan and Barron, 1992; Huber and Caballero, 2011). Climate modelling has been able to reconstruct this climatic pattern only with excessively high pCO2 levels (∼ 4500 ppm: Huber and Caballero, 2011), but such elevated pCO2 atmospheres do not agree with most proxy records. It has therefore been speculated that Eocene climate sensitivity was elevated compared to today and/or that other forcing in addition to high pCO2 was involved (Caballero and Huber, 2013; Hansen et al., 2013). In order to solve this enigma reliable multiple proxy records of pCO2 are of paramount importance.

The fundamental climatic reorganization that occurred close to the Eocene–Oligocene boundary (33.8 Ma), often referred to as the Eocene–Oligocene transition (EOT, 34–33.5 Ma), had drastic consequences for biological systems. These included both terrestrial and marine faunal and floral extinctions accompanied by evolutionary turnover (Prothero, 1994; Coxall and Pearson, 2007; Sheldon et al., 2009: Kunzmann, 2012; Kvaček et al., 2014), although vegetation changes in the European terrestrial record appear to be less dramatic and more gradual (Kvaček et al., 2014; Kunzmann et al., 2015). General circulation models of Palaeogene climate have shown that continuously declining pCO2, amplified by Milankovitch forcing and ice-albedo feedbacks, could cause significant temperature reduction. This could result in a permanent continental Antarctic ice-sheet once a critical pCO2 threshold, generally considered to be < 700 ppm is crossed (e.g. DeConto and Pollard, 2003; Coxall et al., 2005; Pollard and DeConto, 2005; Zachos and Kump, 2005; Pagani et al., 2011; Hansen et al., 2013). Modelling studies thus indicate that lowering of pCO2 may have been the primary forcer of this cooling transition (DeConto and Pollard, 2003; DeConto et al., 2008). However, detailed estimates for pCO2 for the Eocene and the Oligocene are highly variable and sometimes contradictory or showing unexpected relationships with palaeo-temperature proxy records (see Pagani et al., 2005). For example, comparing the pCO2 record of Pearson et al. (2009: Fig. 1), which is based on measurements of boron isotopes in planktonic foraminifera, and the benthic foraminifera oxygen isotope (δ18O) compilations of Zachos et al. (2008), it is evident that in the late Eocene δ18O-inferred deep ocean cooling coincided with decreasing pCO2. In contrast, there is little evidence of warming in the early Oligocene, despite a surprising initial large increase in pCO2. Overall, the pCO2 and O isotope-based temperature records seem to be (largely) coupled in the Eocene, but decoupled in the Oligocene. Pagani et al. (2011) on the other hand recently published compiled alkenone-based pCO2 records and found declining pCO2 before and during the Antarctic glaciation (EOT and earliest Oligocene) (Pagani et al., 2011: Fig. 4), supporting the role of pCO2 as the primary forcing agent of Antarctic glaciation, consistent with model-derived thresholds. A compounding factor of these discrepancies is that the influence of temperature on ice sheet volume is unconstrained and the influence of temperature versus ice volume on the δ18O record is unresolved, with no proxy identified to isolate ice sheet volume changes, complicating further the interpretation of the climate proxy data sets. Independent proxy records of E–O pCO2 are therefore desirable and may support one or the other of the major prevailing scenarios outlined above, or provide alternative information on Cenozoic climate change.

One of the four proxies that have been identified as being particularly useful for Cenozoic pCO2 reconstructions by the Intergovernmental Panel on Climate Change (initially reported in the Fourth IPCC Report; IPCC, 2007) is the terrestrial proxy based on stomatal densities of fossil plants. Previous studies using the stomatal proxy method of pCO2 reconstructions for the part of the Cenozoic relevant here were, however, mostly of low resolution and have been inconclusive. Some suggested that pCO2 was essentially stable at between 300 and 450 parts per million by volume (ppm) during the Eocene, Oligocene and Miocene (Royer, 2001; Royer et al., 2001; Greenwood et al., 2003; Maxbauer et al., 2014) and others suggesting distinct decrease in pCO2 across the Eocene–Oligocene boundary (Retallack, 2001). More recent studies suggest higher and possibly rapidly decreasing pCO2 (ranging ca. 1000–500 ppm) during the late middle Eocene (Doria et al., 2011; Grein et al., 2011). In this issue, Liu et al. report a “late Eocene” pCO2 from a single stratigraphical level of ca. 390 ppm. However, the chronological range they supply for their pCO2 estimate (42.0–38.5 Ma) falls within the late Lutetian to Bartonian in the Middle Eocene, thus recording an unusually low pCO2 estimate for this time-interval characterized by high temperatures (Liu et al., 2015). Closer to the E–O boundary, one study suggests that pCO2 was significantly higher at the EOT than during the early Oligocene (Roth-Nebelsick et al., 2004) and others that early Oligocene to early Miocene pCO2 was ca. 400 ppm throughout (Grein et al., 2013; Roth-Nebelsick et al., 2014).

Here we present a new stomatal proxy-based record with multiple data points spanning the late middle to late Eocene, two sampling levels that according to current available evidence are from the earliest Oligocene, and two samples from later in the Oligocene.

Stomata are pores on plant leaf surfaces through which gas exchange takes place; i.e. carbon is obtained from CO2 and at the same time water vapour and oxygen are lost by diffusion. An inverse relationship exists between the frequency of stomata and pCO2, as established by Woodward (1987) from observations of herbarium material, showing that modern tree species have responded to the anthropogenic rise in pCO2 by reducing their stomatal frequency significantly. The inverse relationship between stomatal frequency, recorded as “stomatal density” (SD = the number of stomata per mm2) or “stomatal index” (SI = the percentage of stomata relative to epidermal cells), and pCO2 has been repeatedly demonstrated for a wide variety of plant taxa from disparate geological and ecological settings from the Palaeozoic until today and is thus established as a strong proxy for palaeo-pCO2 (e.g. Beerling et al., 1998; McElwain, 1998; Retallack, 2001; Royer et al., 2001; Kürschner et al., 2008; Steinthorsdottir et al., 2011b; 2013: Steinthorsdottir and Vajda, 2015). The increasingly close match between stomatal proxy pCO2 results and independent proxy records, actual pCO2 measurements and in some cases climate modelling (e.g. Finsinger and Wagner-Cremer, 2009; Foster et al., 2012; Kürschner et al., 2008; Retallack, 2001; Rundgren and Björck, 2003; Steinthorsdottir and Vajda, 2015) instils growing confidence in stomatal frequency for recording past pCO2. Strongly supporting the validity of the stomatal proxy is also the identification of the mechanism by which plants control their stomatal densities based on atmospheric pCO2. All plants use the enzyme carbonic anhydrase to detect pCO2 around their leaves (Frommer, 2010; Hu et al., 2010); mature leaves (early shoots) then control stomatal development of younger leaves through long-distance signalling (Lake et al., 2002), involving the HIC gene signalling pathway (Brownlee, 2001; Gray et al., 2000).

In order to transform stomatal frequency data derived from fossil plants into palaeo-pCO2 estimates it is usually necessary to compare stomatal data from present-day plants that are either phylogenetically related or in other ways equivalent to the fossil plants. Nearest living relatives (NLRs) should be used when possible, but when these cannot be identified for the fossil plants, nearest living equivalents (NLEs = present-day species that are of comparable ecological setting and/or structural similarity to their fossil counterpart) may be used instead (McElwain and Chaloner, 1995; Barclay et al., 2010; Steinthorsdottir et al., 2011a, b).

There are three stomatal palaeo-pCO2 calibration methods in use. These are (i) the “stomatal ratio method” (McElwain and Chaloner, 1995; McElwain, 1998), which relies on a ratio between stomatal frequencies of fossil plants and their NLE to semi-quantify pCO2; (ii) the “transfer function method”, which relies on herbarium material and/or experimental data sets for NLR/NLE responses to calculate pCO2 curves (e.g. Beerling and Royer, 2002); and (iii) the more recently developed taxon-independent “mechanistic gas exchange modelling” approach (e.g. Wynn, 2003; Konrad et al., 2008; Franks et al., 2014; Grein et al., 2013; Roth-Nebelsick et al., 2014) which all use measurements of stomatal density and pore size to estimate maximum theoretical gas exchange rates, together with various photosynthetic biochemical traits, and in some cases palaeoenvironmental information, to estimate palaeo-CO2. The stomatal ratio method, which is used here, calibrates palaeo-pCO2 based on two so-called standardizations. The first is the “modern” standardization that assumes that the ratio between past and modern pCO2 is 1 (RCO2 = 1) and is applied to young material, typically from the Quaternary. The second is the “Carboniferous” standardization that sets the ratio between past and modern pCO2 at two times preindustrial levels of 300 ppm (RCO2 = 2 = 600) (McElwain and Chaloner, 1995). Both standardizations are usually applied to fossil leaf material of Cenozoic age and older to yield minimum and maximum pCO2 estimates and both standardizations will be used in this paper.

We have chosen not to apply the mechanistic optimization model of Konrad et al. (2008) to our study, because it has recently been shown in a modern test of the model to produce the most accurate pCO2 estimates when used on multiple species, to derive a consensus pCO2 estimate from their area of overlapping pCO2 values (Grein et al., 2013), and we here study a one-species database. The optimization model produces very large and species-dependent uncertainty in pCO2 estimates when applied to individual fossil species (Konrad, 2008; Roth-Nebelsick et al., 2012) and even modern species (Grein et al., 2013) for which all the biochemical, environmental and anatomical parameters required to initialize the model are known (Konrad et al., 2008; Grein et al., 2013; Roth-Nebelsick et al., 2012). We have also not applied the mechanistic stomatal model of Franks et al. (2014) because it is shown to be highly sensitive to initial parametrization of assimilation rate resulting in ±500 ppm error in palaeo-pCO2 estimates (McElwain et al., 2016b). Future work on Eotrigonobalanus furcinervis will aim to constrain likely palaeo-assimilation rate for this extinct taxon by applying available palaeo-assimilation proxies (McElwain et al., 2016a, b; Wilson et al., 2015) and undertaking elevated pCO2 experiments on appropriately selected NLEs.

Eotrigonobalanus furcinervis (Rossmässler, 1840; Kvaček and Walther, 1989), an extinct evergreen Fagaceae (Fig. 1), existed from the middle Eocene to the Oligocene–Miocene boundary and was geographically widely distributed, i.e. from central Europe to Russia, as well as to the Mediterranean area (Mai and Walther, 2000; Velitzelos et al., 1999). It is considered as a thermophilous species that grew in evergreen broadleaved forests as well as in mixed mesophytic forests adapted to humid and warm-temperate to subtropical climate (Mai and Walther, 2000). E. furcinervis was present in megafossil assemblages or “taphocoenoses” derived from riparian forests, back swamps, peat bogs and zonal vegetation and therefore the parent plant tolerated a wide range of water table conditions and soil characteristics. Whereas in the Eocene it often predominated in zonal Fagaceae–Lauraceae forests (Mai and Walther, 2000), in the Oligocene mixed mesophytic forest it was ecologically sub-dominant. Based on the combined fossil record of cupules, seeds and leaves, including cuticles, it is commonly accepted that the fossils represent a single long-lived but rather variable fossil species, although minor changes in leaf anatomy have led to the distinction of two subspecies, ssp. furcinervis (mainly Eocene, rare in Oligocene) and ssp. haselbachensis (only Oligocene; Kvaček and Walther, 1989). The latter is distinguished by the absence of pubescence (trichome clusters) on the abaxial leaf epidermis. Furthermore, a variety of leaf morphotypes can be distinguished that have been interpreted as ecological variants (ecotypes, see Kriegel, 2001).

Except for the material from the Kleinsaubernitz site (Fig. 2), the leaf specimens used here originate from the central German Weißelster Basin (Fig. 2), a coastal alluvial plain at the southern margin of the North German–Polish “Tertiary” Basin (Standke, 2008). This basin is well known for its extensive record of middle Eocene to early Miocene plant assemblages that are mainly derived from azonal vegetation, i.e. riparian and swamp forests (e.g. Mai and Walther, 2000; Kunzmann, 2012). The Knau assemblage represents the fluvial hinterland of the Weißelster lignite swamps (Mai and Walther, 2000).

The leaves used here are derived from a succession of cuticle-rich taphocoenoses that contain E. furcinervis ranging in age from the late middle Eocene to the end of the Oligocene (Table 1, Fig. 3). The database analysed here consists of 233 E. furcinervis leaf cuticle fragments on as many slides, representing 151 separate individual leaf specimens (Supplement and Table 2). All specimens represent material used in previous taxonomic-systematic studies; they are housed in the Senckenberg Natural History Collections Dresden, Germany. The plant fossil assemblages have been positioned on the most recent lithostratigraphy for central and East Germany (Standke, 2008; Standke et al., 2010; Fig. 3, Table 2) using published information on the fossil sites (Mai and Walther, 1991, 2000; Kunzmann and Walther, 2002; Hennig and Kunzmann, 2013; Ferdani, 2014) and personal observations (LK). Information on dating is provided in Sect. 2.2.

One late Oligocene locality, Kleinsaubernitz (Figs. 2 and 3a), lies within the Lausitz Basin, at its southern margin or even in the hinterland (Standke, 2008). Leaf specimens derive from a sediment-filled maar, volcanic in origin, preserving a parautochthonous assemblage mainly representing zonal vegetation (Walther, 1999) in contrast to the mainly azonal vegetation from the coastal plains of the Weißelster Basin.

The relative stratigraphic positions for the samples from the Weißelster Basin (Figs. 2 and 3a) are based on accumulating knowledge from more than 150 years of geological-palaeontological investigations of the respective units – see Walther and Kunzmann (2008) for a summary. The samples are derived from a superposed sequence of four lignite seams and their associated strata (Table 1, Fig. 3a), the subdivisions of which can be readily recognized across different opencast mines.

It is not possible to directly correlate the plant-fossil-bearing horizons in the Weißelster Basin to the global marine stratigraphy. Although there are a number of brackish-marine intercalations (Standke et al., 2010) most of these strata lack fossils suitable for biostratigraphy. As is typical for lignite-bearing non-consolidated sedimentary successions (i.e. gravel, sands, silts, clays) hard parts of mineralized organisms that might be used for biostratigraphy in continental sequences (such as mammals and charophytes) are lacking due to dissolution by humic acids originating from organic material. Non-consolidated sediments do not reveal any casts or moulds of these former fossils. This is also the case for any intercalation of brackish-marine sediments in the Weißelster Basin profile. The lack of common index fossils prevents accurate stratigraphic chronology in the basin and reduces the level of stratigraphic resolution compared with that typically attainable for marine deposits (e.g. Roth-Nebelsick et al., 2014). Furthermore, heterogeneity in facies types (channel, floodplain, tidal deposits, swamps) and in grain sizes of the sediments precludes the use of magnetostratigraphic methods which need longer sequences of fine-grained sediments without facies shifts (e.g. lake sediments) to produce reliable data.

Based on a series of consecutive pollen assemblages in the Weißelster Basin strata a regional phytostratigraphic concept was developed (Krutzsch, 1967) that can be applied to all formations, members and submembers, and also to all lignite seams and even individual seam measures (Krutzsch, 2011). All of our investigated material is unambiguously assigned to a certain unit of the regional lithostratigraphic scheme (Fig. 3a) and thus connected to a respective pollen zone or subzone (Fig. 3a, Table 1). However, the pollen zonation yields only a relative age for a given horizon within the regional palynostratigraphic framework and does not enable correlation to global stratigraphy or to the global timescale. The attempt by Krutzsch (2011) to correlate the Eocene spore-pollen zones with the global timescale is used herein (Fig. 3a) as it is the only available information to interpret our assemblages. A “late” Eocene age (i.e. late Bartonian + Priabonian, Krutzsch 2011) for our respective assemblages has been previously inferred based on floristic comparison to assemblages from the nearby Bohemian basins (Czech Republic) some of which have absolute dates from volcanic rocks (i.e. Kučlin, Staré Sedlo, Roudníky; Kvaček et al., 2014).

In the younger part of the succession, marine deposits have yielded index fossils suitable for biostratigraphy. Marine strata above the Gröbers Member of the Böhlen Formation are placed into regional dinoflagellate zones D13 and D14 (Köthe, 2005; Standke et al., 2010) which are Rupelian in age. The Haselbach horizon of the Gröbers Member, including our assemblage sites Schleenhain 4 and Haselbach 2 (Figs. 2 and 3a), was therefore interpreted to be basalmost Oligocene (Standke et al., 2010; Krutzsch, 2011). However, the only definitive information from the dinoflagellate data is that the samples must be older than mid-Rupelian. Lithofacies changes in the centre of the Weißelster Basin, i.e. the profile in the Schleenhain mine (Kunzmann and Walther, 2002), that indicate major sea level changes below the sample horizon of sites Schleenhain 4 and Haselbach 2 are consistent with those that occur around the Eocene–Oligocene boundary and are documented in other European successions (e.g. Hooker et al., 2009). A basalmost Oligocene age for the Schleenhain 4 and Haselbach 2 sites is also indicated by the first occurrence of Boehlensipollis hohlii in the sampled horizon which places the sample in spore-pollen zone 20A/B sensu Krutzsch (2011). Boehlensipollis hohlii is regarded as a key element for the Oligocene in central and East Germany (Krutzsch, 2011) and had also been treated as such in the International Geological Correlation Programme (Vinken, 1988). However, it should be mentioned that Collinson (1992) reported several occurrences of the species in the late Eocene of the UK and Frederiksen (1980) reported the species ranging from late middle Eocene to Oligocene in the USA. Possibly the species arose in the USA and spread later via the UK into central Europe but further work is needed to securely link the occurrences of Boehlensipollis hohlii with the marine biostratigraphy and the global timescale. In short, there are two independent pieces of evidence (lithofacies, first appearance of Boehlensipollis hohlii) that clearly suggest an early Oligocene age for the Schleenhain 4 and Haselbach 2 samples. However, this is not conclusive evidence and direct linkage to the global marine scale is currently not available. The site at Kleinsaubernitz has been located on Fig. 3 based on its pollen assemblage which is zone 20G (Goth et al., 2003).

In summary, the material from the Weißelster Basin comes from a superposed sequence where relative stratigraphic position is securely known (Table 1). Relative changes of SD (and thus pCO2) through the succession can be placed in context of the spore-pollen zonation. However, the positions of the Eocene–Oligocene boundary and the Oligocene-Miocene boundary cannot be located with certainty in the Weißelster profiles. All age estimates in Figs. 3 and 4 are based on Krutzsch's (2011) proposed correlation of the regional spore-pollen zones to global sea level changes. Independent support is needed for these proposals so they should be regarded as preliminary age information.

Cuticles were prepared at the Senckenberg Natural History Collections Dresden as a part of an existing collection. Cuticle slides were prepared using standard methods for Palaeogene material. Fragments removed from leaf specimens with preparation needles were macerated for 1–4 min in Schulze's solution. Cuticles were then neutralized with NH4OH, washed with distilled water, and upper and lower cuticles were separated using preparation needles. Finally, the cuticles were stained with Safranin and affixed to slides by glycerol jelly. For this study, the slides were examined microscopically by an adaptation of the methodology set out by Poole and Kürschner (1999) in order to determine SD. According to this protocol, counts from mid-lamina are preferable in establishing SD, but the fragmented nature of a proportion of the fossil material did not allow establishing where individual cuticle samples were located on the original leaf surface (see Fig. 1b). Individual epidermal cells were not easily discernible in the majority of the E. furcinervis material, making SI determination impossible. SD was obtained using a Nikon SK Light Microscope at ×200 magnification with a graticule providing a counting field of 0.042 mm2. The graticule was centred over areas where stomata occurred in greatest numbers (away from veins and margins where those were known, sensu Poole and Kürschner, 1999) and up to five individual counts were recorded for each slide, resulting in 659 SD counts for the database of 151 leaf specimens (Table 1 and Supplement). Data were stored in Microsoft Excel 2010 before being statistically manipulated using MINITAB (version 16.1.1 for Windows).

Eotrigonobalanus furcinervis belongs to the Fagaceae, but its phylogenetic position is not well defined. Based on cupule morphology, Eotrigonobalanus belongs to a basal clade of the family, exhibiting intermediate characters between modern Trigonobalanus and Castanopsis (Mai, 1995). However, leaf venation and leaf cuticle micromorphology place Eotrigonobalanus with Trigonobalanus and Lithocarpus, away from Castanopsis (Kvaček and Walther, 1989), an affiliation recently confirmed by Denk et al. (2012). Since the phylogeny of Fagaceae has changed considerably (Manos et al., 2001, 2008), an improved systematic framework is still required to confirm the phylogenetic position of Eotrigonobalanus. Because the exact relationship of Eotrigonobalanus to crown group Fagaceae is unknown, a single nearest living relative (NLR) could not be obtained, hence the nearest living equivalent (NLE) approach has been used for the stomatal proxy-based pCO2 reconstruction.

In this study, Trigonobalanus doichangensis was chosen as the NLE, due to it being a basal species within the Fagaceae family and having leaf macro-morphological and leaf cuticle micro-morphological similarities with E. furcinervis, including cyclocytic stomata and similarly structured trichomes (Kvaček and Walther, 1989; see also Denk et al., 2012). Two herbarium specimens of T. diochangensis, formerly collected in 1988, were sampled at the Kew Herbarium (Royal Botanical Gardens, Kew, Richmond, Surrey, UK). Approximately 1 cm2 was cut from mid-lamina of each leaf specimen and dry mounted onto a slide. Five cuticle images from each slide were taken at 200× magnification using a Leica DM2500 epifluorescent microscope with Leica DFC300FX camera (Leica^® 312 Microsystems, Wetzlar, Germany) and Syncroscopy Automontage (Syncroscopy Ltd, Cambridge, UK) digital imaging software. A 0.09 mm2 square was superimposed on each image and stomatal density was determined within this square following the protocol of Poole and Kürschner (1999). SD was determined to be 546.11 mm2 at pCO2 of 351 ppm (collection year levels according to NOAA ESRL data, available at http://www.esrl.noaa.gov).

Using the stomatal ratio method with T. doichangensis NLE for E. furcinervis, we calibrated palaeo-pCO2 using the equations below to derive minimum and maximum palaeo-pCO2 (“Modern” and “Carboniferous” Standardization of McElwain and Chaloner, 1995), respectively: Palaeo-pCO2min(ppm)=SDNLE=546.11/SDfossil⋅351ppmPalaeo-pCO2max(ppm)=SDNLE=546.11/SDfossil⋅600ppm.

SD of E. furcinervis range between ca. 425 and 740 stomata mm-2. The lowest SD values (signifying highest pCO2) are found in the oldest deposits, late middle to earliest late Eocene (spore-pollen zone 17), and the highest values (signifying lowest pCO2) are found in the later late Eocene (spore-pollen zone 18o), representing the most pronounced SD change during the time period covered by the data set (Table 2, Fig. 3b), with three intermediate samples showing intermediate values (spore-pollen zones 17/18, 18u, 18uo). During this interval SD increases by > 300 stomata mm-2 or by ca. 75 %, a very significant change indicating a sizeable decrease in pCO2 in perhaps ca. 3.5 million years. Stomatal densities then decrease slightly again and remain around 600–650 mm-2 in the latest Eocene and in samples that may be earliest Oligocene as well as in the late Oligocene (spore-pollen zones 19, 20A/B, 20G, II). At the end of the Oligocene, SD decreases again to ca. 570 mm-2.

Palaeo-pCO2 calibrated using the stomatal densities of E. furcinervis will be discussed as average values and evaluated in terms of relative change, as introduced above. The largest change in palaeo-pCO2 is the decrease from the late middle to earliest late Eocene of > 250 ppm, from ca. 630 to ca. 365 ppm – a decrease in pCO2 of ca. 40 % (Fig. 3b; Table 2). Concentrations of CO2 then increase again by ca. 45 to ca. 410 ppm in the latest Eocene and possibly earliest Oligocene, and further to between ca. 430–475 ppm in the late and latest Oligocene (Fig. 3b; Table 2).

The Saxony fossil leaf database is unique in that this relatively large database derives from a well-constrained stratigraphic succession and consists of a single species throughout – E. furcinervis – which is the most ideal situation when using fossil leaf material to reconstruct palaeo-pCO2, since inter-species variability is eliminated and stomatal responses to pCO2 are likely to be consistent through time. The procurement of a single-species data set from multiple stratigraphic levels across several million years is not common, in particular when the stratigraphy represents time intervals of significant climate and/or environmental change, as is the case here. The principal challenge concerning the Saxony stomatal density record was translating the stomatal signal into reliable levels of pCO2. One of the main limitations associated with the use of palaeo-proxies is the preservational state of fossil material and in this case the preservation of fossil leaves did not allow palaeo-pCO2 reconstruction using gas exchange models for independent comparison of the results using the stomatal ratio method because stomatal pore length could not be measured in all samples with confidence. Additionally, there is a lack of available transfer functions for potential NLEs of E. furcinervis, so it was not possible either to obtain independent pCO2 reconstructions using the transfer function method. The stomatal ratio method has however been shown to closely match results produced with transfer function methods (Beerling and Royer, 2002; Royer, 2003; Barclay et al., 2010; Steinthorsdottir et al., 2011b) and is seen as a good alternative where detailed estimates of other photosynthetic parameters, which are required to initialize mechanistic models, are not readily available (McElwain, unpublished data).

The absence of an obvious NLE for E. furcinervis – an extinct species of uncertain phylogenetic affinity – further introduces potential errors in pCO2 calibration. Although we consider T. doichangensis the best available NLE, there is no guarantee that its stomatal density and degree of response to pCO2 closely mirrors that of its distant fossil relative. The pCO2 levels calibrated here appear somewhat low compared to most previously published pCO2 data sets, although broadly comparable to stomatal pCO2 records (Fig. 4a). When testing three additional potentially suitable NLE species for reconstructing pCO2 using the Saxony database (Trigonobalanus verticillata, Castanopsis cuspidata and Lithocarpus henryi), the resulting palaeo-pCO2 values were extremely low – considerably lower than when using the chosen NLE T. doichangensis – in many cases being lower than minimum pCO2 levels required to maintain sufficient plant growth and reproduction (i.e. below the ecological compensation point). This indicates that, for some reason (e.g. species-specific responses) the stomatal proxy-derived pCO2 estimates presented here based on E. furcinervis may be artificially low.

Palaeoclimate reconstructions based on central European megafloras reveal a sharp decline in continental cold month mean temperature (Mosbrugger et al., 2005) and mean annual temperature (Moraweck et al., 2015; Kvaček et al., 2014) in the late Eocene (Fig. 4b) which is consistent with the timing of the pCO2 decline that we report here (Fig. 4a and b), and with global sea surface temperature trends as recorded by marine oxygen-isotopes (Fig. 4c). The marine isotope curve also shows a gradual decrease of temperatures in the late Eocene, but in contrast with the terrestrial records, the most pronounced and abrupt change coincides with the Eocene–Oligocene boundary (Fig. 4c), suggesting that pCO2 drawdown may have taken place gradually before the slow feedback ice sheet growth was initiated and global temperatures dropped suddenly in response. The possibility remains that future terrestrial proxy reconstructions of pCO2 will record a transient major drawdown of pCO2 at the Eocene–Oligocene boundary. In order to resolve this, more proxy records from well-constrained early Oligocene sites must be added.

Furthermore, palaeo-vegetation analysis of the Weißelster and North Bohemian basins reveals that gradual restructuring of dominantly evergreen forests by immigration of deciduous species such as Platanus neptuni, Trigonobalanopsis rhamnoides and Taxodium dubium (Kunzmann et al. 2015) took place in the late Bartonian to early Priabonian interval around ca. 38 Ma (Kvaček, 2010; Teodoridis and Kvaček, 2015). The temporal coincidence of pCO2 decline and major vegetation transition – from angiosperm-dominated notophyllous evergreen forests to mixed mesophytic forests – suggests a potential causal role of pCO2 decline in the changing ecological composition of forests. It may have been in part triggered by differential responses of evergreen and deciduous taxa to declining pCO2 (Fig. 4a and b), explaining the lag between “temperatures” indicated by terrestrial vegetation and sea surface temperatures recorded by marine oxygen-isotopes (Fig. 4c). The functional trait of deciduousness is an adaptation to episodic cooling (Zanne et al., 2014). However, it has also been demonstrated experimentally (McElwain et al., 2016b) and on theoretical grounds (Niinemets et al., 2011) that taxa with low leaf mass per area or LMA (i.e. those that are deciduous or herbaceous) and high stomatal conductance have faster photosynthetic rates than evergreens at lower atmospheric pCO2. In contrast, evergreens have higher responsiveness in terms of photosynthetic rates at elevated pCO2 (Niinemets et al., 2011). A transition from elevated to lower CO2 atmospheres would therefore favour the ecophysiology of deciduous or low-LMA taxa over evergreen high-LMA species. Further experimental investigation is now required to tease apart the relative importance of “CO2 starvation” and increased temperature seasonality on the late Bartonian to early Priabonian vegetation transition.

Previously published stomatal proxy-based pCO2 records from the part of the Cenozoic relevant to this paper do not always agree, but instead report highly elevated (McElwain, 1998; Doria et al., 2011; Grein et al., 2011; Smith et al., 2010), intermediate (Retallack, 2009) or similar to modern (Royer et al., 2001) pCO2 for the Eocene. Similarly high variability in estimated pCO2 levels exists for the Oligocene as well as the Miocene (Grein et al., 2013; Kürschner et al., 2008; Roth-Nebelsick et al., 2014; Royer et al., 2001). The results reported here are the highest stratigraphic resolution pCO2 estimates derived from the late Eocene to early Miocene basins in Saxony (see Table 2, Figs. 1 and 3). Previous studies have tended to only report temporal trends on stomatal parameters (Roth-Nebelsick et al., 2004) or to lump pCO2 estimates from single Saxony localities into coarse temporal bins making cross comparison difficult (Roth-Nebelsick et al., 2012). However, where individual site pCO2 data are reported (Grein et al., 2013) our estimates are in very good agreement with previous studies despite differences in species and calibration approach (Table 2). For example, Grein et al. (2013) report pCO2 estimates of ∼ 400 ppm and between ∼ 430 and ∼ 530 ppm respectively for the sites Kleinsaubernitz and Witznitz (Fig. 3) using the Konrad et al. (2008) stomatal optimization model in a consensus approach on multiple species (3–4) including E. furcinervis (Table 2). The optimization model produces a very large range of pCO2 estimates however (∼ 270 to 710 ppm), when applied to E. furcinervis alone from stratigraphically lumped samples from Haselbach and Profen (Table 2) (Roth-Nebelsick et al., 2012). In comparison with the study of Roth-Nebelsick et al. (2012), we report seven stratigraphically well-resolved pCO2 estimates spanning the same interval for which they report a single lumped average (∼ 470 ppm) for two sites (Table 2). This is thus the first study to resolve a significant drop in palaeo-pCO2 in the late Eocene, prior to the E–O boundary from a stratigraphically well-constrained and relatively high-resolution record.

Using a rigorous generalized statistical framework, Beerling et al. (2009) revised previously published pCO2 estimates based on Ginkgo and Metasequoia from the early Eocene and middle Miocene upwards by 150–250 ppm. Based on this revision, average stomatal proxy-based pCO2 is 450–700 ppm in the Palaeogene and 500–600 ppm in the Neogene (Beerling et al., 2009). Interestingly, the younger set of pCO2 estimates was fully compatible to marine proxy data and modelling results (e.g. Pagani et al., 2005; Hansen et al., 2008), whereas the older set of estimates seemed to underestimate pCO2 compared to the other approaches, even after the upwards revision of stomatal pCO2 values (see Fig. 4 in Beerling et al., 2009). However, Kürschner et al. (2008) indicated that an upwards correction of 150–200 ppm – a so-called “correction factor” – was necessary also when reconstructing Miocene palaeo-pCO2 with two species from the Lauraceae family. Recently discrepancies between the various pCO2 proxies have narrowed significantly, and a coherent pattern of long-term Cenozoic pCO2 has emerged, indicating pCO2 mostly in the hundreds rather than thousands of ppm, although shorter-term inter-proxy discrepancies remain (see Beerling and Royer, 2011, Fig. 1). It has thus become evident that pCO2 values reconstructed using the stomatal proxy do not require a correction factor.

Pearson et al. (2009) reconstructed pCO2 for the late Eocene to early Oligocene using the planktonic foraminifera boron isotope pH proxy and found that the main reduction in pCO2 took place before the main phase of EOT ice growth (ca. 33.6 Ma: DeConto et al., 2008), followed by a sharp recovery to pre-transition levels and then a more gradual decline. Their results thus support the central role of declining pCO2 in Antarctic ice sheet initiation and development and agree broadly with carbon cycle modelling (e.g. Merico et al., 2008). The quantitative estimates of pCO2 varied greatly however, according to which δ11B value was used to derive pH, with geochemical models of the boron cycle suggesting a range of 37–39 ‰ for sea water (sw) δ11B during this time (Simon et al., 2006). The range of pCO2 values spanned from ca. 2000–1500 ppm at the upper end and ca. 620–450 ppm at the lower end (Pearson et al., 2009). Recently published alkenone-based pCO2 records found significantly declining pCO2 before, as well as during, the Antarctic glaciation (EOT and earliest Oligocene), supporting the pCO2 pattern of Pearson et al. (2009) and the role of pCO2 as the primary forcing agent of Antarctic glaciation, consistent with model-derived thresholds (DeConto et al., 2008; Pagani et al., 2011; Zhang et al., 2013). The alkenone-derived data set values are overall higher – but not much higher – than those derived by stomatal densities, with late Eocene values of ca. 1000 ppm, minimum value of ca. 670 at 33.57 Ma and then gradual decline to ca. 350 ppm at the Oligocene-Miocene boundary.

In general therefore, Cenozoic stomatal proxy-based pCO2 values, reconstructed using the available methods, tend to report somewhat lower pCO2 values than alkenone- or boron-based proxies as well as those from mass balance modelling. As discussed above, isotope-based proxy records depend on a range of assumptions that influence the output interpretation to a large extent. In addition, it has recently been shown that the modelled pCO2 threshold for Antarctic glaciation at the EOT, routinely referred to be ca. 700 ppm (DeConto and Pollard, 2003), is in fact highly dependent on the type of climate model used and the configurations of the model (Gasson et al., 2014), implying that the range of Cenozoic pCO2 may be due for an update. It is noteworthy that most existing stomatal proxy-based pCO2 records report a similar range of low pCO2 values for this time interval and an internally consistent pattern is emerging for the Cenozoic (see Fig. 4a). Stomatal proxy-based pCO2 records that are independently calibrated using different species/genera and families usually agree with one another and show Eocene–Miocene pCO2 in the range of 800–300 ppm (Fig. 4a). Although this discrepancy between proxies needs to be better understood before significant reevaluation of the role of pCO2 in Cenozoic climate change is warranted, it should not be a priori rejected that collectively stomatal proxy records may accurately indicate lower pCO2 levels during the Cenozoic than previously assumed.

The concept of Earth's climate sensitivity – usually defined as the equilibrium surface temperature response to doubling of pCO2 (2 × CO2) – is a key parameter for understanding the mechanisms of future climate change. Recently there has been much focus on accurately and uniformly defining this concept, but although progress has been made, discrepancies still remain. The term most in use for predicting future climate change is “equilibrium climate sensitivity”, defined as the response of global mean surface temperatures to a 2 × CO2 radiative forcing after all the fast feedbacks have occurred (changes in atmospheric temperatures, clouds, water vapour, winds, snow, sea ice, etc.), but before the slow feedbacks occur (mainly ice sheet, vegetation and the carbon cycle responses) and often estimated to be ca. 3 ∘C (Rohling et al., 2012; Royer et al., 2012; Hansen et al., 2013; Huber et al., 2014; the Intergovernmental Panel on Climate Change report 2013 best estimate 1.5–4.5 ∘C). When studying palaeo-climate sensitivity, which has the potential to be accurately inferred from high-resolution palaeo-climate proxy archives, both fast and slow feedbacks must be considered to define a related concept – the “Earth system sensitivity”, where e.g. pCO2 may act both as forcer and as feedback, and which depends to a large degree on the initial climate state (Royer et al., 2007; Hansen et al., 2013). In the Cenozoic, pCO2 is involved in climate change both as forcing and feedback, with evidence of increased climate sensitivity in warm climates, rather than cool ones (Hansen et al., 2013).

The Eocene–Oligocene global cooling transition is represented by a large increase in deep-sea benthic foraminiferal oxygen isotope values, reflecting simultaneously decrease in temperatures and increased ice sheet growth, with as of yet no proxy to accurately separate the relative effects of the two (Zachos et al., 2001, 2008). Constraining the decrease in temperature that occurred during the transition is thus a work in progress, but consensus is emerging around a ca. 2–5 ∘C cooling in sea surface as well as mean annual air temperature (e.g. Lear et al., 2009; Zachos et al., 2008; Liu et al., 2009; Bohaty et al., 2012; Wade et al., 2012; Hren et al., 2013; Inglis et al., 2015; Petersen and Schrag, 2015). The EOT cooling and glaciation was forced by a decrease in pCO2 from ca. 1000 to ca. 600 ppm based on marine isotopes and climate modelling (e.g. DeConto et al., 2008; Pearson et al., 2009; Pagani et al., 2011) or ca. 800 to ca. 400 ppm based on stomatal records (e.g. Beerling and Royer, 2011; this data set) – a decrease of at least ca. 40 % in pCO2 in < 5 Ma. A simple estimation of Earth system sensitivity during the EOT suggests elevated sensitivity compared to today, implying an enhancing factor by fast and/or slow feedbacks, such as ice sheet growth, but the radiative contribution of each is presently unknown (Lunt et al., 2010; Goldner et al., 2014; Gasson et al., 2014; Maxbauer et al., 2014). The transition in Earth's climate mode from the Eocene greenhouse to the Oligocene icehouse was driven by changes in pCO2 (and associated feedbacks) that largely fall within the range of modern to predicted future pCO2 – albeit in the opposite direction. Understanding how the Earth system responds to radiative forcing within this range (i.e. understanding Earth system sensitivity) is of considerable interest, with the input and correlation of multiple palaeo-pCO2 proxy records being of crucial importance.