5.1. Comparison With Other CO2 Records
 Because CO2 is well mixed in the atmosphere and because surface ocean carbonate chemistry at site 999 is close to equilibrium with the atmosphere, we consider our aqueous pCO2 estimates approximate to atmospheric pCO2, hereafter referred to as δ11B-based pCO2. At 2.0 Ma our estimates overlap with pCO2 estimates from the eastern equatorial Atlantic obtained via the same method [Hönisch et al., 2009] (Figure 4), which raises confidence that boron isotope reconstructions from oligotrophic oceanic regions indeed yield reliable estimates of atmospheric pCO2.
 Few stomata index-based CO2 estimates of 280 and 370 μatm between 3.3 and 2.0 Ma [Kürschner et al., 1996] are also in general agreement with our data (Figure 4c). Pagani et al.  reconstructed surface ocean pCO2 based on alkenone ɛp from 6 locations in the oligotrophic and mesotrophic Atlantic and Pacific Oceans (not shown). For all these locations, their estimates show consistently higher pCO2 for the Pliocene, although the absolute values differ between sites and only two of their records approach ice core CO2 values of 170–280 ppmv during the Pleistocene. Pagani et al.  estimated average pCO2 values of 365–415 μatm during the warm early Pliocene, which is similar to our CO2 estimates (380–420 μatm), and described a decrease of atmospheric pCO2 by 45–144 μatm or 41–216 μatm, depending on their assumption of nutrient supply, over the last 4.5 Ma. Our boron isotope reconstruction is restricted to the period 4.2–2.0 Ma and registers a relatively similar decrease by 100 μatm.
 Pearson and Palmer's  Cenozoic record is also based on planktic foraminiferal δ11B, using a different analytical technique than in this study. Although their Pliocene estimates show a similar decline in atmospheric CO2 from 280 to 210 ppm between 3.87 Ma to 3.0 Ma, their low temporal resolution and use of mixed planktic foraminifer species preclude any meaningful comparison. A more detailed comparison can be made with recently published pCO2 estimates from site 999 based on alkenone ɛp and δ11B [Seki et al., 2010]. In general both data sets are in good agreement and show similar CO2 trends and absolute values through time (Figure 4c). However, the sampling strategy applied by Seki et al.  appears to have favored interglacial times, as they did not yield any pCO2 estimates lower than 250 μatm during the Pleistocene. Similarly, the pCO2 estimates from alkenone ɛp only reproduce intermediate values (∼240 μatm) if low nutrient concentrations are assumed. Their favored scenario of elevated nutrient concentrations agrees well with the boron isotope ratios measured on the same samples but suggests pCO2 > 250 μatm for the entire record (Figure 4c).
 Finally, reconstructions of Pliocene pCO2 include estimates from foraminiferal B/Ca ratios (not shown) spanning 3.4–2.4 Ma at western tropical Pacific site 806 [Tripati et al., 2009]. Their CO2 data present a maximum for the Pliocene epoch of only 300 μatm (±50 μatm) around 3.2 Ma and a decrease to 150 μatm (±50 μatm) around 2.8 Ma. This range seems biased toward too low values compared to other Pliocene records, which record atmospheric pCO2 higher than the preindustrial level for this warm epoch [Seki et al., 2010; Pagani et al., 2010; Kürschner et al., 1996; Raymo et al., 1996], including this present study. Because various calibration data sets for a pH effect on foraminiferal B/Ca ratios show conflicting results [Yu et al., 2007; Foster, 2008; Tripati et al., 2009], much more information is needed on this new proxy before past pCO2 reconstructions can be approached with confidence [Tripati et al., 2011].
 With the exception of the B/Ca estimates, all records agree that interglacial Pliocene pCO2 was higher by about +110 μatm compared to the Pleistocene, although small differences exist in the details of each record. In agreement with studies referenced above, our record shows that a gradual decline in CO2 by ∼100 μatm from 4.1 to 2.0 Ma coincides with the late Pliocene onset and intensification of Northern Hemisphere Glaciations, supporting the link between glaciation and decrease in atmospheric CO2 concentration, as previously postulated [Raymo et al., 1996; Seki et al., 2010; Pagani et al., 2010] and modeled [Lunt et al., 2008]. However, the higher temporal resolution of this record offers a more detailed perspective on the timing and the supposed causes of the atmospheric CO2 changes between 2.0 and 3.6 Ma, as described below.
5.2. Implications for the Onset and Intensification of Northern Hemisphere Glaciations
 Prior to the start of Northern Hemisphere glaciations, early Pliocene δ11B-based pCO2 values between 4.6 and 3.6 Ma averaged maxima of 410 μtam and minima of 310 μatm, both above a suggested threshold value of 280 μatm below which Northern Hemisphere glaciations are possible [DeConto et al., 2008]. This may explain in part why large-scale glaciations on Greenland were not recorded during this time. Alternatively, Koenig et al.  defined other CO2 threshold values between 200 and 400 μatm for the initiations of glaciations on Greenland by taking into account vegetation changes (i.e., forest versus tundra) over an initially ice-free Greenland. They concluded on a critical role of decreasing atmospheric CO2 in the glaciation of Greenland at 2.74 Ma, although other albedo-related feedbacks were also significant, including the vegetation cover on Greenland and the sea ice cover in Greenland and Labrador Seas [Koenig et al., 2011]. If the CO2 threshold defined by Koenig et al.  is correct, Miocene and early Pliocene atmospheric CO2 estimates <400 μatm [Seki et al., 2010; Pagani et al., 2010; Kürschner et al., 1996; this study] would explain why ice-rafted detritus deposition in the North Atlantic could start as early as 5–10 Ma [Wolf and Thiede, 1991; Jansen and Sjøholm, 1991].
 After 4.2 Ma and until 2.0 Ma, minima/maxima amplitudes of estimated pCO2 were as high as 100 μatm (Figure 4c), similar to the Pleistocene glacial/interglacial CO2 amplitude [Hönisch et al., 2009]. While the Pleistocene amplitude of 100 μatm is still not fully understood [Kohfeld et al., 2005], it is clear that glacial/interglacial CO2 variations are mainly controlled by the Southern Ocean through ventilation of the deep ocean and biological pump strength [see Sigman et al., 2010]. Although the estimated Pliocene minima/maxima CO2 amplitudes of 100 μatm require validation by other atmospheric CO2 records, such amplitudes could be controlled by the state of the Southern Ocean. Indeed, glacial/interglacial oscillations in the size of the West Antarctic Ice Sheet during the early Pliocene [Naish et al., 2009] may have influenced the overturning of the Southern Ocean and the sequestration of atmospheric CO2 during the early Pliocene. Similarly, it was suggested by Sigman et al.  that overturning of the polar oceans could be responsible for large amounts of CO2 degassing during the late Oligocene and middle Miocene.
 After the start of Northern Hemisphere glaciations around 3.6 Ma, maximal pCO2 estimates transiently decreased from 410 μatm to 260 μatm on average between 3.4 and 3.32 Ma, while minimal pCO2 estimates gradually decreased from 310 μatm to 245 μatm. A transient increase in glacial and interglacial ice volume is also recorded between 3.43 and 3.32 Ma by an increase of 0.20‰ in the benthic δ18O stack LR04 (Figure 4b) and/or by a transient increase of 12 m equivalent sea level (0.11‰ δ18O) around 3.3 Ma (Figure 3c) [Mudelsee and Raymo, 2005]. The transient glaciation on Greenland at this time also coincides with a transient decrease in North Atlantic SST between 3.3 and 3.5 Ma by 2°C during interglacials and 5°C during glacials at site U1313 [Naafs et al., 2010] and at site 982 [Lawrence et al., 2009]. Increased ice rafted-detritus deposition at 3.3 Ma in the North Atlantic [Jansen et al., 2000; Kleiven et al., 2002] and in the Labrador Sea [Sarnthein et al., 2009] also evidenced the transient increase in the size of the Greenland and Laurentide ice sheets. Although the transient decline in pCO2 coincides with transient glaciations on Greenland between 3.43 and 3.32 Ma, the maximal pCO2 estimates between 3.6 and 3.2 Ma are as low as Pleistocene interglacial values, an observation that may be unexpected from comparison with the benthic δ18O stack LR04 (Figure 4b). However, several lines of evidence suggest that the low pCO2 estimates could indicate temporary disequilibrium between surface seawater at site 999 and the atmosphere during this time. Increased nutrient supply to site 999 between 3.5 and 3.1 Ma [Kameo, 2002] and increased surface seawater primary productivity between 3.4 and 3.35 Ma at nearby site 502A [Bornmalm et al., 1999] have been documented for this time, suggesting a possible lowering of surface seawater pCO2 compared to the atmosphere via increased primary productivity [Wanninkhof et al., 2007]. Similarly, low surface water salinity at site 999 between 3.5 and 3.1 Ma could suggest an inflow of low-salinity and nutrient-rich Amazon and Orinoco river water, brought to the Caribbean Sea by the Guyana Current [Kameo et al., 2004]. In any case, the decrease in pCO2 estimates appears to have occurred during the start of Northern Hemisphere glaciations at 3.3–3.6 Ma.
 Between 3.2 and 2.7 Ma, maximal pCO2 estimates returned to values of 400 μatm on average until 2.77 Ma (Figure 4c), similar to present-day CO2 level. Interglacial ice volume also resumed to previous level during this time (Figure 4b). Although it was not the specific focus of this paper, we estimate high pCO2 of 410 μatm (MIS K1) and 350 μatm (MIS G19) during interglacials of the mid-Pliocene warm period (3.29–2.97 Ma). This confirms the earlier notion that the warmer climate characteristic of the mid-Pliocene warm period may have been partly driven by a stronger greenhouse effect due to higher atmospheric CO2 [Raymo et al., 1996]. Lunt et al.  discussed in detail the implication of such a result for the estimation of Earth sensitivity through time. Between 3.2 and 2.7 Ma, minimal CO2 estimates stayed at a level of 245 μatm (Figure 4c), while maximal glacial ice volume also stayed constant between 3.1 and 2.8 Ma, as evidenced by the LR04 benthic δ18O stack (Figure 4b). Ice-rafted detritus deposition was also low in the North Atlantic after the transient increase around 3.3 Ma until 2.7 Ma [Kleiven et al., 2002]. Moreover, North Atlantic SST records off South Iceland increased by 2–3°C during interglacial and glacial stages at site 984 [Bartoli et al., 2005] and site 982 [Lawrence et al., 2009] between 3.05 and 2.80 Ma. This suggests that warming over the northern North Atlantic, associated with high atmospheric CO2 concentrations and low obliquity [Laskar et al., 2004] jointly prevented the further growth of the Greenland ice sheet between 3.2 and 2.7 Ma.
 During the late Pliocene transition at 2.73 Ma, global ice volume increased by 22 m equivalent sea level [Sosdian and Rosenthal, 2009], glacial/interglacial cycles intensified [Lisiecki and Raymo, 2007], deepwater cooled by 2°C [Sosdian and Rosenthal, 2009], and ice-rafted detritus deposition increased over the North Atlantic, thus confirming the increase of the Greenland, Laurentide, and Scandinavian ice sheets [Kleiven et al., 2002]. After 2.7 Ma, our minimal pCO2 estimates decreased by 45 μatm (Figure 4c) and reached values similar to early Pleistocene glacial values of 200 μatm, while maximal pCO2 estimates decreased by 50 μatm but stayed +50 μatm higher than during early Pleistocene interglacials. Our results are consistent with the timing of increased stratification in the subarctic North Pacific [Sigman et al., 2004] and in the Southern Ocean [Hodell and Venz-Curtis, 2006; Waddell et al., 2009] at 2.73 Ma. It has been suggested that increased polar stratification led to enhanced sequestration of atmospheric CO2 in the oceanic abyss after 2.73 Ma, by preventing CO2-rich deep waters to reach the surface. Stratification thus could have played a major role in the onset of large-scale ice sheets in the Northern Hemisphere [Sigman et al., 2004]. North Pacific stratification alone could account for a decrease of 30–40 μatm [Haug et al., 1999], which is close to the 45 μatm glacial decrease registered in this study between 2.70 and 2.68 Ma. In addition, the strong North Pacific halocline may have persisted during both glacials and interglacial in the North Pacific [Swann, 2010] and thus may have played a role in the 50 μatm decrease during interglacials after 2.7 Ma. Recently, Martínez-Garcia et al.  presented a record of Aeolian iron input to the Southern Ocean that suggests increased iron fertilization during glacials after 2.7 Ma. In addition to polar stratification, increased export production and nutrient utilization in the Southern Ocean thus may have contributed to the glacial decrease in atmospheric CO2 [Martínez-Garcia et al., 2011]. Increased dust supply to the North Pacific after 2.75 Ma [Bailey et al., 2011] may have had a similar fertilization effect on the North Pacific and adjacent seas, leading to higher CO2 drawdown in these regions as well. Our results suggest that a threshold was crossed after 2.7 Ma, when the onset of polar stratification and glacial iron fertilization in the North Pacific and Southern Ocean were responsible for the glacial atmospheric CO2 decrease to values characteristic of Pleistocene glacials. Interestingly, tropical SST changes in various ocean basins show a strong increase in the 41 ky orbital periodicity at 2.7 Ma [Herbert et al., 2010]. This suggests that atmospheric CO2 and glacial feedbacks after 2.7 Ma were strong enough to imprint a 41 ky periodicity on tropical SST records as compared to prior to 2.7 Ma [Herbert et al., 2010]. Our study agrees well with a decrease in glacial atmospheric CO2 to Pleistocene levels after 2.7 Ma.
 In comparison, maximal pCO2 estimates decreased by only 50 μatm after 2.7 Ma in our record (Figure 4c) but another high level of 380 μatm is observed at 2.2 Ma, suggesting that interstadial pCO2 stayed above early Pleistocene values until 2.0 Ma. A single estimate from stomata also records 360 μatm around 2.0 Ma (Figure 4c). Our findings may suggest that after the sequestration of atmospheric CO2 in the oceanic abyss after 2.7 Ma, another mechanism was responsible for the lowering interglacial atmospheric CO2 after 2.2 Ma, such as the intensification of the Benguela upwelling system between 2.1 and 1.9 Ma, which would have increased the efficiency of the biological pump [Marlow et al., 2000]. Alternatively, Etourneau et al.  suggested that increased upwelling off Namibia between 2.4 and 2.0 Ma would have increased atmospheric CO2 by releasing deep water CO2 to the surface and thereby counteracting the effects of atmospheric CO2 sequestration elsewhere in the oceans until 2.0 Ma. This mechanism is also supported by our data showing increased atmospheric CO2 at 2.2 Ma compared to 2.5 Ma.
 In summary, the increase in the size of the Greenland ice sheet during the Pliocene epoch was short lived between 3.43 and 3.32 Ma [Lisiecki and Raymo, 2005] and permanent after 2.7 Ma [Haug et al., 2005], and is matched by decreased atmospheric CO2 concentrations as compared to the early Pliocene. Following the pattern given by the LR04 benthic δ18O stack (Figure 4b) and tropical SSTs [Herbert et al., 2010, Figure 6], minimal atmospheric CO2 concentrations decreased faster than maximal atmospheric CO2 concentrations over the course of the Plio-Pleistocene. It seems that an important climatic threshold was crossed at 2.7 Ma, when estimated pCO2 minima decreased by ∼45 μatm and approached early Pleistocene glacial values. This decrease and its timing are consistent with the onset of polar stratification [Haug et al., 1999; Sigman et al., 2004] associated with iron fertilization in the Southern Ocean [Martínez-Garcia et al., 2011] and in the North Pacific [Bailey et al., 2011]. It is possible that Southern Ocean stratification starting around 3.3 Ma [Hillenbrand and Cortese, 2006] has contributed to the early atmospheric CO2 decrease between 3.43 and 3.32 Ma, although the role of terrestrial weathering, in particular of the Tibetan-Himalayan Plateau between 4.2 and 2.7 Ma [Zhang et al., 2009], and other possible mechanisms also have to be considered. Based on the timing of the precursor and final closures of Panama dated at 3.15–3.3 Ma and at 2.82–2.95 Ma [Bartoli et al., 2005], the final closure of Panama did not influence the decrease in atmospheric CO2 recorded here. It can be postulated that a decrease in atmospheric CO2 concentrations was necessary in order to sustain large-scale glaciations in the Northern Hemisphere after 2.7 Ma, as opposed to the ephemeral glaciations of the Pliocene prior to 3.2 Ma.