Descent toward the Icehouse: Eocene sea surface cooling inferred from GDGT distributions

The TEX86 proxy, based on the distribution of marine isoprenoidal glycerol dialkyl glycerol tetraether lipids (GDGTs), is increasingly used to reconstruct sea surface temperature (SST) during the Eocene epoch (56.0–33.9 Ma). Here we compile published TEX86 records, critically reevaluate them in light of new understandings in TEX86 palaeothermometry, and supplement them with new data in order to evaluate long-term temperature trends in the Eocene. We investigate the effect of archaea other than marine Thaumarchaeota upon TEX86 values using the branched-to-isoprenoid tetraether index (BIT), the abundance of GDGT-0 relative to crenarchaeol (%GDGT-0), and the Methane Index (MI). We also introduce a new ratio, % GDGTRS, which may help identify Red Sea-type GDGT distributions in the geological record. Using the offset between TEX86 H and TEX86 L (ΔH-L) and the ratio between GDGT-2 and GDGT-3 ([2]/[3]), we evaluate different TEX86 calibrations and present the first integrated SST compilation for the Eocene (55 to 34 Ma). Although the available data are still sparse some geographic trends can now be resolved. In the high latitudes (>55°), there was substantial cooling during the Eocene (~6°C). Our compiled record also indicates tropical cooling of ~2.5°C during the same interval. Using an ensemble of climate model simulations that span the Eocene, our results indicate that only a small percentage (~10%) of the reconstructed temperature change can be ascribed to ocean gateway reorganization or paleogeographic change. Collectively, this indicates that atmospheric carbon dioxide (pCO2) was the likely driver of surface water cooling during the descent toward the icehouse.


Introduction
Throughout the Phanerozoic, and possibly throughout geological time, the Earth's climate has oscillated between greenhouse and icehouse climate states, where greenhouse climates are characterized by high atmospheric carbon dioxide (pCO 2 ) [Pearson and Palmer, 2000;Pagani et al., 2005;Lowenstein and Demicco, 2006;Pearson et al., 2009], high sea surface temperatures (SST) [Pearson et al., 2007;Bijl et al., 2009;Hollis et al., 2012], and the absence of continental ice sheets [Francis and Poole, 2002;Contreras et al., 2013], and icehouse climates are characterized by the opposite: reduced pCO 2 , reduced SST, and presence of continental ice sheets [Zachos et al., 1992;Pearson and Palmer, 2000;DeConto and Pollard, 2003;Pagani et al., 2005;Lear et al., 2008;Zhang et al., 2013]. The most recent transition, from a greenhouse to an icehouse climate state, occurred during the Eocene-Oligocene transition (EOT;~33.6-34.1 Ma). It is thought to have been driven by either a long-term decline in pCO 2 [Pagani et al., 2005;Zhang et al., 2013] and/or changes in ocean circulation and heat distribution as a result of ocean gateway reorganization [Kennett and Exon, 2004;Stickley et al., 2004;Bijl et al., 2013]. The generation of long-term, regional temperature records is essential for developing a more detailed picture of global cooling during the Eocene and elucidating the driving mechanisms responsible. calibration introduced two separate indices and calibrations for (1) the entire data set TEX 86 L ) and (2) a subset of the data set that excluded GDGT distributions from high-latitude sediments (GDGT ratio-2; TEX 86 H ) [Kim et al., 2010]. Kim et al. [2010] recommended applying TEX 86 H when SSTs are estimated to have been above 15°C and TEX 86 L where SSTs may have ranged below 15°C. However, this approach has been questioned [Taylor et al., 2013], and it is unclear which of the two calibrations is most appropriate for a given setting. The most recent TEX 86 calibration is based upon the original TEX 86 [Schouten et al., 2002] and calibrated to SST using a spatially varying, Bayesian regression model (BAYSPAR) [Tierney and Tingley, 2014].
The TEX 86 proxy is based upon the assumption that GDGTs in sediments are largely derived from Thaumarchaeota living in the upper water column [Schouten et al., 2002;Pearson and Ingalls, 2013]. However, Thaumarchaeota are not restricted to these settings and inputs of GDGTs to sediments from alternative sources will affect TEX 86 SST estimates. For example, Groups I.1a and I.1b Thaumarchaeota are present in the terrestrial environment [Jurgens et al., 1997;Ochsenreiter et al., 2003] and can bias TEX 86 SST estimates in areas with high terrigenous input Sluijs et al., 2006;Weijers et al., 2006b;. Considerable work has also explored the potential for sedimentary GDGT production to affect TEX 86 values. Particular interest has focused upon methanotrophic Weijers et al., 2011;Y. G. Zhang et al., 2011] and methanogenic [Schouten et al., 2002;Blaga et al., 2009] archaea, yet these sources are rarely discussed in deep-time investigations.
Paleotemperature reconstructions based on TEX 86 assume that Thaumarchaeota in modern oceans are representative of those living in ancient settings. In most open ocean settings, thaumarchaeotal assemblages are dominated by Group I.1a Thaumarchaeota [Pester et al., 2011] which are the putative biological source of the sedimentary GDGTs that define the TEX 86 core-top calibration. In the Red Sea, however, phylogenetically distinct archaeal communities occur both above and below the thermocline [Eder et al., 2002;Ionescu et al., 2009;Qian et al., 2011] and correspond to core-top sediments in which TEX 86 values consistently overestimate satellite-derived SST by 6-8°C [Trommer et al., 2009].
Here, we critically evaluate new and previously published GDGT distributions from Eocene sediments in order to understand the drivers of long-term cooling. Using the Methane Index (MI) [Y. G. , % GDGT-0 , and the branched and isoprenoidal tetraether (BIT) index Weijers et al., 2006b], we assess the impact of archaea other than marine Thaumarchaeota upon Eocene TEX 86 values. We also propose a new index (%GDGT RS ) which we use to tentatively identify Red Sea-type GDGT distributions within the geological record. We use ΔH-L offsets and [2]/[3] ratios [Taylor et al., 2013] to suggest the most appropriate TEX 86 calibration for a given setting. Based on those observations, we use new and previously published TEX 86 SST estimates to reconstruct spatial patterns of cooling during the Eocene (55-34 Ma) and earliest Oligocene (33-34 Ma). We compare our results with an ensemble of climate model simulations and investigate the most likely driving mechanism of long-term cooling during the descent toward the icehouse.

GDGT Analyses
To complement our data compilation, we have determined GDGT distributions from ODP Site 929 (Ceara Rise), ODP Site 913 (Greenland Basin), South Dover Bridge (Atlantic Coastal Plain), and Hampden Beach (New Zealand) using methods similar to those of previous studies [Liu et al., 2009;Hollis et al., 2012] ( Figure 1). Approximately 35-70 g of ground sediment was extracted via Soxhlet apparatus for 24 h using dichloromethane/methanol (2:1 vol/vol) as the organic solvent. The total lipid extract was subsequently separated over silica into neutral and fatty acid fractions using chloroform-saturated ammonia and chloroform:acetic acid (100:1 vol/vol), respectively [Dickson et al., 2009]. The neutral fraction was further fractionated over alumina into apolar and polar fractions using Hexane:DCM (9:1 vol/vol) and DCM:MeOH (1:2 vol/vol), respectively. The polar fraction, containing the GDGTs, was dissolved in hexane/isopropanol (99:1, vol/vol) and passed through 0.45 μm PTFE filters. Fractions were analyzed by high-performance liquid chromatography/atmospheric pressure chemical ionization-mass spectrometry (HPLC/APCI-MS) using a ThermoFisher Scientific Accela Quantum Access. Normal phase separation was achieved on an Alltech Prevail Cyano column (150 mm × 2.1 mm; 3 μm i.d.) with a flow rate of 0.2 ml min À1 . Initial solvent was hexane/isopropanol 99:1 (vol/vol), eluted isocratically for 5 min, and followed by a linear gradient to 1.8% isopropanol over 45 min. Analyses were performed in selective ion monitoring mode (SIM) to increase sensitivity and reproducibility, and [M+H] + (protonated molecular ion) GDGT peaks were integrated.

GDGT-Based SST Indices
To reconstruct SST, Kim et al. [2010] invoke two separate TEX 86 -based SST indices and calibrations. TEX 86 H uses the same combination of GDGTs as in the original TEX 86 relationship [Schouten et al., 2002;Kim et al., 2008] and is defined as where numbers refer to individual GDGT structures shown in Figure 2. GDGT index-2 is correlated to SST using the calibration equation: GDGT index-1 is correlated to SST using the calibration equation:  [Taylor et al., 2013]. Despite this, unexpectedly large ΔH-L offsets exist during the Eocene [e.g., Hollis et al., 2012]. Hollis et al. [2012] also observed that Eocene TEX 86 H SSTs are higher than those derived from inorganic proxies (i.e., Mg/Ca ratios and δ 18 O values for planktic foraminifera). As a result, Hollis et al. [2012] developed an Eocene or "paleo" calibration based on the relationship between these inorganic SST proxies and GDGT ratio-2: SST ¼ 39:036* GDGT-ratio 2 ð Þ þ 36:455 r 2 ¼ 0:87 À Á This relationship (defined as pTEX 86 ) [Hollis et al., 2012] is derived from four Eocene records in which TEX 86 indices and SSTs based on well-preserved, mixed layer planktic foraminifera have been determined for the same samples [Zachos et al., 2006;Pearson et al., 2007;Burgess et al., 2008;Hollis et al., 2009]. In the SW Pacific, this yields SST estimates that are consistently lower than TEX 86 H but are generally similar to those derived using TEX 86 L [Hollis et al., 2012]. Taylor et al. [2013] argue that the ΔH-L offset is a function of the GDGT-2/GDGT-3 ratio ([2]/[3] ratio). As this ratio is markedly higher in deeper waters than the mixed layer [Taylor et al., 2013], it is governed by export dynamics [Hernández-Sánchez et al., 2014] but also partly related to water depth. For example, deep settings (>1000 m) in the modern ocean are characterized by low ΔH-L offsets (<3.0) and high [2]/[3] ratios (>5.0), whereas shallow settings (<1000 m) are characterized by high ΔH-L values (>3.0) and low [2]/[3] ratios (<5.0). Other recent developments in TEX 86 palaeothermometry include the expansion of the core-top data set into subpolar and polar regions [Ho et al., 2014] and the development of a spatially varying, TEX 86 Bayesian regression model (BAYSPAR) [Tierney and Tingley, 2014]. In deep-time settings, BAYSPAR searches the modern core-top data set for TEX 86 values which are similar to the measured TEX 86 value and draws regression parameters from these modern "analogue" locations. SSTs are derived using an online graphical use interface (GUI) (www.whoi. edu/bayspar) [Tierney and Tingley, 2014]. Using this approach, an Eocene high-latitude site will draw analogues from a modern-day midlatitude site and so on. However, BAYSPAR does not resolve the problem of high ΔH-L offsets, as the SSTs tend to be similar to those derived from TEX 86 H [Tierney and Tingley, 2014]. This is not surprising as BAYSPAR is based upon the original TEX 86 ratio.

Other GDGT-Based Indices
A number of indices have been developed to screen for potential secondary influences on TEX 86 . The ratio of branched GDGTs to crenarchaeol ( Figure 2) in marine and lacustrine sediments is a function of terrestrial input, expressed as the Branched versus Isoprenoid Tetraether (BIT) index: Numbers refer to individual GDGT structures shown in Figure 2. It has been argued that TEX 86 estimates with BIT values >0.3 should not be used for SST reconstruction due to the potential influence of soil-derived GDGTs on temperature estimates [Weijers et al., 2006b]. Although the BIT has been applied within deeptime settings [Sluijs et al., 2011;Jenkyns et al., 2012], it is unclear whether a threshold of 0.3 remains applicable.
The Methane Index (MI) was proposed to distinguish the relative input of methanotrophic Euryarchaeota versus ammonia-oxidizing Thaumarchaeota in settings characterized by gas-hydrate-related anaerobic oxidation of methane (AOM) Wakeham et al., 2003;Stadnitskaia et al., 2008;Y. G. Zhang et al., 2011]: High MIs (>0.5) reflect high rates of gas-hydrate-related AOM and low values (<0.3) suggest normal sedimentary conditions (i.e., no appreciable AOM input); by extension, TEX 86 values should be excluded when MI values > 0.5.

Statistical Analysis
During the Eocene, TEX 86 SST records have different sampling densities and/or span different intervals [Pearson et al., 2007;Burgess et al., 2008;Bijl et al., 2009;Hollis et al., 2009;Liu et al., 2009;Hollis et al., 2012;Bijl et al., 2013]. To address this problem, time series which spanned the majority of the investigated time window (i.e., ODP 925, ODP 929, ODP 913, ODP 1172, IODP 1356, TDP, SDB, Mid-Waipara, and Hampden Beach) were grouped into low-(<30°) or high-latitude (>55°) bins. Using TEX 86 H , each time series was then turned into a relative temperature (ΔT) by comparison to the warmest temperature in that time series. In order to determine the long-term mean SST evolution in each bin (high and low latitudes) with an associated uncertainty, separate nonparametric LOESS regressions were fitted to both the low-and highlatitude TEX 86 H ΔSST compilations using the R software package (http://www.R-project.org/). The degree of smoothing (i.e., the span term) was optimized for each time series using generalized cross validation, and an uncertainty envelope (±95% confidence intervals) was calculated based upon the observed scatter of data around the best fit line. Sequential removal of one time series at a time (jackknifing) was also performed to examine the influence of each record on the long-term mean SST (see supporting information). Ma) geological stages and run for 1422 model years in total to allow surface conditions to approach equilibrium, reducing the error from model drift relative to shorter simulations (see Figure S5). Mean climate state is produced from the final 50 years of the simulation. Following an initial 50 years at 280 ppmv, atmospheric CO 2 is prescribed at 1120 ppmv (4× preindustrial level) for each simulation and with an appropriate solar constant [Gough, 1981] representative of each geologic stage defined. The initial~500 years of the model simulations have a purely baroclinic ocean circulation to ensure stability during spin-up; the barotropic circulation is initialized after 500 years. The barotropic solver in the ocean model requires the definition of continental islands, around which the net ocean flow is nonzero; the defined islands in the model are shown in Figure S6. Note that Antarctica has not been defined as an island in any of these simulations, resulting in a net ocean flow of zero around the margins of Antarctica, even though the palaeogeographic reconstruction implies a possible pathway for circum-Antarctic transport. Due to the small latitudinal extent and shallow depth of the Drake's and Tasman gateways at this time, we do not expect this to greatly affect our results. More details of the climate model itself are described in Loptson et al. [2014]; their simulation 4 × DYN is carried out with an identical model to the one used here.

Results and Discussion
For each site, including new and previously published data sets, we have determined TEX 86 SSTs during the Eocene and the Oligocene. All of these data sets are described in detail within the supporting information. Using a combination of parameters (BIT, MI, %GDGT-0), we investigate the sedimentary GDGT distributions and discard samples that are potentially problematic with respect to those prospective criteria (see sections 3-5 and supporting information). We then compare ΔH-L offsets against [2]/[3] ratios to explore the applicability of TEX 86 L before investigating spatial patterns of cooling during the Eocene (see section 6).
Based upon our findings, we also reinvestigate cooling trends during the EOT (see section 7).

Impact of Terrestrial Input Upon Eocene TEX 86 Values
The observation that branched GDGTs occur predominantly in soils whereas crenarchaeol occurs predominantly in the marine environment led to the development of the branched-to-isoprenoidal tetraether (BIT) index . Although this was originally used to elucidate the relative input of terrestrial organic matter into the marine realm, it can also provide insights into the efficacy of TEX 86 estimates Weijers et al., 2006b;Fietz et al., 2011]. Weijers et al. [2006b] show that when BIT values exceed 0.2-0.3, temperature estimates are~1°C higher than expected, and when BIT values exceed 0.4, temperature estimates can be >2°C higher. However, those observations are specific to that depositional system (the Congo Fan), and the impact of terrigenous GDGTs on reconstructed SST will depend on the nature and temperature of the source catchment. Using our Eocene and Oligocene compilation, we examine the apparent effect of terrestrial input upon TEX 86 SST estimates.
BIT values from the modern core-top data set do not exceed 0.25 in marine settings ,   Figure 1) resulted in a significant temperature deviation. They removed GDGT-3 from the original TEX 86 and developed a new index (TEX 86 ′) which was calibrated to the modern core-top data set . However, we suggest that elevated GDGT-3 is not the only impact of terrigenous OM inputs on isoprenoidal GDGT distributions; i.e., an increase in GDGT-3 due to terrestrial input will also be associated with an increase in the abundance of other isoprenoidal GDGTs. As a result, we argue that TEX 86 ′ is not a reliable alternative to SST reconstruction when terrestrial input is high.
TEX 86 SST estimates from some deep (e.g., ODP Site 929 and ODP Site 925) and shallow (e.g., South Dover Bridge) water settings are relatively unaffected by enhanced terrestrial input. Intriguingly, the few sediment samples from those sites with high BIT values (>0.4) generally yield similar SSTs as those with low BIT values (<0.1). This could be fortuitous, with terrigenous input not causing significant deviations from marine distributions, but it does suggest that the threshold of 0.4 is conservative in some settings.
In the modern core-top data set, the Methane Index (MI) spans a narrow range (0.03-0.23) and averages 0.15 (n = 426; σ = 0.07) (Figure 4b). MIs exceed 0.3 in < 1% of samples and do not exceed 0.5. As with %GDGT-0 values, this is expected for core-top sediments which are likely unaffected by methane cycling [Martens and Berner, 1974]. In gas-hydrate-impacted and/or methane-rich environments, MIs are higher (>0.6) and span a larger range (~0.6-1.0). In such settings, high MIs are associated with the presence of 13 C-depleted biphytanes, providing further evidence for a methanotrophic source Wakeham et al., 2004;Bouloubassi et al., 2006;Pancost et al., 2008;Y. G. Zhang et al., 2011]. Elevated MIs also occur in older sediments of continental marginal settings characterized by high sedimentation rate and organic matter flux [Aquilina et al., 2010;Weijers et al., 2011]. MIs span a larger range (0.08-0.82) in our Eocene and Oligocene data set (Figure 4b; n = 686) and yield a slightly higher average value (0.22; σ = 0.08) than modern core-top sediments. MIs exceed 0.3 in~8% of samples and exceed 0.5 in <2% of samples, suggesting that most Eocene and Oligocene sediments, despite their continental margin locations, are relatively unaffected by diffusive methane flux and associated anaerobic oxidation of methane.
In the Eocene and Oligocene data set, a nonlinear, positive correlation exists between MI and %GDGT-0 ( Figure S1). This is expected because sediment profiles characterized by methanogenesis will likely also have experienced some amount of anaerobic oxidation of methane [Sivan et al., 2007]. This relationship is almost certainly driven by methane cycling rather than temperature, because the latter-by decreasing % GDGT-0 and increasing MIs-would yield a negative rather than positive correlation.

Red Sea-Type GDGT Distributions
In the modern core-top calibration, sediments from the Red Sea yield much warmer TEX 86 SST estimates than observed values [Trommer et al., 2009;Ionescu et al., 2009] and are excluded from the global core-top calibration data sets of Kim et al. [2008] and Kim et al. [2010] but not the BAYSPAR calibration data set of Tierney and Tingley [2014]. Red Sea GDGT distributions are characterized by a low fractional abundance of GDGT-0 relative to Crenarchaeol regioisomer (Cren.′). To identify a typical Red Sea-type distribution within the geological record, we propose the following ratio: However, we propose this only as an approximate evaluation tool, because other factors, such as temperature [Schouten et al., 2002;Kim et al., 2010], can affect %GDGT RS indices (see later). Thus, we suggest it is initially employed to identify sediments with unusually low amounts of GDGT-0 relative to crenarchaeol regioisomer. Further evaluation of a putative Red Sea-type GDGT signature can be based on the entirety of the GDGT distribution [Trommer et al., 2009].
%GDGT RS values from the modern core-top data set (n = 396) [Kim et al., 2010] do not exceed 24, except for the Red Sea, where values range from 32 to 59 (n = 30; Figure S2) [Trommer et al., 2009]. As such, we propose hat a Red Sea-type contribution should be considered for %GDGT RS >30. In our Eocene compilation, these  Figure S4). During the Early Eocene Climatic Optimum (EECO), high %GDGT RS values become more geographically widespread, occurring at ODP Site 1172 [Bijl et al., 2009], Mid-Waipara [Hollis et al., 2009[Hollis et al., , 2012, Hampden Beach (this paper), and South Dover Bridge (this paper). At these sites, %GDGT RS values gradually increase during the EECO, attain highest values during peak EECO warmth, and then gradually decrease following the EECO ( Figure S4). Similarly, %GDGT RS values increase at the onset of the PETM at Wilson Lake [Zachos et al., 2006;Sluijs et al., 2007], ODP Site 1172 [Sluijs et al., 2011] (Figure 5), and South Dover Bridge (this paper). GDGT-0 was not detected at Bass River [Sluijs et al., 2007;. Unfortunately, it appears that most of the Red Sea GDGT characteristics are indistinguishable from those expected for temperatures in excess of~30°C (based on projecting correlations to temperatures beyond the modern limits). Therefore, we cannot currently untangle these effects on GDGT distributions in the sedimentary record.
Aside from temperature, the underlying ecological controls that govern the occurrence of these distributions remain unclear. At ODP Site 1172, the dinocyst genus Eocladopyxis, a member of the extant family Goniodomidae that mainly inhabits low-latitude lagoonal environments, peaks during the PETM and the EECO [Sluijs et al., 2011] ( Figure 5). A peak in Eocladopyxis spp. also occurs prior to and immediately after the onset of the PETM at Bass River and Wilson Lake . At all three sites, the occurrence of hypersaline dinocysts coincides with an increase in %GDGT RS values. The presence of Eocladopyxis in the Recent has been explained by hyperstratification and the development of lagoonal conditions in the open ocean [Reichart et al., 2004;. At Mid-Waipara River, the dinocyst genus Homotryblium, a similar "lagoonal" indicator genus, is also present in low abundances during the early Eocene [Hollis et al., 2009] while other high-salinity, lagoonal dinocysts, such as Heteraulacacysta and Polysphaeidium, are identified during the PETM at Bass River and Wilson Lake . Although the presence of hypersaline and/or lagoonal dinocysts is consistent with an increase in salinity, they rarely dominate the dinocyst assemblage [e.g., Sluijs et al., 2011] and it is possible that other factors exert a control upon Red Sea-type GDGT distributions.
Pure cultures of Nitrosopumilis Maritimus, a marine group I.1a thaumarchaeon, indicate that nutrient availability can influence GDGT distributions [Elling et al., 2014]. However, this contrasts with Trommer et al. [2009] who correlated Red Sea TEX 86 values with nitrate concentrations at 100 m depth and found no obvious correlation. Alternatively, Kim et al. [2015] argue that modern Red Sea GDGT distributions originate from a deep-water (>1000 m) thaumarchaeotal community. Using core-top sediments from the Mediterranean and the Red Sea, Kim et al. [2015] recently developed a regional TEX 86 SST calibration for deep-water (>1000 m), restricted basins. This yields lower TEX 86 SSTs, both in the modern and during the Eocene. However, as Eocene Red Sea-type GDGT distributions are restricted to shallow water settings (typically <500 m), this calibration is deemed unsuitable here.   [Kuypers et al., 2001;Kuypers et al., 2002]. There, a range of biomarker evidence has shown that deposition of organic-rich sediments represents an unusual and widespread expansion of archaea [Kuypers et al., 2002]. The most diagnostic biomarkers for OAE1b archaeal assemblages, i.e., tetramethylicosane (TMI), have not been reported for the Eocene sediments discussed here nor the Red Sea. This could provide additional evidence for extreme Palaeogene and Mesozoic warmth; i.e., they reflect additional changes in the GDGT distribution beyond those reflected by TEX 86 values. Alternatively, they could reflect the same factors that influence Red Sea distributions and that overestimate SST. As Red Sea GDGT characteristics are indistinguishable from those expected for temperatures in excess of~30°C; we continue to include high %GDGT RS values within our longterm Eocene compilation.

Interrogating GDGT Distributions
BIT, %GDGT-0, MI, and %GDGT RS are useful tools which can be used to flag potentially problematic TEX 86 values. However, there are limitations to a single numerical representation of these complex GDGT distributions. Figure 6a shows two sets of Eocene GDGT distributions with identical TEX 86 values (0.70). Sample 2 has a much higher %GDGT-0 value than Sample 1 and suggests an additional, potentially methanogenic, source of isoprenoidal GDGTs. Otherwise, the GDGT distribution is very similar to Sample 1 and suggests the SST reconstructions are valid. In Figure 6b, Samples 3 and 4 also have identical  latitude "greenhouse" environments. However, in mesocosm studies, the fractional abundance of the crenarchaeol isomer is ∼14 fold lower than expected and we thus argue against applying this calibration in deep-time settings. The application of a linear [Schouten et al., 2002;Tierney and Tingley, 2014], logarithmic [Kim et al., 2010] or reciprocal [Liu et al., 2009] calibration will also impact SST reconstructions, particularly in low-latitude greenhouse environments. However, the linear calibration yields unrealistically high SST values (>30°C) during the Holocene [Kim et al., 2010], and we therefore argue against its application in modern and ancient (sub)tropical climates. BAYSPAR, which also utilizes a linear calibration, does not appear to yield unrealistically high SST values during the Quaternary [e.g., Tierney and Tingley, 2014]; however, it does not formally addresses regional/oceanographic variations in deeper time reconstructions. This is because the analogue is generated by sampling within TEX 86 space (as discussed in section 2.3) rather than on the basis of oceanographic (productivity regime) or regional (water depth, circulation, seasonal) considerations. The reciprocal approach [Liu et al., 2009], which yields similar SST estimates as the logarithmic approach [Kim et al., 2010], is associated with a maximum temperature of 35°C and is therefore also unsuitable for low-latitude greenhouse environments.  [Taylor et al., 2013;Kim et al., 2015]. After discarding TEX 86 values with potentially problematic GDGT distributions (as discussed above and shown in the supporting information), we use [2]/[3] ratios and ΔH-L offsets [Taylor et al., 2013] to evaluate the various TEX 86 calibrations for each site.
In the SW Pacific (ODP Site 1172, IODP Site 1356, Mid-Waipara River, and Hampden Beach; Figure 1), high ΔH-L offsets and low [2]/[3] ratios are consistent with sediments deposited in a relatively shallow water setting. It has been shown that the lower TEX 86 L -derived SSTs are similar to inorganic and modeled SST estimates [Hollis et al., 2012;Bijl et al., 2013]. TEX 86 L -derived SSTs exhibit a stronger latitudinal temperature gradient (~10°C) than TEX 86 H , which yields much warmer SW Pacific SSTs (~27-33°C) and a low-latitudinal SST gradient. pTEX 86 , which has been calibrated to inorganic proxies, gives SW Pacific temperatures similar to those of TEX 86 L . All three calibrations exhibit a similar timing and magnitude of cooling through the Eocene [Bijl et al., 2009;Hollis et al., 2012;Bijl et al., 2013]. In the South Atlantic (Seymour Island; Figure 1), [2]/[3] ratios and ΔH-L offsets are also consistent with sediments deposited in a relatively shallow water setting. There, SSTs derived from inorganic proxies, in this case clumped isotope paleothermometry, are similar to TEX 86 L -derived SSTs but colder than TEX 86 H .
In the North Atlantic (ODP Site 913; Figure 1), TEX 86 H and TEX 86 L yield similar SSTs, consistent with sediments deposited in a deeper water setting [Myhre et al., 1995;Eldrett et al., 2004]. In contrast, SST estimates derived from pTEX 86 are significantly colder. In the western tropical Atlantic (ODP Site 925, ODP Site 929; Figure 1) In the Indian Ocean (Tanzania; Figure 1) (Figure 1), which are both shallow water settings with relatively small ΔH-L offsets. This reinforces previous arguments that water depth is not the primary control on differences between TEX 86 H and TEX 86 L -derived SSTs [Taylor et al., 2013;Kim et al., 2015]. Instead, we argue that differences are controlled by the magnitude of the subsurface GDGT contribution to sediments, which can be related to water depth but is also governed by the range of factors related to export productivity [Hernández-Sánchez et al., 2014].

10.1002/2014PA002723
Our data also challenge the simple framework that TEX 86 L is most applicable in shallow water settings. In the Atlantic (South Dover Bridge; Figure 1) and Gulf Coastal Plain [Keating-Bitonti et al., 2011], [2]/[3] ratios and ΔH-L offsets are consistent with samples deposited in a shallow setting. However, TEX 86 L SST estimates are unexpectedly low for a subtropical setting (22°C) and are, in fact, 2-3°C colder than contemporary SST estimates [Levitus and Boyer, 1994]. A similar problem has been observed in the Gulf of Mexico Coastal Plain during the late Paleocene (~15°C) and PETM (~25°C) [Sluijs et al., 2013]. At Hampden Beach, [2]/[3] ratios and ΔH-L offsets are consistent with samples deposited in a shallow setting. However, there are large variations in TEX 86 L SST estimates which are inconsistent with inorganic and organic SST estimates from nearby sites [Bijl et al., 2009;Hollis et al., 2009;Creech et al., 2010;Hollis et al., 2012;Bijl et al., 2013]. These estimates may reflect local variations in SST; alternatively, they may be related to the TEX 86 L index which is far more sensitive to contributions from other archaea and, in particular, the fractional abundance of GDGT-3.
Thus, although TEX 86 L does agree with inorganic proxies in some shallow water settings [Hollis et al., 2009;Hollis et al., 2012;Douglas et al., 2014], there are exceptions. Modern water column investigations suggest that the TEX 86 L calibration should be used with great caution. Recently, Taylor et al. [2013] showed that the increase in [2]/[3] ratios with depth is a globally widespread feature of GDGT distributions in the water column, possibly due to the predominance of different Thaumarchaeota communities in the surface mixed layer and subsurface [Villanueva et al., 2014]. The implication is that subsurface export has a markedly stronger impact on TEX 86 L values than on TEX 86 H and, by extension, that the depth-related difference between TEX 86 L -and TEX 86 H -derived SSTs is due to complexities associated with the former. As a result, the following section is restricted to the discussion of TEX 86 H -derived SSTs.

Sea Surface Temperature Change During the Eocene
Present-day SST rarely exceeds 28-29°C (except in some isolated basins), which some have suggested indicates a homeostatic limit to tropical SST [Ramanathan and Collins, 1991;Kleypas et al., 2008]. This has however been shown to be ill-founded [Pierrehumbert, 1995;van Hooidonk and Huber, 2009;Williams et al., 2009] and is not supported by SST records in the more recent geological past [O'Brien et al., 2014]. During the early and middle Eocene, SST estimates from Tanzania [Pearson et al., 2007], Ceara Rise (ODP Site 925; ODP Site 929) [Liu et al., 2009] and the Atlantic Coastal Plain (South Dover Bridge) regularly exceed this modern limit, with TEX 86 H -derived SSTs > 32°C (Figure 7). TEX 86 H SSTs, which are clearly higher than those of today, do not support the existence of a tropical "thermostat" [O'Brien et al., 2014;Pagani, 2014], at least insofar as it is most strictly defined [Ramanathan and Collins, 1991].
Previous work stipulated that if SSTs were truly~35°C in Tanzania [Pearson et al., 2007], then some tropical regions (e.g., the Western Pacific Warm Pool (WPWP) must have been much hotter [Huber, 2008]. Indeed, our modeling simulations indicate that the WPMP (~34°C) was~3-4°C warmer than Tanzania (~30-31°C) ( Figure 10). Moderately higher tropical temperatures relative to today (>2°C) will significantly increase evaporation rates, latent heat transport [Huber and Sloan, 2000], and the frequency and/or the strength of tropical cyclones [Sriver and Huber, 2007]. Tropical cyclones help to induce ocean mixing which enhances meridional overturning and ocean heat transport. This can reduce the latitudinal temperature gradient by up to 6°C and warm high-latitude oceans by as much as 10°C [Sriver and Huber, 2007;Thomas et al., 2014].
Our record also suggests tropical cooling during the Eocene, albeit of much lesser magnitude than that observed at high southern latitudes [see later ;Bijl et al., 2009;Hollis et al., 2009;Creech et al., 2010;Hollis et al., 2012;Bijl et al., 2013]. TEX 86 H indicates ≤2°C of tropical cooling within the Indian Ocean during the middle and late Eocene (45-34 Ma; Figure 8), 3-4°C of cooling within the western equatorial Atlantic during the middle and late Eocene (40-34 Ma; Figure 8), and 4-5°C of cooling within the subtropical Atlantic Coastal Plain between the early and middle Eocene (53-41 Ma; Figure 8). Crucially, middle and late Eocene (47.8-34.0 Ma) tropical cooling is apparent regardless of the calibration. By fitting a nonparametric LOESS regression to our compiled data set, we are able to determine that there was~2.5°C of long-term tropical surface water cooling between the early and late Eocene (Figure 9b) Jackknifing (the sequential removal of one record at a time) revealed that no single time series overly influences the magnitude of Eocene cooling determined by LOESS regression; however, removal of the South Dover Bridge record does change the pattern of the low-latitude long-term cooling ( Figure S7) suggests slightly cooler temperatures, perhaps coupled with increasing ice volume, in the late Eocene and early Oligocene [Pearson et al., 2007].
For comparison, a nonparametric LOESS regression was fitted through the compiled high-latitude data set. This approach indicates~6°C of high-latitude cooling between the early and late Eocene ( Figure 9c). As with the lowlatitude compilation, jackknifing revealed that no single record influences the overall magnitude of long-term high-latitude cooling determined by LOESS regression ( Figure S8). However, because the IODP 1356 time series has a very high sampling density around the EECO, its removal causes the general cross validation optimization routine to choose a relatively low degree of smoothing, such that the long-term mean high-latitude SST determined without this record exhibits more structure in the Mid and Late Eocene ( Figure S8). Nonetheless, long-term average high-latitude cooling, as indicated by TEX 86 H (and also BAYSPAR), is also in agreement with inorganic Mg/Ca SST estimates [Creech et al., 2010;Hollis et al., 2012] and δ 18 O BWT estimates  which indicate amplified polar cooling during the Eocene epoch.

Latitudinal SST Gradients During the Eocene
Our revised SST compilation provides new insights into global cooling during the descent toward the icehouse. During the early Eocene (56.0-47.8 Ma), the temperature difference (ΔT) between the tropics (2.5-4.5°N) and the SW Pacific (~55-65°) is very low (ΔT: <2°C) (Figure 7) when compared with modern conditions, as has been extensively noted and discussed elsewhere [Bijl et al., 2009;Hollis et al., 2009;Hollis et al., 2012]. Gradual cooling in the SW Pacific during the middle Eocene (47.8-38.0 Ma) progressively strengthens the southern hemisphere SST gradient (Figure 7). During the late Eocene (38.0-33.9 Ma), the latitudinal SST gradient between the SW Pacific (ODP Site 1772) and the tropics is markedly stronger than the early Eocene (ΔT:~9°C) (Figure 7) but remains much smaller than observed today (ΔT: >25°C) .
During the late middle Eocene (41.3-38.0 Ma), the temperature difference between the equatorial Atlantic (2.5-4.5°N) and the South Atlantic (52-67°S) is relatively large (ΔT: 14°C) (Figure 7). Although there is cooling in the South Atlantic during the middle late and late Eocene, the latitudinal temperature gradient Previous studies have shown that latitudinal temperature gradients of less than 20°C are difficult for climate models to simulate and require large changes in latitudinal heat transport and/or substantial positive feedbacks acting at high latitudes [Huber and Sloan, 1999;Bice et al., 2000;Huber et al., 2003;Lunt et al., 2012]. As a result, the application of TEX 86 H in high-latitude sites cannot be reconciled with modeled SSTs during the early Eocene [Hollis et al., 2012;Sijp et al., 2014]. However, a closer agreement between proxies and models can be obtained via changes in the physical parameters of the model (e.g., cloud cover) [Sagoo et al., 2013].
The apparent tropical SST stability observed by Pearson et al. [2007] suggests that mechanisms such as gateway reorganization  may have been important in regulating high-latitude cooling during the Eocene [Bijl et al., 2009. However, we note that Pearson et al. [2007] never argued that tropical SSTs were constant during the Eocene, only that SST change was much smaller than inferred from the oxygen isotopic composition of diagenetically altered foraminifera [Bralower et al., 1995;Dutton et al., 2005]. In fact, a small cooling trend (perhaps coupled with minor ice growth) is apparent in the wellpreserved foraminifera in Tanzanian sediments during the middle Eocene (47.8-38.0 Ma) [Pearson et al., 2007]. Although this is not reflected in the original low-resolution Tanzanian TEX 86 data, our new higherresolution TEX 86 data (Figures 7 and 8) and compiled tropical SST record fitted with a nonparametric LOESS regression (Figure 9) indicate the tropics cooled during the middle and late Eocene (47.8-34.0 Ma).   [Stickley et al., 2004;Bijl et al., 2013] ( Figure  10). The Drake Passage (DP) is open throughout the Ypresian and Lutetian (Figure 10), in contrast with tectonic and geochemical evidence which suggests that the DP remained closed until the early Bartonian (~41 Ma) [Scher and Martin, 2006;Livermore et al., 2007;Borrelli et al., 2014]. Despite this the total rate of transport (1.3-3 Sverdrups (Sv); Table S1) across the DP during the Ypresian and the Lutetian simulations is very small when compared to modern observations (~130 Sv) [Chidichimo et al., 2014]. Our constant-pCO 2 model simulations indicate that on a regional scale, low-latitude (<30°) SSTs decrease bỹ 0.3°C between the early and late Eocene (Figure 9a). During the same interval, compiled, proxy-derived SSTs decrease by 2.5°C (Figure 9b). Based upon this, and assuming the model and boundary conditions are not fundamentally flawed, changes in gateways and palaeogeography can only account for~10% of the low-latitude, proxy-derived cooling between the early and late Eocene. Although the magnitude of modelderived SST change varies on a site-by-site basis (see Tables S2 and S3), our results indicate that oceanographic change related to palaeogeographic change cannot account for the majority of tropical cooling. Bathymetric change (such as gateway openings) may have been responsible for other specific regional features. For example, Sijp et al. [2009] argue that opening the DP can account for~5°C of Antarctic cooling under modern-day bathymetries. However, later studies, using inferred Eocene bathymetry, indicate that the magnitude of Antarctic cooling associated with DP opening is negligible (<0.5°C) [Zhang et al., 2010;Z. Zhang et al., 2011;Lefebvre et al., 2012;Goldner et al., 2014]. Bijl et al. [2013] argue that initial deepening of the Tasman Gateway~49-50 Ma coincided with westward throughflow of the proto-Antarctic Circumpolar Current (ACC), resulting in surface water and continental cooling in the SW Pacific and along the East Antarctic margin [Pross et al., 2012;Bijl et al., 2013]. Evidence from neodymium isotopes [Scher and Martin, 2006], clumped isotope, and TEX 86 paleothermometry [Bijl et al., 2009;Douglas et al., 2014] and model simulations of intermediate complexity ] also indicate that initial opening of the Tasman Gateway is linked to the intensification of deep-water formation in the Ross Sea [Bijl et al., 2014]. Our model simulations indicate that on a regional scale, high-latitude (>55°) SSTs increase by~0.4°C between the early and late Eocene (Figure 9a). During the same interval, compiled, high-latitude proxyderived SSTs decrease by~6°C (Figure 9b). Based upon this, changes in paleogeography cannot account for the observed high-latitude, proxy-derived cooling during the Eocene (Tables S2-S4). On a local scale, high-latitude, HadCM3L-derived SSTs remain relatively stable (e.g., at the site of ACEX, 913) or increase during the Eocene (e.g., at the site of 1172, Hampden, 1356) (Table S3), indicating that changes in paleogeography are unable to explain the entirety of high-latitude cooling and that other mechanisms, such as CO 2 drawdown, must be invoked. However, it should be noted that models often struggle to replicate specific oceanographic features. For example, the subtropical East Antarctic Current (EAC) may have extended as far south as~54°during the early Eocene and could have been responsible for warming the surface waters of ODP Site 1172 and New Zealand [Hollis et al., 2012]. In contrast, many models struggle to replicate this phenomena [e.g., Lunt et al., 2012, and references therein]. HadCM3L also exhibits a relatively strong early Eocene latitudinal SST gradient compared to other models (e.g., ECHAM5 or CCSM3) [Lunt et al., 2012], in contradiction to several lines of evidence from proxies [e.g., Bijl et al., 2009].
The evolution of pCO 2 during the Eocene remains poorly constrained, particularly during the early Eocene [Beerling et al., 2011;Hyland and Sheldon, 2013]. Using TEX 86 and an ensemble of climate model simulations which span the Eocene, we conclude that the some portion of tropical cooling (~10%) can be explained by changes in paleogeography and/or ocean gateways. However, the majority of high-latitude cooling cannot be explained by changes in ocean gateways and, in the absence of other plausible forcing mechanisms, indicates that CO 2 was primarily responsible for global surface water cooling during the Eocene.

Descent Into the Icehouse
Long-term gradual cooling during the Eocene culminated in the establishment of permanent ice sheets on the Antarctic continent in the earliest Oligocene. This relatively rapid ice sheet expansion may have been driven by southern ocean gateway opening [Katz et al., 2008[Katz et al., , 2011, declining pCO 2 concentrations [DeConto and Pollard, 2003;Pearson et al., 2009;Pagani et al., 2011], or a combination of the two. During this interval, tropical TEX 86 SST estimates decrease by up to 13°C [Liu et al., 2009]. However, these values are hard to reconcile with Mg/Ca SST estimates  and U K' 37 SST estimates [Liu et al., 2009]. This suggests that parameters other than SST are controlling TEX 86 values during the EOT. Based upon our earlier discussion, we reinvestigate this possibility using the TEX 86 H proxy.
From the latest Eocene (~34-37 Ma) into the earliest Oligocene (~33-34 Ma), low-latitude TEX 86 H SST estimates decrease, on average, between 0.2 and 5.6°C. However, this does not take into account the full range of cooling which can exceed 10°C within tropical ODP Sites 998 and 803. Both sites are characterized by very high [2]/[3] ratios and low-to-negative ΔH-L offsets, suggesting the presence of "deep-water" Thaumarchaeota throughout the late Eocene and early Oligocene [Taylor et al., 2013;Kim et al., 2015]. As deep-water GDGTs can be incorporated into the sedimentary GDGT pool [e.g., Kim et al., 2015], this could account for some of the observed temperature change in tropical settings across the EOT. The intensification of Antarctic bottom water formation and enhanced equatorward transport of Antarctic intermediate water associated with Antarctic glaciation [Katz et al., 2011;Goldner et al., 2014] could have also influenced the depth of GDGT production during this interval. It certainly could have impacted the depth of and temperature change across the tropical thermocline, both of which could have impacted subsurface GDGT production, export, and recorded temperature. Other tropical settings, such as ODP 925 and ODP 929, are characterized by relatively modest cooling (~3°C) and do not appear to be affected by changes in deep-water export of GDGTs. Future studies should attempt to exploit depositional settings which are less likely to be affected by deep-water GDGT export.

Conclusions
Using new and previously published GDGT distributions, we have generated a composite TEX 86 SST record for the Eocene (55-34 Ma). To investigate the influence of archaea other than marine Thaumarchaeota upon Eocene (and Oligocene) TEX 86 values, we compiled and compared BIT indices, MIs, and %GDGT-0 values from modern and ancient sediments. Our results indicate that Eocene and Oligocene sediments have similar average values as the modern core-top data set but larger standard deviations. Nonetheless, it appears that the effect of archaea other than marine Thaumarchaeota upon Eocene and Oligocene TEX 86 values is minimal. Our compiled TEX 86 compilation indicates that between the early and late Eocene, high-latitudes SSTs cooled by~6°C and low-latitudes SST cooled by~2.5°C. Global sea surface cooling during the Eocene is not in agreement with by fixed-CO 2 HadCM3L model simulations. Therefore, our study provides indirect evidence that drawdown of CO 2 (or some, as of yet unidentified, other factor(s)) was the primary forcing for long-term climatic cooling during the Eocene. Our data set, combined with forthcoming model simulations under a range of different CO 2 levels, paves the way to reconstructing atmospheric CO 2 evolution through the Eocene.