《生物技术与人类》课程教学资源（经典阅读）Global warming preceded by increasing carbon dioxide concentrations during the last deglaciation

ARTICLE doi:10.1038/nature10915 Global warming preceded by increasing carbon dioxide concentrations during the last deglaciation Jeremy D.Shakun2,Peter U.Clark3,Feng He4,Shaun A.Marcott3,Alan C.Mix3,Zhengyu Liu4.5.6,Bette Otto-Bliesner7, Andreas Schmittner'&Edouard Bard The covariation of carbon dioxide(CO2)concentration and temperature in Antarctic ice-core records suggests a close link between CO2and climate during the Pleistocene ice ages.The role and relative importance of CO2 in producing these climate changes remains unclear,however,in part because the ice-core deuterium record reflects local rather than global temperature.Here we construct a record of global surface temperature from 80 proxy records and show that temperature is correlated with and generally lags CO,during the last(that is,the most recent)deglaciation.Differences between the respective temperature changes of the Northern Hemisphere and Southern Hemisphere parallel variations in the strength of the Atlantic meridional overturning circulation recorded in marine sediments.These observations, together with transient global climate model simulations,support the conclusion that an antiphased hemispheric temperature response to ocean circulation changes superimposed on globally in-phase warming driven by increasing CO2 concentrations is an explanation for much of the temperature change at the end of the most recent ice age. Understanding the causes of the Pleistocene ice ages has been a sig- increase in COz concentration over the last deglaciation,and that nificant question in climate dynamics since they were discovered in variations in the Atlantic meridional overturning circulation the mid-nineteenth century.The identification of orbital frequencies (AMOC)caused a seesawing of heat between the hemispheres, in the marine 150/160 record,a proxy for global ice volume,in the supporting an early hypothesis that identified potentially important 1970s demonstrated that glacial cycles are ultimately paced by astro- roles for these mechanisms".These findings,supported by transient nomical forcing'.Initial measurements of air bubbles in Antarctic ice simulations with a coupled ocean-atmosphere general circulation cores in the 1980s revealed that greenhouse gas concentrations also model,can explain the lag of COz behind Antarctic temperature in increased and decreased over the last glacial cycle2,suggesting they the ice-core record and are consistent with an important role for CO too may be part of the explanation.The ice-core record now extends in driving global climate change over glacial cycles. back 800,000 yr and shows that local Antarctic temperature was strongly correlated with and seems to have slightly led changes in Global temperature CO2 concentration.The implication of this relationship for under- We calculate the area-weighted mean of 80 globally distributed,high- standing the role of CO2 in glacial cycles,however,remains unclear. resolution proxy temperature records to reconstruct global surface For instance,proxy data have variously been interpreted to suggest temperature during the last deglaciation(Methods and Fig.1).The that COz was the primary driver of the ice ages",a more modest global temperature stack shows a two-step rise,with most warming feedback on warmingor,perhaps,largely a consequence rather than occurring during and right after the Oldest Dryas and Younger Dryas cause of past climate change.Similarly,although climate models intervals and relatively little temperature change during the Last generally require greenhouse gases to explain globalization of the Glacial Maximum (LGM),the Bolling-Allerod interval and the early ice-age signal,they predict a wide range (one-third to two-thirds)in Holocene epoch (Fig.2a).The atmospheric CO2 record from the the contribution of greenhouse gases to ice-age cooling,with addi- EPICA Dome C ice core,which has recently been placed on a more tional contributions from ice albedo and other effects".10.Moreover, accurate timescale,has a similar two-step structure and is strongly models have generally used prescribed forcings to simulate snapshots correlated with the temperature stack(r2=0.94(coefficient of deter- in time and thus by design do not distinguish the timing of changes in mination),P=0.03;Fig.2a). various forcings relative to responses. Lag correlations quantify the timing of change in the temperature Global temperature reconstructions and transient model simula- stack relative to CO2 from 20-10kyr ago,an interval that spans the tions spanning the past century and millennium have been essential to period during which low LGM CO2 concentrations increased to the attribution of recent climate change,and a similar strategy would almost pre-industrial values.Our results indicate that CO2 probably probably improve our understanding of glacial cycle dynamics.Here leads global warming over the course of the deglaciation(Fig.2b).A we use a network of proxy temperature records that provide broad comparison of the global temperature stack with Antarctic temper- spatial coverage to show that global temperature closely tracked the ature provides further support for this relative timing,in showing that Department of Earth and Planetary Sciences,Harvard University.Cambridge,,USALamont-Doherty Earth Observatory,Columbia University,Palisades,New York 10964,USA College of Earth,Ocean,and Atmospheric Sciences,Oregon State University,Corvallis,Oregon 97331,USACenter for Climatic Research,University of Wisconsin,Madison,Wisconsin 53706,USA Department of Atmospheric and Oceanic Sciences University of Wisconsin,Madison,Wisconsin 53706,USALaboratory for Ocean-Atmosphere Studies Peking University.Beijing 10871,China Climate and Global Dynamics Division,National Center for Atmospheric Research,Boulder,Colorado 80307-3000,USACEREGE,College de France,CNRS-Universite Aix-Marseille,Europole del'Arbois, 13545 Aix-en-Provence.France. 5 APRIL 2012 VOL 484 NATURE 49 2012 Macmillan Publishers Limited.All rights reserved

ARTICLE doi:10.1038/nature10915 Global warming preceded by increasing carbon dioxide concentrations during the last deglaciation Jeremy D. Shakun1,2, Peter U. Clark3 , Feng He4 , Shaun A. Marcott3 , Alan C. Mix3 , Zhengyu Liu4,5,6, Bette Otto-Bliesner7 , Andreas Schmittner3 & Edouard Bard8 The covariation of carbon dioxide (CO2) concentration and temperature in Antarctic ice-core records suggests a close link between CO2 and climate during the Pleistocene ice ages. The role and relative importance of CO2 in producing these climate changes remains unclear, however, in part because the ice-core deuterium record reflects local rather than global temperature. Here we construct a record of global surface temperature from 80 proxy records and show that temperature is correlated with and generally lags CO2 during the last (that is, the most recent) deglaciation. Differences between the respective temperature changes of the Northern Hemisphere and Southern Hemisphere parallel variations in the strength of the Atlantic meridional overturning circulation recorded in marine sediments. These observations, together with transient global climate model simulations, support the conclusion that an antiphased hemispheric temperature response to ocean circulation changes superimposed on globally in-phase warming driven by increasing CO2 concentrations is an explanation for much of the temperature change at the end of the most recent ice age. Understanding the causes of the Pleistocene ice ages has been a sig￾nificant question in climate dynamics since they were discovered in the mid-nineteenth century. The identification of orbital frequencies in the marine 18O/16O record, a proxy for global ice volume, in the 1970s demonstrated that glacial cycles are ultimately paced by astro￾nomical forcing1 . Initial measurements of air bubbles in Antarctic ice cores in the 1980s revealed that greenhouse gas concentrations also increased and decreased over the last glacial cycle2,3, suggesting they too may be part of the explanation. The ice-core record now extends back 800,000 yr and shows that local Antarctic temperature was strongly correlated with and seems to have slightly led changes in CO2 concentration4 . The implication of this relationship for under￾standing the role of CO2 in glacial cycles, however, remains unclear. For instance, proxy data have variously been interpreted to suggest that CO2 was the primary driver of the ice ages5 , a more modest feedback on warming6,7 or, perhaps, largely a consequence rather than cause of past climate change8 . Similarly, although climate models generally require greenhouse gases to explain globalization of the ice-age signal, they predict a wide range (one-third to two-thirds) in the contribution of greenhouse gases to ice-age cooling, with addi￾tional contributions from ice albedo and other effects9,10. Moreover, models have generally used prescribed forcings to simulate snapshots in time and thus by design do not distinguish the timing of changes in various forcings relative to responses. Global temperature reconstructions and transient model simula￾tions spanning the past century and millennium have been essential to the attribution of recent climate change, and a similar strategy would probably improve our understanding of glacial cycle dynamics. Here we use a network of proxy temperature records that provide broad spatial coverage to show that global temperature closely tracked the increase in CO2 concentration over the last deglaciation, and that variations in the Atlantic meridional overturning circulation (AMOC) caused a seesawing of heat between the hemispheres, supporting an early hypothesis that identified potentially important roles for these mechanisms11. These findings, supported by transient simulations with a coupled ocean–atmosphere general circulation model, can explain the lag of CO2 behind Antarctic temperature in the ice-core record and are consistent with an important role for CO2 in driving global climate change over glacial cycles. Global temperature We calculate the area-weighted mean of 80 globally distributed, high￾resolution proxy temperature records to reconstruct global surface temperature during the last deglaciation (Methods and Fig. 1). The global temperature stack shows a two-step rise, with most warming occurring during and right after the Oldest Dryas and Younger Dryas intervals and relatively little temperature change during the Last Glacial Maximum (LGM), the Bølling–Allerød interval and the early Holocene epoch (Fig. 2a). The atmospheric CO2 record from the EPICA Dome C ice core12, which has recently been placed on a more accurate timescale13, has a similar two-step structure and is strongly correlated with the temperature stack (r 2 5 0.94 (coefficient of deter￾mination), P 5 0.03; Fig. 2a). Lag correlations quantify the timing of change in the temperature stack relative to CO2 from 20–10 kyr ago, an interval that spans the period during which low LGM CO2 concentrations increased to almost pre-industrial values. Our results indicate that CO2 probably leads global warming over the course of the deglaciation (Fig. 2b). A comparison of the global temperature stack with Antarctic temper￾ature provides further support for this relative timing, in showing that 1 Department of Earth and Planetary Sciences, Harvard University, Cambridge, Massachusetts 02138, USA. 2 Lamont-Doherty Earth Observatory, Columbia University, Palisades, New York 10964, USA. 3 College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, Corvallis, Oregon 97331, USA. 4 Center for Climatic Research, University of Wisconsin, Madison, Wisconsin 53706, USA. 5 Department of Atmospheric and Oceanic Sciences, University of Wisconsin, Madison, Wisconsin 53706, USA. 6 Laboratory for Ocean-Atmosphere Studies, Peking University, Beijing 100871, China. 7 Climate and Global Dynamics Division, National Center for Atmospheric Research, Boulder, Colorado 80307-3000, USA. 8 CEREGE, Colle`ge de France, CNRS-Universite´ Aix-Marseille, Europole de l’Arbois, 13545 Aix-en-Provence, France. 5 APR IL 2012 | VOL 484 | NATURE | 49 ©2012 Macmillan Publishers Limited. All rights reserved

RESEARCH ARTICLE 90°N ords 6 0 0 30N U ●MB工C 60 0° 90° 0.05 204 0 120 180 0 Fraction of planet area Longitude Figure 1 Proxy temperature records.a,Location map.CBT,cyclization U,alkenone unsaturation index.b,Distribution of the records by latitude ratio of branched tetraethers;MBT,methylation index of branched tetraethers; (grey histogram)and areal fraction of the planet in 5 steps (blue line). TEXss tetraether index of tetraethers consisting of 86 carbon atoms; although the structure of the global stack is similar to the pattern of suggests that changes in COz concentration were either synchronous Antarctic temperature change,it lags Antarctica by several centuries with or led global warming during the various steps of the deglaciation to a millennium throughout most of the deglaciation(Fig.2a).Thus, (Supplementary Table 2).An important exception is the onset of the small apparent lead of Antarctic temperature over CO2 in the ice- deglaciation,which features about 0.3C of global warming before core records does not apply to global temperature.An additional the initial increase in CO2~17.5kyr ago.This finding suggests that evaluation of this result comes from an objective identification of CO2 was not the cause of initial warming.We return to this point inflection points in the CO2 and global temperature records,which below.Nevertheless,the overall correlation and phasing of global temperature and CO2 are consistent with CO2 being an important a driver of global warming during the deglaciation,with the centennial- 260 0 scale lag of temperature behind CO2 being consistent with the thermal inertia of the climate system owing to ocean heat uptake and ice 240 Antarctic co melting's Although other mechanisms contributed to climate change during the ice ages,climate models suggest that their impacts were regional 220 and thus cannot explain the global extent of temperature changes 8 documented by our stacked record alone.This conclusion is sup- 200 ported by the distinct differences,relative to the temperature stack,in -3 the temporal variabilities of other likely climate change agents(Fig.3). 180 For example,insolation is a smoothly varying sinusoid that is in LGM OD Holocene antiphase between the hemispheres and sums to near zero globally at 22 20 18 16 14 12 10 8 the top of the atmosphere(Fig.3f).Although spatial and temporal Age(kyr) asymmetries in albedo could convert insolation to a non-zero forcing at Earth's surface,it is unlikely to account for much of the step-like 150- structure and global nature of the temperature stack. Temperature leads CO, Temperature lags CO2 Similarly,although ice-sheet extent and its associated albedo(from Global ice cover and emergent continental shelves)and orographic forcing 100 460±340 decreased through the deglaciation,globalice volume and area changed SH only slowly or not at all during intervals of pronounced global warming 8 620±660 NH such as the Oldest Dryas and Younger Dryas,and the greatest volume 50 720±330 or area loss in fact occurred during intervals of little or no warming around 19kyr ago and the Bolling-Allerod(Fig.3a,b).This distinction 0 is particularly notable during the early Holocene,when the temperature -2.000-1.500-1,000-50005001.0001,5002.000 stack had reached interglacial levels while nearly one-third of the excess Lag (yr) global ice still remained,although we note that any ice-driven warming Figure 2 CO2 concentration and temperature.a,The global proxy would have been partly offset by decreasing greenhouse gas forcing temperature stack(blue)as deviations from the early Holocene (11.5-6.5 kyr (Fig.3c and Supplementary Fig.29a).The apparently small influence ago)mean,an Antarctic ice-core composite temperature record(red),and of ice-sheet forcing on the temperature stack is consistent with general atmospheric CO2 concentration (refs 12,13;yellow dots).The Holocene, circulation models that suggest its effect was largely confined to the Younger Dryas(YD),Bolling-Allerod(B-A),Oldest Dryas(OD)and Last northern mid to high latitudes and was otherwise modest in the areas Glacial Maximum (LGM)intervals are indicated.Error bars,1(Methods); p.p.m.v.parts per million by volume.b,The phasing of CO2 concentration and sampled by our proxy network which is biased away from the ice temperature for the global (grey),Northern Hemisphere (NH;blue)and sheets.Our results,therefore,do not preclude an important contri- Southern Hemisphere(SH;red)proxy stacks based on lag correlations from bution to global mean warming from ice-sheet retreat,but suggest that 20-10kyr ago in 1,000 Monte Carlo simulations(Methods).The mean and1 much of this warming was spatially restricted and may be inherently of the histograms are given.CO,concentration leads the global temperature under-represented owing to the lack of suitable palaeotemperature stack in 90%of the simulations and lags it in 6% records from and proximal to areas formerly covered by ice. 50 I NATURE I VOL 4845 APRIL 2012 2012 Macmillan Publishers Limited.All rights reserved

although the structure of the global stack is similar to the pattern of Antarctic temperature change, it lags Antarctica by several centuries to a millennium throughout most of the deglaciation (Fig. 2a). Thus, the small apparent lead of Antarctic temperature over CO2 in the ice￾core records12,14 does not apply to global temperature. An additional evaluation of this result comes from an objective identification of inflection points in the CO2 and global temperature records, which suggests that changes in CO2 concentration were either synchronous with or led global warming during the various steps of the deglaciation (Supplementary Table 2). An important exception is the onset of deglaciation, which features about 0.3 uC of global warming before the initial increase in CO2 ,17.5 kyr ago. This finding suggests that CO2 was not the cause of initial warming. We return to this point below. Nevertheless, the overall correlation and phasing of global temperature and CO2 are consistent with CO2 being an important driver of global warming during the deglaciation, with the centennial￾scale lag of temperature behind CO2 being consistent with the thermal inertia of the climate system owing to ocean heat uptake and ice melting15. Although other mechanisms contributed to climate change during the ice ages, climate models suggest that their impacts were regional and thus cannot explain the global extent of temperature changes documented by our stacked record alone9,16,17. This conclusion is sup￾ported by the distinct differences, relative to the temperature stack, in the temporal variabilities of other likely climate change agents (Fig. 3). For example, insolation is a smoothly varying sinusoid that is in antiphase between the hemispheres and sums to near zero globally at the top of the atmosphere (Fig. 3f). Although spatial and temporal asymmetries in albedo could convert insolation to a non-zero forcing at Earth’s surface, it is unlikely to account for much of the step-like structure and global nature of the temperature stack. Similarly, although ice-sheet extent and its associated albedo (from ice cover and emergent continental shelves) and orographic forcing decreased through the deglaciation, global ice volume and area changed only slowly or not at all during intervals of pronounced global warming such as the Oldest Dryas and Younger Dryas, and the greatest volume or area loss in fact occurred during intervals of little or no warming around 19 kyr ago and the Bølling–Allerød (Fig. 3a, b). This distinction is particularly notable during the early Holocene, when the temperature stack had reached interglacial levels while nearly one-third of the excess global ice still remained, although we note that any ice-driven warming would have been partly offset by decreasing greenhouse gas forcing (Fig. 3c and Supplementary Fig. 29a). The apparently small influence of ice-sheet forcing on the temperature stack is consistent with general circulation models that suggest its effect was largely confined to the northern mid to high latitudes and was otherwise modest in the areas sampled by our proxy network16–18, which is biased away from the ice sheets. Our results, therefore, do not preclude an important contri￾bution to global mean warming from ice-sheet retreat, but suggest that much of this warming was spatially restricted and may be inherently under-represented owing to the lack of suitable palaeotemperature records from and proximal to areas formerly covered by ice. 90° N 90° S180° 120° W 60° W 0° 60° E 120° E 180° 60° N 60° S 30° N 30° S 0° a b 90° N 60° N 30° N 0° 30° S 60° S 90° S Latitude 048 No. records 0 0.05 Fraction of planet area Mg/Ca U37 k′ Ice core TEX86 Microfossils Pollen MBT/CBT Longitude Latitude Figure 1 | Proxy temperature records. a, Location map. CBT, cyclization ratio of branched tetraethers; MBT, methylation index of branched tetraethers; TEX86, tetraether index of tetraethers consisting of 86 carbon atoms; Uk0 37, alkenone unsaturation index. b, Distribution of the records by latitude (grey histogram) and areal fraction of the planet in 5u steps (blue line). 22 20 18 16 14 12 10 8 Age (kyr) –4 –3 –2 –1 0 Proxy global temperature (°C) –1 0 1 Antarctic composite (σ) 180 200 220 240 260 CO2 (p.p.m.v.) LGM OD YDB–A Holocene –2,000 –1,500 –1,000 –500 0 500 1,000 1,500 2,000 Lag (yr) 0 50 100 150 Count Temperature leads CO2 SH –620 ± 660 NH 720 ± 330 Global 460 ± 340 Temperature lags CO2 a b Figure 2 | CO2 concentration and temperature. a, The global proxy temperature stack (blue) as deviations from the early Holocene (11.5–6.5 kyr ago) mean, an Antarctic ice-core composite temperature record42 (red), and atmospheric CO2 concentration (refs 12, 13; yellow dots). The Holocene, Younger Dryas (YD), Bølling–Allerød (B–A), Oldest Dryas (OD) and Last Glacial Maximum (LGM) intervals are indicated. Error bars, 1s (Methods); p.p.m.v., parts per million by volume. b, The phasing of CO2 concentration and temperature for the global (grey), Northern Hemisphere (NH; blue) and Southern Hemisphere (SH; red) proxy stacks based on lag correlations from 20–10 kyr ago in 1,000 Monte Carlo simulations (Methods). The mean and 1s of the histograms are given. CO2 concentration leads the global temperature stack in 90% of the simulations and lags it in 6%. RESEARCH ARTICLE 50 | NATURE | VOL 484 | 5 APRIL 2012 ©2012 Macmillan Publishers Limited. All rights reserved

ARTICLE RESEARCH 0- land and ocean proxy records during the last deglaciation,but that the timing may be slightly earlier in the marine records20(Supplementary E 20 Fig.4).Likewise,the pattern of temperature changes at upwelling sites, 40 where reservoir ages may be more variable,is similar to that at non- 40 upwelling sites but again seems somewhatolder(Supplementary Fig.4). These relationships imply that marine reservoir corrections may have 60 -80 been underestimated,which would shift the temperature stack to 260 80 later times in some intervals and increase its average lag relative to CO,.We also evaluated the EPICA Dome C CO,chronology'3 by -120 100 comparing the Dome C methane record on this timescale with the 240 8 more precisely dated Greenland composite methane record on the GICC05 timescale21.This comparison suggests that the EPICA Dome CCO2 age model may be one to two centuries too young during parts of the deglaciation(Supplementary Fig.7),which would further increase the lead of COz over global temperature.We thus regard the 200 lag of global temperature behind CO,reported here as conservative. 0 180 Hemispheric temperatures The lead of Antarctic temperature over global temperature indicates -1 spatial variability in the pattern of deglacial warming.To examine this 雪 spatial variability further,we calculated separate temperature stacks -2 for the Northern Hemisphere and Southern Hemisphere and found ature that the magnitude of deglacial warming in the two hemispheric 65°N June21 -3 stacks is nearly identical(Fig.4b).Given that greater LGM cooling probably occurred in the areas affected by the Northern Hemisphere ice sheets7,this result provides additional support for our inference Global annual that the proxy network under-represents the regional impact of the ice 65°SDec21 sheets.Each hemispheric stack also shows a two-step warming as seen in the global stack and the CO,record (Fig.4a).Otherwise,the hemi- LGM OD Holocene spheric stacks differ in two main ways.First,lag correlations suggest 22 20 161412 10 8 that whereas Southern Hemisphere temperature probably leads CO2, Age(kyr) consistent with the Antarctic ice-core results12,Northern Hemisphere temperature lags CO2(Fig.2b).Second,the Northern Hemisphere Figure 3 Global temperature and climate forcings.a,Relative sea level2 shows modest coolings coincident with the onset of Southern (diamonds).b,Northern Hemisphere ice-sheet area (line)derived from summing the extents of the Laurentide'3,Cordilleran and Scandinavian(R Hemisphere warmings,and the warming steps are concave-up in Gyllencreutz and J.Mangerud,personal communication)ice sheets through the north but are concave-down in the south (Fig.4b). time.c,Atmospheric CO2 concentration.d,Global proxy temperature stack. Calculating the temperature difference,AT,between the two hemi- e,Modelled globaltemperature stacks from the ALL(blue),CO2(red)and ORB spheric stacks yields an estimate of the heat distribution between the (green)simulations.Dashed lines show global mean temperatures in the hemispheres,and reveals two large millennial-scale oscillations that simulations,using sea surface temperatures over ocean and surface air are one-quarter to one-third of the glacial-interglacial range in global temperatures over land.f,Insolation forcing for latitudes 65N(purple)and temperature (Fig.4d).We attribute the variability in AT to variations 65S (orange)at the local summer solstice,and global mean annual insolation in the strength of the AMOC and its attendant effects on cross- (dashed black)#.Error bars,16. equatorial heat transport2.A strong correlation of ATwith a kinematic proxy(Pa/Th,the protactinium/thorium ratio)for the strength of the Unlike these regional-scale forcings,methane,nitrous oxide and AMOC24(r2=0.79,P=0.03)supports this interpretation (Fig.4g). possibly dust are global in nature.Because greenhouse gas forcing We find that ATdecreases during the Oldest Dryas and Younger Dryas was dominated by CO,(ref.19;Supplementary Fig.29a),and because intervals,when the Pa/Th record suggests that the AMOC is weak and at the onsets of the Bolling-Allerod,Younger Dryas and Holocene the heat transfer between the hemispheres is reduced,and that AT methane and nitrous oxide records have small step changes like those of increases during the LGM,the Bolling-Allerod and the Holocene, the global temperature stack,including these greenhouse gases leaves when the AMOC is stronger and transports heat from the south to the correlation with the stack essentially unchanged (r=0.93)and the north.Recalculating AT for Atlantic-only records yields the same slightly decreases the temperature lag(250+340 yr)(Supplementary relations,but they are more pronounced and better correlated with Fig.29).Global dust forcing is poorly constrained,however,and we Pa/Th(r=0.86,P=0.01),as would be expected given the importance cannot dismiss it as a potentially important driver ofglobal temperature of the AMOC in this ocean (Fig.4d).We note that this seesawing of independent of greenhouse warming.Vegetation forcing is likewise heat between the hemispheres explains the contrast between the lead of difficult to assess and may have significantly contributed to global Antarctic temperature over CO2 and the lag of global (and Northern warming.These uncertainties notwithstanding,we suggest that the Hemisphere)temperature behind CO2. increase in CO,concentration before that of global temperature is consistent with CO2 acting as a primary driver of global warming, Transient modelling although its continuing increase is presumably a feedback from changes We evaluate potential physicalexplanations for the correlations between in other aspects of the climate system. temperature,CO2 concentration and AMOC variability in three tran- The global temperature lag behind COz identified here relies sient simulations of the last deglaciation using the Community Climate critically on the chronological accuracy of these records.The largest System Model version 3(CCSM3;ref.25)of the US National Center for uncertainty in the proxy age models is associated with radiocarbon Atmospheric Research.The first simulation (ALL)runs from 22 to reservoir corrections,which affect marine,but generally not terrestrial, 6.5kyr ago and is driven by changes in greenhouse gases,insolation, records.A recent synthesis found similar temperature variabilities in ice sheets and freshwater fluxes(the last of which is adjusted iteratively 5 APRIL 2012 I VOL 484 NATURE 51 2012 Macmillan Publishers Limited.All rights reserved

Unlike these regional-scale forcings, methane, nitrous oxide and possibly dust are global in nature. Because greenhouse gas forcing was dominated by CO2 (ref. 19; Supplementary Fig. 29a), and because at the onsets of the Bølling–Allerød, Younger Dryas and Holocene the methane and nitrous oxide records have small step changes like those of the global temperature stack, including these greenhouse gases leaves the correlation with the stack essentially unchanged (r 2 5 0.93) and slightly decreases the temperature lag (250 6 340 yr) (Supplementary Fig. 29). Global dust forcing is poorly constrained19, however, and we cannot dismiss it as a potentially important driver of global temperature independent of greenhouse warming. Vegetation forcing is likewise difficult to assess19 and may have significantly contributed to global warming. These uncertainties notwithstanding, we suggest that the increase in CO2 concentration before that of global temperature is consistent with CO2 acting as a primary driver of global warming, although its continuing increase is presumably afeedbackfrom changes in other aspects of the climate system. The global temperature lag behind CO2 identified here relies critically on the chronological accuracy of these records. The largest uncertainty in the proxy age models is associated with radiocarbon reservoir corrections, which affect marine, but generally not terrestrial, records. A recent synthesis found similar temperature variabilities in land and ocean proxy records during the last deglaciation, but that the timing may be slightly earlier in the marine records20 (Supplementary Fig. 4). Likewise, the pattern of temperature changes at upwelling sites, where reservoir ages may be more variable, is similar to that at non￾upwelling sites but again seems somewhat older (Supplementary Fig. 4). These relationships imply that marine reservoir corrections may have been underestimated, which would shift the temperature stack to later times in some intervals and increase its average lag relative to CO2. We also evaluated the EPICA Dome C CO2 chronology13 by comparing the Dome C methane record on this timescale with the more precisely dated Greenland composite methane record on the GICC05 timescale21. This comparison suggests that the EPICA Dome C CO2 age model may be one to two centuries too young during parts of the deglaciation (Supplementary Fig. 7), which would further increase the lead of CO2 over global temperature. We thus regard the lag of global temperature behind CO2 reported here as conservative. Hemispheric temperatures The lead of Antarctic temperature over global temperature indicates spatial variability in the pattern of deglacial warming. To examine this spatial variability further, we calculated separate temperature stacks for the Northern Hemisphere and Southern Hemisphere and found that the magnitude of deglacial warming in the two hemispheric stacks is nearly identical (Fig. 4b). Given that greater LGM cooling probably occurred in the areas affected by the Northern Hemisphere ice sheets9,17, this result provides additional support for our inference that the proxy network under-represents the regional impact of the ice sheets. Each hemispheric stack also shows a two-step warming as seen in the global stack and the CO2 record (Fig. 4a). Otherwise, the hemi￾spheric stacks differ in two main ways. First, lag correlations suggest that whereas Southern Hemisphere temperature probably leads CO2, consistent with the Antarctic ice-core results12, Northern Hemisphere temperature lags CO2 (Fig. 2b). Second, the Northern Hemisphere shows modest coolings coincident with the onset of Southern Hemisphere warmings, and the warming steps are concave-up in the north but are concave-down in the south (Fig. 4b). Calculating the temperature difference, DT, between the two hemi￾spheric stacks yields an estimate of the heat distribution between the hemispheres, and reveals two large millennial-scale oscillations that are one-quarter to one-third of the glacial–interglacial range in global temperature (Fig. 4d). We attribute the variability in DT to variations in the strength of the AMOC and its attendant effects on cross￾equatorial heat transport22,23.A strong correlation ofDTwith a kinematic proxy (Pa/Th, the protactinium/thorium ratio) for the strength of the AMOC24 (r 2 5 0.79, P 5 0.03) supports this interpretation (Fig. 4g). We find thatDT decreases during the Oldest Dryas and Younger Dryas intervals, when the Pa/Th record suggests that the AMOC is weak and heat transfer between the hemispheres is reduced, and that DT increases during the LGM, the Bølling–Allerød and the Holocene, when the AMOC is stronger and transports heat from the south to the north. Recalculating DT for Atlantic-only records yields the same relations, but they are more pronounced and better correlated with Pa/Th (r 2 5 0.86, P 5 0.01), as would be expected given the importance of the AMOC in this ocean (Fig. 4d). We note that this seesawing of heat between the hemispheres explains the contrast between the lead of Antarctic temperature over CO2 and the lag of global (and Northern Hemisphere) temperature behind CO2. Transient modelling We evaluate potential physical explanationsfor the correlations between temperature, CO2 concentration and AMOC variability in three tran￾sient simulations of the last deglaciation using the Community Climate System Model version 3 (CCSM3; ref. 25) of the US National Center for Atmospheric Research. The first simulation (ALL) runs from 22 to 6.5 kyr ago and is driven by changes in greenhouse gases, insolation, ice sheets and freshwater fluxes (the last of which is adjusted iteratively 1416182022 1012 8 Age (kyr) –3 –2 –1 0 Proxy global temperature (°C) 180 200 220 240 260 CO2 (p.p.m.v.) 100 80 60 40 20 0 NH ice-sheet area (%) –120 –80 –40 0 Relative sea level (m) –5 0 5 10 Insolation (% relative to present) –3 –2 –1 0 Model global temperature (°C) 65° N June 21 65° S Dec. 21 Global annual a b c d e LGM OD YDB–A Holocene f Figure 3 | Global temperature and climate forcings. a, Relative sea level26 (diamonds). b, Northern Hemisphere ice-sheet area (line) derived from summing the extents of the Laurentide43, Cordilleran43 and Scandinavian (R. Gyllencreutz and J. Mangerud, personal communication) ice sheets through time. c, Atmospheric CO2 concentration. d, Global proxy temperature stack. e, Modelled global temperature stacksfrom the ALL (blue), CO2 (red) and ORB (green) simulations. Dashed lines show global mean temperatures in the simulations, using sea surface temperatures over ocean and surface air temperatures over land. f, Insolation forcing for latitudes 65u N (purple) and 65u S (orange) at the local summer solstice, and global mean annual insolation (dashed black)44. Error bars, 1s. ARTICLE RESEARCH 5 APRIL 2012 | VOL 484 | NATURE | 51 ©2012 Macmillan Publishers Limited. All rights reserved

RESEARCH ARTICLE forcing.We sample the model output at the locations of the 80 proxy 8℃车260 records,recording sea surface temperatures for the marine records and 209 surface air temperatures for the land records,and stacking these sampled model time series just as we did the data.We also correct sea surface temperature time series from the ALL simulation for sea- 0 level changes by scaling the eustatic sea-level curve26 to a warming by 0.32C at the LGM lowstand".To simulate uncertainties in the model 200 results comparable to those of the data,we generated 1,000 Monte Carlo simulations in which the modelled time series were perturbed 180 with random age-model (+300yr,1o)and temperature (1C,1a) errors at each site. The ALLmodel temperature stack from the 80 sites is similar to the global mean temperature from the model(r=0.97),suggesting that the proxy sites represent the globe fairly well,although the amplitude 》 of warming is slightly smaller in the stack(Fig.3e and Supplementary Fig.12).The ALL model stack is also similar to the CO2 model stack in shape and amplitude (r2=0.98;Fig.3e).Because the CO2 model stack reflects a response to only greenhouse gas forcing,its similarity to the ALL stack suggests that greenhouse gases can explain most of the mean warming at these 80 sites.The ORB model stack,by contrast, shows only minor warming,consistent with a modest role for orbital forcing in directly driving global temperature changes. Calculating the difference in model temperature between the Northern Hemisphere and the Southern Hemisphere at the proxy sites in the three simulations yields ATtime series that are strongly correlated with variations in modelled AMOC strength in each simu- 20. Model lation(r=0.95 for ALL,0.98 for CO2;Fig.4f).Only AT for the ALL simulation,however,shows millennial-scale variability similar to that 16 1 seen in the proxy AT time series and the Pa/Th record (Fig.4d,e,g). These results suggest that ocean circulation changes driven primarily 12 by freshwater flux,rather than by direct forcing from greenhouse gases or orbits,are plausible causes of the hemispheric differences in temperature change seen in the proxy records.Furthermore,in the ALLsimulation the Southern Hemisphere temperature stackleads Northern Hemisphere(and global)temperature during the two degla- 005 cial warming steps(Fig.4c),supporting our inference that AMOC- driven internal heat redistributions explain the Antarctic temperature -0.06 lead and global temperature lag relative to COz.Lag correlations from -0.07 231Pa/ 20-10 kyr ago suggest that the modelled global temperature lags CO2 concentration by 120 yr,which is within the uncertainty range of the -0.08 proxy-based lag. 0.09 The trigger for deglacial warming LGM OD B-A YD Holocene -0.1 The proxy database provides an opportunity to explore what triggers deglacial warming.Substantial temperature change at all latitudes 2220 18 1614 12 10 8 Age (kyr) (Fig.5b),as well as a net global warming of about 0.3C(Fig.2a), precedes the initial increase in COz concentration at 17.5kyr ago, Figure 4Hemispheric temperatures.a,Atmospheric CO2 concentration suggesting that COz did not initiate deglacial warming.This early b,Northern Hemisphere(blue)and Southern Hemisphere (red)proxy global warming occurs in two phases:a gradual increase between temperature stacks.c,Modelled Northern Hemisphere (blue)and Southern Hemisphere(red)temperature stacks from the ALL simulation.d,Northern 21.5 and 19kyr ago followed by a somewhat steeper increase between Hemisphere minus Southern Hemisphere proxy temperature stacks(dark 19 and 17.5 kyr ago(Fig.2a).The first increase is associated with mean purple).North Atlantic minus South Atlantic region proxy temperature stacks warming of the northern mid to high latitudes,most prominently in (light purple).e,Modelled Northern Hemisphere minus Southern Hemisphere Greenland,as there is little change occurring elsewhere at this time temperature stacks in the ALL(blue),CO2(red)and ORB(green)simulations. (Fig.5and Supplementary Fig.20).The second increase occurs during f,Modelled AMOC strength in the ALL(blue),CO2 (red)and ORB (green) a pronounced interhemispheric seesaw event(Fig.5),presumably simulations.g.North Atlantic sediment core OCE326-GGC52Pa/Th(ref. related to a reduction in AMOC strength,as seen in the Pa/Th record 24).Temperatures are given as deviations from the early Holocene(11.5- and our modelling(Fig.4f,g).Tropical and Southern Hemisphere 6.5kyr ago)mean.Error bars,16. warming seem to have more than offset northern extratropical cool- ing,however,perhaps as a result of an asymmetry in the response of and thus is a tunable parameter).The second simulation (CO2)is feedbacks such as Southern Ocean sea ice or tropical water vapour forced only by imposed changes in greenhouse gases (COz,methane leading to the global mean response.Alternatively,this non-zero-sum and nitrous oxide),and the third simulation(ORB)is forced only by response may reflect proxy biases,as tropical warming is not equally orbitally driven insolation variations.All other forcing factors for the evident in all proxies(Supplementary Fig.20).In any event,we second and third simulations,which run from 17 to7kyr ago,are held suggest that these spatiotemporal patterns of temperature change constant at their values at 17 kyr ago.All three simulations include are consistent with warming at northern mid to high latitudes,leading dynamic vegetation feedback and a fixed annual cycle of aerosol to a reduction in the AMOC at~19kyr ago,being the trigger for the 52 NATURE I VOL 484 5 APRIL 2012 2012 Macmillan Publishers Limited.All rights reserved

and thus is a tunable parameter). The second simulation (CO2) is forced only by imposed changes in greenhouse gases (CO2, methane and nitrous oxide), and the third simulation (ORB) is forced only by orbitally driven insolation variations. All other forcing factors for the second and third simulations, which run from 17 to 7 kyr ago, are held constant at their values at 17 kyr ago. All three simulations include dynamic vegetation feedback and a fixed annual cycle of aerosol forcing. We sample the model output at the locations of the 80 proxy records, recording sea surface temperatures for the marine records and surface air temperatures for the land records, and stacking these sampled model time series just as we did the data. We also correct sea surface temperature time series from the ALL simulation for sea￾level changes by scaling the eustatic sea-level curve26 to a warming by 0.32 uC at the LGM lowstand27. To simulate uncertainties in the model results comparable to those of the data, we generated 1,000 Monte Carlo simulations in which the modelled time series were perturbed with random age-model (6300 yr, 1s) and temperature (61 uC, 1s) errors at each site. The ALL model temperature stack from the 80 sites is similar to the global mean temperature from the model (r 2 5 0.97), suggesting that the proxy sites represent the globe fairly well, although the amplitude of warming is slightly smaller in the stack (Fig. 3e and Supplementary Fig. 12). The ALL model stack is also similar to the CO2 model stack in shape and amplitude (r2 5 0.98; Fig. 3e). Because the CO2 model stack reflects a response to only greenhouse gas forcing, its similarity to the ALL stack suggests that greenhouse gases can explain most of the mean warming at these 80 sites. The ORB model stack, by contrast, shows only minor warming, consistent with a modest role for orbital forcing in directly driving global temperature changes. Calculating the difference in model temperature between the Northern Hemisphere and the Southern Hemisphere at the proxy sites in the three simulations yields DT time series that are strongly correlated with variations in modelled AMOC strength in each simu￾lation (r 2 5 0.95 for ALL, 0.98 for CO2; Fig. 4f). Only DT for the ALL simulation, however, shows millennial-scale variability similar to that seen in the proxy DT time series and the Pa/Th record (Fig. 4d, e, g). These results suggest that ocean circulation changes driven primarily by freshwater flux, rather than by direct forcing from greenhouse gases or orbits, are plausible causes of the hemispheric differences in temperature change seen in the proxy records. Furthermore, in the ALL simulation the Southern Hemisphere temperature stack leads Northern Hemisphere (and global) temperature during the two degla￾cial warming steps (Fig. 4c), supporting our inference that AMOC￾driven internal heat redistributions explain the Antarctic temperature lead and global temperature lag relative to CO2. Lag correlations from 20–10 kyr ago suggest that the modelled global temperature lags CO2 concentration by 120 yr, which is within the uncertainty range of the proxy-based lag. The trigger for deglacial warming The proxy database provides an opportunity to explore what triggers deglacial warming. Substantial temperature change at all latitudes (Fig. 5b), as well as a net global warming of about 0.3 uC (Fig. 2a), precedes the initial increase in CO2 concentration at 17.5 kyr ago, suggesting that CO2 did not initiate deglacial warming. This early global warming occurs in two phases: a gradual increase between 21.5 and 19 kyr ago followed by a somewhat steeper increase between 19 and 17.5 kyr ago (Fig. 2a). The first increase is associated with mean warming of the northern mid to high latitudes, most prominently in Greenland, as there is little change occurring elsewhere at this time (Fig. 5 and Supplementary Fig. 20). The second increase occurs during a pronounced interhemispheric seesaw event (Fig. 5), presumably related to a reduction in AMOC strength, as seen in the Pa/Th record and our modelling (Fig. 4f, g). Tropical and Southern Hemisphere warming seem to have more than offset northern extratropical cool￾ing, however, perhaps as a result of an asymmetry in the response of feedbacks such as Southern Ocean sea ice or tropical water vapour, leading to the global mean response. Alternatively, this non-zero-sum response may reflect proxy biases, as tropical warming is not equally evident in all proxies (Supplementary Fig. 20). In any event, we suggest that these spatiotemporal patterns of temperature change are consistent with warming at northern mid to high latitudes, leading to a reduction in the AMOC at ,19 kyr ago, being the trigger for the 16182022 101214 8 Age (kyr) –4 –3 –2 –1 0 Proxy hemispheric temperature (°C) 180 200 220 240 260 CO2 (p.p.m.v.) –2 –1 0 1 Model ΔT (°C) 0.1 0.09 0.08 0.07 0.06 0.05 231Pa/230Th –3 –2 –1 0 1 Proxy ΔT (°C) 0 4 8 12 16 20 Model AMOC (Sv) –4 –3 –2 –1 0 Model hemispheric temperature (°C) a b c e f g d LGM OD YDB–A Holocene Figure 4 | Hemispheric temperatures. a, Atmospheric CO2 concentration. b, Northern Hemisphere (blue) and Southern Hemisphere (red) proxy temperature stacks. c, Modelled Northern Hemisphere (blue) and Southern Hemisphere (red) temperature stacks from the ALL simulation. d, Northern Hemisphere minus Southern Hemisphere proxy temperature stacks (dark purple). North Atlantic minus South Atlantic region proxy temperature stacks (light purple). e, Modelled Northern Hemisphere minus Southern Hemisphere temperature stacks in the ALL (blue), CO2 (red) and ORB (green) simulations. f, Modelled AMOC strength in the ALL (blue), CO2 (red) and ORB (green) simulations. g, North Atlantic sediment core OCE326-GGC5 231Pa/230Th (ref. 24). Temperatures are given as deviations from the early Holocene (11.5– 6.5 kyr ago) mean. Error bars, 1s. RESEARCH ARTICLE 52 | NATURE | VOL 484 | 5 APRIL 2012 ©2012 Macmillan Publishers Limited. All rights reserved

ARTICLE RESEARCH Temperature trend (C kyr1) latitudes over the entire deglaciation (Fig.5b)is difficult to reconcile a -2 -1 0 with hypotheses invoking a southern high-latitude trigger for degla- 90°N ciation Our global temperature stack and transient modelling point to CO2 60°N- as a key mechanism of global warming during the last deglaciation. Furthermore,our results support an interhemispheric seesawing of 30°N- heat related to AMOC variability and suggest that these internal heat redistributions explain the lead of Antarctic temperature over COz while global temperature was in phase with or slightly lagged CO2. 30°S Lastly,the global proxy database suggests that parts of the northern mid to high latitudes were the first to warm after the LGM,which 60°S could have initiated the reduction in the AMOC that may have ulti- mately caused the increase in CO,concentration. 90°SJ METHODS SUMMARY The data set compiled in this study contains most published high-resolution (median resolution,200 yr),well-dated(n=636 radiocarbon dates)temperature Onset of records from the last deglaciation(see Supplementary Information for the full 0.75 CO,rise database).Sixty-seven records are from the ocean and are interpreted to reflect sea surface temperatures,and the remaining 13 record air or lake temperatures on 6 0.5 Onset of 60-90°N land.All records span 18-11 kyr ago and ~85%of them span 22-6.5kyr ago.We seesaw 30-60°N recalibrated all radiocarbon dates with the IntCal04 calibration (Supplementary 0-30°N Information)and converted proxy units to temperature using the reservoir cor. 0.25 0-30°S rections and proxy calibrations suggested in the original publications.An excep- 30-60°S tion to this was the alkenone records,which were recalibrated with a global 60-90°S core-top calibration".The data were projected onto a 55 grid,linearly interpolated to 100-yr resolution and combined as area-weighted averages.We used Monte Carlo simulations to quantify pooled uncertainties in the age models -0.25= 22 and proxy temperatures,although we do not account for analytical uncertainties 20 16 14 12 10 8 or uncertainties related to lack ofglobal coverage and spatial bias in the data set.In Age(kyr) particular,the records are strongly biased towards ocean margins where high Figure 5 Temperature change before increase in CO2 concentration sedimentation rates facilitate the development of high-resolution records.Given a,Linear temperature trends in the proxy records from 21.5-19 kyr ago (red) these issues,we focus on the temporal evolution of temperature through the and 19-17.5kyr ago (blue)averaged in 10latitude bins with 1g uncertainties. deglaciation rather than on its amplitude of change.The global temperature stack b,Proxy temperature stacks for 30 latitude bands with 1o uncertainties.The is not particularly sensitive to interpolation resolution,areal weighting,the stacks have been normalized by the glacial-interglacial(G-IG)range in each number of proxy records,radiocarbon calibration,infilling of missing data or time series to facilitate comparison. proxy type.Details on the experimental design of the transient model simulations can be found in ref.25.The temperature stacks and proxy data set are available in Supplementary Information. global deglacial warming that followed,although more records will be required to confirm the extent and magnitude of early warming at Full Methods and any associated references are available in the online version of the paper at www.nature.com/nature. such latitudes.A possible forcing model to explain this sequence of events starts with rising boreal summer insolation driving northern Received 16 September 2011;accepted 1 February 2012. warming2.This leads to the observed retreat of Northern Hemisphere ice sheets26 and the increase in sea level2 commencing ~19kyr ago 1. Hays,J.D.Imbrie,J.Shackleton,N.J.Variations in the Earth's orbit:pacemaker (Fig.3a,b),with the attendant freshwater forcing causing a reduction of the ice ages.Science 194,1121-1132(1976) 2 Delmas,R.J.,Ascencio,J.M.Legrand,M.Polar ice evidence that atmospheric in the AMOC that warms the Southern Hemisphere through the CO2 20,000 yr BP was 50%of present.Nature 284,155-157(1980). bipolar seesaw 3 Neftel,A.Oeschger,H.,Schwander,J.,Stauffer,B.Zumbrunn,R.Ice core sample Recent studies of the deglaciation3132 have shown a strong correla- measurements give atmospheric CO2 content during the past 40,000 yr.Nature 295.220-223(1982) tion between times of minima in the AMOC and maxima in CO LOthi,D.et al.High-resolution carbon dioxide concentration record 650,000- release,consistent with our AT proxy for AMOC strength(Fig.4d), 300,000 years before present Nature 453,379-382(2008). suggesting that a change in the AMOC may have also contributed to Shackleton,N.J.The 100,000 year ice-age cycle identified and found to lag CO2 degassing from the deep Southern Ocean though its influence on temperature,carbon dioxide and orbital eccentricity.Science 289,1897-1902 (2000). the extent of Southern Ocean sea ice,the position of the southern 6. Imbrie..et al.On the structure and origin of major glaciation cycles.2.The westerliesor the efficiency of the biological pump.Further insight 100,000-year cycle.Paleoceanography 8,699-735(1993). into this relationship is provided by meridional differences in the Alley,R.B.Clark,P.U.The deglaciation of the northern hemisphere:a global perspective.Annu.Rev.Earth Planet.Sci.27,149-182 (1999). timing of proxy temperature change following the reduction in Toggweiler,J.R.Lea,D.W.Temperature differences between the hemispheres AMOC after ~19kyr ago.A near-synchronous seesaw response is and ice age climate variability.Paleoceanography 25,PA2212(2010). seen from the high northern latitudes to the mid southern latitudes, 9. Weaver,A.J.Eby,M.,Fanning.A.F.Wiebe,E.C.Simulated influence of carbon whereas strong Antarctic warming and the increase in CO2 concen- dioxide,orbital forcing and ice sheets on the climate of the Last Glacial Maximum. Nature394,847-853(1998). tration lag the AMOC change%(Figs 2a and 5b).This lag suggests that 10.Schneider von Deimling.T.,Held,H.,Ganopolski,A.Rahmstorf,S.Climate the high-southern-latitude temperature response to an AMOC per- sensitivity estimated from ensemble simulations of glacial climate.Clim.Dyn.27, 149-163(2006. turbation may involve a time constant such as that from Southern Ocean thermal inertia2337,whereas the CO2 response requires a 11.Mix,A.C.,Ruddiman,W.F.Mclntyre,A.Late Quatemary paleoceanography of the tropical Atlantic,1:spatial variability of annual mean sea-surface temperatures, threshold in AMOC reduction to displace southern winds or sea ice -20,000 years B.P.Paleoceanography 1,43-66(1986). sufficiently or to perturb the ocean's biological pump.We also 12.Monnin,E etal.Atmospheric CO2 concentrations over the last glacial termination suggest that the delay of Antarctic warming that follows the AMOC Science291,112-114(2001). 13.Lemieux-Dudon,B.et al.Consistent dating for Antarctic and Greenland ice cores seesaw event 19kyr ago and occurs relative to the mid southern Quat.Sci..Rev.29,8-20(2010) 5 APRIL 2012IVOL 484 2012 Macmillan Publishers Limited.All rights reserved

global deglacial warming that followed, although more records will be required to confirm the extent and magnitude of early warming at such latitudes. A possible forcing model to explain this sequence of events starts with rising boreal summer insolation driving northern warming28. This leads to the observed retreat of Northern Hemisphere ice sheets26 and the increase in sea level29 commencing ,19 kyr ago (Fig. 3a, b), with the attendant freshwater forcing causing a reduction in the AMOC that warms the Southern Hemisphere through the bipolar seesaw30. Recent studies of the deglaciation31,32 have shown a strong correla￾tion between times of minima in the AMOC and maxima in CO2 release, consistent with our DT proxy for AMOC strength (Fig. 4d), suggesting that a change in the AMOC may have also contributed to CO2 degassing from the deep Southern Ocean though its influence on the extent of Southern Ocean sea ice33, the position of the southern westerlies34 or the efficiency of the biological pump35. Further insight into this relationship is provided by meridional differences in the timing of proxy temperature change following the reduction in AMOC after ,19 kyr ago. A near-synchronous seesaw response is seen from the high northern latitudes to the mid southern latitudes, whereas strong Antarctic warming and the increase in CO2 concen￾tration lag the AMOC change36 (Figs 2a and 5b). This lag suggests that the high-southern-latitude temperature response to an AMOC per￾turbation may involve a time constant such as that from Southern Ocean thermal inertia23,37, whereas the CO2 response requires a threshold in AMOC reduction to displace southern winds or sea ice sufficiently38 or to perturb the ocean’s biological pump35. We also suggest that the delay of Antarctic warming that follows the AMOC seesaw event 19 kyr ago and occurs relative to the mid southern latitudes over the entire deglaciation (Fig. 5b) is difficult to reconcile with hypotheses invoking a southern high-latitude trigger for degla￾ciation39,40. Our global temperature stack and transient modelling point to CO2 as a key mechanism of global warming during the last deglaciation. Furthermore, our results support an interhemispheric seesawing of heat related to AMOC variability and suggest that these internal heat redistributions explain the lead of Antarctic temperature over CO2 while global temperature was in phase with or slightly lagged CO2. Lastly, the global proxy database suggests that parts of the northern mid to high latitudes were the first to warm after the LGM, which could have initiated the reduction in the AMOC that may have ulti￾mately caused the increase in CO2 concentration. METHODS SUMMARY The data set compiled in this study contains most published high-resolution (median resolution, 200 yr), well-dated (n 5 636 radiocarbon dates) temperature records from the last deglaciation (see Supplementary Information for the full database). Sixty-seven records are from the ocean and are interpreted to reflect sea surface temperatures, and the remaining 13 record air or lake temperatures on land. All records span 18–11 kyr ago and ,85% of them span 22–6.5 kyr ago. We recalibrated all radiocarbon dates with the IntCal04 calibration (Supplementary Information) and converted proxy units to temperature using the reservoir cor￾rections and proxy calibrations suggested in the original publications. An excep￾tion to this was the alkenone records, which were recalibrated with a global core-top calibration41. The data were projected onto a 5u 3 5u grid, linearly interpolated to 100-yr resolution and combined as area-weighted averages. We used Monte Carlo simulations to quantify pooled uncertainties in the age models and proxy temperatures, although we do not account for analytical uncertainties or uncertainties related to lack of global coverage and spatial bias in the data set. In particular, the records are strongly biased towards ocean margins where high sedimentation rates facilitate the development of high-resolution records. Given these issues, we focus on the temporal evolution of temperature through the deglaciation rather than on its amplitude of change. The global temperature stack is not particularly sensitive to interpolation resolution, areal weighting, the number of proxy records, radiocarbon calibration, infilling of missing data or proxy type. Details on the experimental design of the transient model simulations can be found in ref. 25. The temperature stacks and proxy data set are available in Supplementary Information. Full Methods and any associated references are available in the online version of the paper at www.nature.com/nature. Received 16 September 2011; accepted 1 February 2012. 1. Hays, J. D., Imbrie, J. & Shackleton, N. J. Variations in the Earth’s orbit: pacemaker of the ice ages. Science 194, 1121–1132 (1976). 2. Delmas, R. J., Ascencio, J. M. & Legrand, M. Polar ice evidence that atmospheric CO2 20,000 yr BP was 50% of present. Nature 284, 155–157 (1980). 3. Neftel, A., Oeschger, H., Schwander, J., Stauffer, B. & Zumbrunn, R. Ice core sample measurements give atmospheric CO2 content during the past 40,000 yr. Nature 295, 220–223 (1982). 4. Lu¨thi, D. et al. High-resolution carbon dioxide concentration record 650,000– 800,000 years before present. Nature 453, 379–382 (2008). 5. Shackleton, N. J. The 100,000 year ice-age cycle identified and found to lag temperature, carbon dioxide and orbital eccentricity. Science 289, 1897–1902 (2000). 6. Imbrie, J. et al. On the structure and origin of major glaciation cycles. 2. The 100,000-year cycle. Paleoceanography 8, 699–735 (1993). 7. Alley, R. B. & Clark, P. U. The deglaciation of the northern hemisphere: a global perspective. Annu. Rev. Earth Planet. Sci. 27, 149–182 (1999). 8. Toggweiler, J. R. & Lea, D. W. Temperature differences between the hemispheres and ice age climate variability. Paleoceanography 25, PA2212 (2010). 9. Weaver, A. J., Eby, M., Fanning, A. F. & Wiebe, E. C. Simulated influence of carbon dioxide, orbital forcing and ice sheets on the climate of the Last Glacial Maximum. Nature 394, 847–853 (1998). 10. Schneider von Deimling, T., Held, H., Ganopolski, A. & Rahmstorf, S. Climate sensitivity estimated from ensemble simulations of glacial climate. Clim. Dyn. 27, 149–163 (2006). 11. Mix, A. C., Ruddiman, W. F. & McIntyre, A. Late Quaternary paleoceanography of the tropical Atlantic, 1: spatial variability of annual mean sea-surface temperatures, 0–20,000 years B.P. Paleoceanography 1, 43–66 (1986). 12. Monnin, E. et al. Atmospheric CO2 concentrations over the last glacial termination. Science 291, 112–114 (2001). 13. Lemieux-Dudon, B. et al. Consistent dating for Antarctic and Greenland ice cores. Quat. Sci. Rev. 29, 8–20 (2010). 90° N 60° N 30° N 0° 30° S 60° S 90° S Latitude –2 –1 0 1 2 Temperature trend (°C kyr–1) 1 0.75 0.5 0.25 0 Temperature (fraction of G–IG range) –0.25 22 20 18 16 14 12 10 8 Age (kyr) a b Onset of CO2 rise Onset of seesaw 60–90° N 30–60° N 0–30° S 0–30° N 30–60° S 60–90° S Figure 5 | Temperature change before increase in CO2 concentration. a, Linear temperature trends in the proxy records from 21.5–19 kyr ago (red) and 19–17.5 kyr ago (blue) averaged in 10u latitude bins with 1s uncertainties. b, Proxy temperature stacks for 30u latitude bands with 1s uncertainties. The stacks have been normalized by the glacial–interglacial (G–IG) range in each time series to facilitate comparison. ARTICLE RESEARCH 5 APRIL 2012 | VOL 484 | NATURE | 53 ©2012 Macmillan Publishers Limited. All rights reserved

RESEARCH ARTICLE 14.Fischer.H..Wahlen.M..Smith.J..Mastroianni.D.Deck.B.Ice core records of 37.Schmittner,A.Saenko,0.Weaver,A.J.Coupling of the hemispheres in atmospheric CO2 around the last three glacial terminations.Science 283, observations and simulations of glacial climate change.Quat Sci.Rev.22 1712-1714(1999). 659-671(2003). 15.Hansen,J.et al Climate response times:dependence on climate sensitivity and 38.Anderson,R.F.etal Wind-driven upwelling in the Southem Ocean and the ocean mixing.Science 229,857-859 (1985). deglacial rise in atmospheric CO2.Science 323,1443-1448 (2009). 16.Manabe,S.Broccoli,A.J.The influence of continental ice sheets on the climate of 39.Stott,L,Timmermann,A.Thunell,R.Southern hemisphere and deep-sea an ice age.J.Geophys.Res.90,2167-2190 (1985). warming led deglacial atmospheric CO2 rise and tropical warming.Science 318 17.Broccoli,A.J.Tropical cooling at the Last Glacial Maximum:an atmosphere-mixed 435-438(2007). layer ocean model simulation.J.Clim.13,951-976(2000). 18.Chiang.J.C.H.Bitz,C.M.Influence of high latitude ice cover on the marine 40.Huybers,P.&Denton,G.Antarctic temperature at orbital timescales controlled by Intertropical Convergence Zone.Clim.Dyn.25,477-496(2005). local summer duration.Nature Geosci.1,787-792 (2008). 19.Jansen.E etal.in Climate Change 2007:The Physical Science Basis(eds Solomon. 41.Muller,P.J.Kirst,G.Ruhland,G.von Storch,I.Rosell-Mele,A.Calibration of the alkenone paleotemperature index Ua based on core-tops from the easter S.et al)433-497(Cambridge Univ.Press,2007) 20.Clark,P.U.et al Global climate evolution during the last deglaciation.Proc.Natl South Atlantic and the global ocean(60N-60S).Geochim.Cosmochim.Acta 62. 1757-1772(1998. Acad.Sci.USA advance online publication doi:10.1073/pnas.1116619109 (13 February 2012). 42.Pedro,J.B.etal.The last deglaciation:timing the bipolar seesaw.Clim.Past Discuss 21.Blunier,T.et al.Sy nchronization of ice core records via atmospheric gases.Clim 7.397-430(2011). Past3.325-330(2007). Dyke,AS.in Quatemary Glaciations:Extent and Chronology Vol.2b(eds Ehlers,J.& 22.Crowley.T.J.North Atlantic Deep Water cools the Southern Hemisphere. Gibbard,P.L)373-424 (Elsevier,2004). Paleoceanography 7,489-497(1992). 44. Laskar,J.et al.A long term numerical solution for the insolation quantities of the 23.Stocker.T.F.Johnsen,S.J.A minimum thermodynamic model for the bipolar Earth.Astron.Astrophys.428,261-285(2004). seesaw.Paleoceanography 18,1087(2003). 24.McManus.J.F..Francois,R.Gherardi,J.-M.,Keigwin,L D.Brown-Leger,S. Supplementary Information is linked to the online version of the paper at www.nature.com/nature. Collapse and rapid resumption of Atlantic meridional circulation linked to deglacial climate changes.Nature 428,834-837(2004) Acknowledgements Discussions with numerous people,including E.J.Brook. 25.Liu,Z.et al.Transient simulation of last deglaciation with a new mechanism for A.E.Carlson,N.G.Pisias and J.Shaman,contributed to this research.We acknowledge Bolling-Allered warming.Science 325,310-314 (2009). the palaeoclimate community for generating the proxy data sets used here.In 26.Clark.P.U.et al The Last Glacial Maximum.Science 325,710-714 (2009) particular,we thank S.Barker,T.Barrows,E.Calvo.J.Kaiser,A.Koutavas,Y.Kubota. 27. Schmittner,A.et al.Climate sensitivity estimated from temperature V.Peck C.Pelejero,J.-R.Petit,J.Sachs.E SchefuB.J.Tierney and G.Wei for providing reconstructions of the Last Glacial Maximum.Science 334,1385-1388 (2011). proxy data,and R.Gyllencreutz and J.Mangerud for providing unpublished results of 28.Alley,R.B.Brook,E.J.Anandakrishnan,S.A northern lead in the orbital band: the DATED Project on the retreat history of the Eurasian ice sheets.The NOAA NGDC north-south phasing of Ice-Age events.Quat Sci.Rev.21,431-441 (2002). and PANGAEA databases were also essential to this work.This research used resources of the Oak Ridge Leadership Computing Facility,located in the National Center for Computational Sciences at Oak Ridge National Laboratory,which is supported by the (2000). Office of Science of the Department of Energy under contract no. 30.Clark,P.U.McCabe,A M.,Mix,A.C.Weaver,A.J.Rapid rise of sea level 19,000 years ago and its global implications.Science 304,1141-1144 (2004). DE-AC05-000R22725.NCAR is sponsored by the NSF.J.D.S.is supported by a NOAA 31.Marchitto,T.M.Lehman,S.J.Ortiz,J.D.Fluckiger,.&van Geen,A.Marine Climate and Global Change Postdoctoral Fellowship.This research was supported by the NSF Paleoclimate Program for the Paleovar Project through grant AGS-0602395 radiocarbon evidence for the mechanism of deglacial atmospheric CO2 rise. Science316,1456-1459(2007) Author Contributions J.D.S.designed the study,synthesized and analysed data,and 32.Skinner,L.C..Fallon,S.,Waelbroeck,C.Michel,E.Barker,S.Ventilation of the wrote the manuscript with P.U.C.F.H.,Z.L.and B.O.-B.did the transient modelling. deep Southern Ocean and deglacial CO2 rise.Science 328,1147-1151(2010) SAM.and AC.M.contributed to data analysis.AS.helped interpret AMOC-CO2 33.Stephens,B.B.Keeling.R.F.The influence of Antarctic sea ice on glacial- linkages.E.B.provided data and discussion on the radiocarbon calibration.All authors interglacial CO variations.Nature 404,171-174(2000). discussed the results and provided input on the manuscript 34.Toggweiler,J.R,Russell,J.L.Carson,S.R.Midlatitude westerlies,atmospheric CO2,and climate changeduring the ice ages.Paleoceanography 21,PA2005(2006). Author Information Reprints and permissions information is available at 35.Schmittner,A.Galbraith,E.D.Glacial greenhouse-gas fluctuations controlled by www.nature.com/reprints.The authors declare no competing financial interests ocean circulation changes.Nature 456,373-376(2008). Readers are welcome to comment on the online version of this article at 36.Barker,S.et al Interhemispheric Atlantic seesaw response during the last www.nature.com/nature.Correspondence and requests for materials should be deglaciation.Nature 457,1097-1102 (2009). addressed to J.D.S.(shakun@fas.harvard.edu). 54 NATURE VOL APRIL 2012 2012 Macmillan Publishers Limited.All rights reserved

14. Fischer, H., Wahlen, M., Smith, J., Mastroianni, D. & Deck, B. Ice core records of atmospheric CO2 around the last three glacial terminations. Science 283, 1712–1714 (1999). 15. Hansen, J. et al. Climate response times: dependence on climate sensitivity and ocean mixing. Science 229, 857–859 (1985). 16. Manabe, S. & Broccoli, A. J. The influence of continental ice sheets on the climate of an ice age. J. Geophys. Res. 90, 2167–2190 (1985). 17. Broccoli, A. J. Tropical cooling at the Last Glacial Maximum: an atmosphere-mixed layer ocean model simulation. J. Clim. 13, 951–976 (2000). 18. Chiang, J. C. H. & Bitz, C. M. Influence of high latitude ice cover on the marine Intertropical Convergence Zone. Clim. Dyn. 25, 477–496 (2005). 19. Jansen, E. et al. in Climate Change 2007: The Physical Science Basis (eds Solomon, S. et al.) 433–497 (Cambridge Univ. Press, 2007). 20. Clark, P. U. et al. Global climate evolution during the last deglaciation. Proc. Natl Acad. Sci. USA advance online publication doi:10.1073/pnas.1116619109 (13 February 2012). 21. Blunier, T. et al. Synchronization of ice core records via atmospheric gases. Clim. Past 3, 325–330 (2007). 22. Crowley, T. J. North Atlantic Deep Water cools the Southern Hemisphere. Paleoceanography 7, 489–497 (1992). 23. Stocker, T. F. & Johnsen, S. J. A minimum thermodynamic model for the bipolar seesaw. Paleoceanography 18, 1087 (2003). 24. McManus, J. F., Francois, R., Gherardi, J.-M., Keigwin, L. D. & Brown-Leger, S. Collapse and rapid resumption of Atlantic meridional circulation linked to deglacial climate changes. Nature 428, 834–837 (2004). 25. Liu, Z. et al. Transient simulation of last deglaciation with a new mechanism for Bølling-Allerød warming. Science 325, 310–314 (2009). 26. Clark, P. U. et al. The Last Glacial Maximum. Science 325, 710–714 (2009). 27. Schmittner, A. et al. Climate sensitivity estimated from temperature reconstructions of the Last Glacial Maximum. Science 334, 1385–1388 (2011). 28. Alley, R. B., Brook, E. J. & Anandakrishnan, S. A northern lead in the orbital band: north-south phasing of Ice-Age events. Quat. Sci. Rev. 21, 431–441 (2002). 29. Yokoyama, Y., Lambeck, K., De Deckker, P., Johnston, P. & Fifield, L. K. Timing of the Last Glacial Maximum from observed sea-level minima. Nature 406, 713–716 (2000). 30. Clark, P. U., McCabe, A. M., Mix, A. C. & Weaver, A. J. Rapid rise of sea level 19,000 years ago and its global implications. Science 304, 1141–1144 (2004). 31. Marchitto, T. M., Lehman, S. J., Ortiz, J. D., Fluckiger, J. & van Geen, A. Marine radiocarbon evidence for the mechanism of deglacial atmospheric CO2 rise. Science 316, 1456–1459 (2007). 32. Skinner, L. C., Fallon, S., Waelbroeck, C., Michel, E. & Barker, S. Ventilation of the deep Southern Ocean and deglacial CO2 rise. Science 328, 1147–1151 (2010). 33. Stephens, B. B. & Keeling, R. F. The influence of Antarctic sea ice on glacial￾interglacial CO2 variations. Nature 404, 171–174 (2000). 34. Toggweiler, J. R., Russell, J. L. & Carson, S. R. Midlatitude westerlies, atmospheric CO2, and climate changeduring the ice ages.Paleoceanography21,PA2005 (2006). 35. Schmittner, A. & Galbraith, E. D. Glacial greenhouse-gas fluctuations controlled by ocean circulation changes. Nature 456, 373–376 (2008). 36. Barker, S. et al. Interhemispheric Atlantic seesaw response during the last deglaciation. Nature 457, 1097–1102 (2009). 37. Schmittner, A., Saenko, O. & Weaver, A. J. Coupling of the hemispheres in observations and simulations of glacial climate change. Quat. Sci. Rev. 22, 659–671 (2003). 38. Anderson, R. F. et al. Wind-driven upwelling in the Southern Ocean and the deglacial rise in atmospheric CO2. Science 323, 1443–1448 (2009). 39. Stott, L., Timmermann, A. & Thunell, R. Southern hemisphere and deep-sea warming led deglacial atmospheric CO2 rise and tropical warming. Science 318, 435–438 (2007). 40. Huybers, P. & Denton, G. Antarctic temperature at orbital timescales controlled by local summer duration. Nature Geosci. 1, 787–792 (2008). 41. Mu¨ller, P. J., Kirst, G., Ruhland, G., von Storch, I. & Rosell-Mele, A. Calibration of the alkenone paleotemperature index U37K’ based on core-tops from the eastern South Atlantic and the global ocean (60uN-60uS). Geochim. Cosmochim. Acta 62, 1757–1772 (1998). 42. Pedro, J. B. et al.The last deglaciation: timing the bipolar seesaw.Clim. Past Discuss. 7, 397–430 (2011). 43. Dyke, A. S. in Quaternary Glaciations: Extent and Chronology Vol. 2b (eds Ehlers, J. & Gibbard, P. L.) 373–424 (Elsevier, 2004). 44. Laskar, J. et al. A long term numerical solution for the insolation quantities of the Earth. Astron. Astrophys. 428, 261–285 (2004). Supplementary Information is linked to the online version of the paper at www.nature.com/nature. Acknowledgements Discussions with numerous people, including E. J. Brook, A. E. Carlson, N. G. Pisias and J. Shaman, contributed to this research. We acknowledge the palaeoclimate community for generating the proxy data sets used here. In particular, we thank S. Barker, T. Barrows, E. Calvo, J. Kaiser, A. Koutavas, Y. Kubota, V. Peck, C. Pelejero, J.-R. Petit, J. Sachs, E. Schefuß, J. Tierney and G. Wei for providing proxy data, and R. Gyllencreutz and J. Mangerud for providing unpublished results of the DATED Project on the retreat history of the Eurasian ice sheets. The NOAA NGDC and PANGAEA databases were also essential to this work. This research used resources of the Oak Ridge Leadership Computing Facility, located in the National Center for Computational Sciences at Oak Ridge National Laboratory, which is supported by the Office of Science of the Department of Energy under contract no. DE-AC05-00OR22725. NCAR is sponsored by the NSF. J.D.S. is supported by a NOAA Climate and Global Change Postdoctoral Fellowship. This research was supported by the NSF Paleoclimate Program for the Paleovar Project through grant AGS-0602395. Author Contributions J.D.S. designed the study, synthesized and analysed data, and wrote the manuscript with P.U.C. F.H., Z.L. and B.O.-B. did the transient modelling. S.A.M. and A.C.M. contributed to data analysis. A.S. helped interpret AMOC–CO2 linkages. E.B. provided data and discussion on the radiocarbon calibration. All authors discussed the results and provided input on the manuscript. Author Information Reprints and permissions information is available at www.nature.com/reprints. The authors declare no competing financial interests. Readers are welcome to comment on the online version of this article at www.nature.com/nature. Correspondence and requests for materials should be addressed to J.D.S. (shakun@fas.harvard.edu). RESEARCH ARTICLE 54 | NATURE | VOL 484 | 5 APR IL 2012 ©2012 Macmillan Publishers Limited. All rights reserved

ARTICLE RESEARCH METHODS 500-yr resolution but this did not change the time series or its uncertainty. Age control.All radiocarbon dates were recalibrated using Calib 6.0.1 with the Differences in areal weighting affect the glacial-interglacial amplitude of the stack IntCal04 calibration and the reservoir corrections suggested in the original pub but have little impact on its structure.Jackknifing suggests that the stack is not lications.Age models based on tuning were left unchanged from the original particularly sensitive to the number of records used.A leave-one-out proxy publications.We used the GICC05 timescale for NGRIP and GRIP,the timescale jackknifing approach suggests that the correlation (r2=0.90-0.95)and lead- of ref.13 for EDML and Dome C,and glaciological age models for the Dome F lag relation (300-600-yr temperature lag)between global temperature and CO, and Vostok ice cores. concentration are not sensitive to proxy type.Statistical infilling of missing data Proxy temperatures.We converted proxy units to temperature for all alkenone, values has negligible impact on the results.Although we here use the IntCal04 Mg/Ca and TEX86 records using the calibrations suggested by the original authors radiocarbon calibration,we tested the sensitivity of our results to this choice by for Mg/Ca and TEX86 and the global core-top calibration for alkenone records" recalibrating radiocarbon dates using the IntCal09 calibration.This makes the We used the published temperature reconstructions for Antarctic ice-core,pollen, global stack up to 350 yr older during the Heinrich 1 interval,and shifts the overall microfossil assemblage and MBT/CBT records and the GISP2 borehole calibration phase lag relative to CO2 concentration from 460+340 to 350+340yr.We for the Greenland ice corests.Missing data values near the beginningand end of the consider the IntCal04 calibration to be more accurate for the reasons discussed ~15%of records not spanning the entire study interval were infilled using the in Supplementary Information.Lag correlations suggest that Antarctic temper. method of regularized expectation maximization. ature led CO:concentration slightly throughout the deglaciation,whereas global Stacking.The proxy data were projected ontoa55grid,linearly interpolated temperature led CO2 concentration at the onset of deglaciation but lagged behind to 100-yr resolution and combined as area-weighted averages.We do not it thereafter.The lead-lag relation between CO2 concentration and the global otherwise account for spatial biases in the dataset or lack of global coverage. temperature stack is not significantly changed by detrending the time series to Uncertainty analysis.There are two main sources of uncertainty in the proxy remove the deglacial ramp in each quantity.The significance levels of the correla- records:age models and temperature calibration.We used a Monte Carlo tions between global temperature and CO2 concentration and between Pa/Th and approach to generate 1,000 realizations of each proxy temperature record after AT were determined by calculating effective sample sizes of these highly auto- perturbing the records with chronological and temperature errors.These per- correlated time series (CO2 concentration,4.3;global temperature,4.1;Pa/Th turbed records were then averaged to yield 1,000 realizations of the global and 5.4:AT,6.2;Atlantic AT,5.5)using equation 6.26 of ref.54.See Supplementary hemispheric temperature stacks.The error bars on the temperature stacks repres- Information for more discussion of these tests. ent the standard deviations of these 1,000 realizations,which provide an estimate Model freshwater forcing.Whereas the forcing from insolation,greenhouse of the propagated uncertainty due to uncertainties in the individual proxy records. gases and ice sheets during the deglaciation are fairly well constrained,freshwater A similar approach was applied to the transient model output to develop the forcing is comparatively uncertain.Several model freshwater schemes were modelled temperature stacks and error bars.We developed continuous uncer- tested,and the final run was based on the meltwater scenario (Supplementary tainty estimates for radiocarbon-based chronologies,taking into account radio- Fig.30)that produced North Atlantic climate variability in best agreement with carbon date errors as well as interpolation uncertainty between dates using a proxy reconstructions.The raw modelled AMOC time series(Fig.4f,thin lines) random walk model".Age-model uncertainties for tuned records,the Dome F were effectively smoothed with a Monte Carlo approach similar to the one used to and Vostok ice cores,and regional temperature reconstructions for Beringiawere develop the modelled temperature stacks(Fig.4e),to facilitate direct comparison assumed to be 2%(1).We used the Dome C and EDML ice-core age-model of the two.More specifically,the smoothed AMOC time series(Fig.4f,bold lines) uncertaintiesand GICCo5 maximum counting errors as 2 uncertainties are the means of 1,000 realizations of the raw AMOC time series generated by for the NGRIP and GRIP ice cores as suggested in ref.50.We used the following perturbing them with 300-yr(1)age-model errors. 1o temperature calibration uncertainties:alkenones,T=(U-0.044 +0.016)/ (0.033±0.001)(ref.41Mg/Ca=±0.02Bexp(±0.003AT),where A and B are 45.Cuffey,K.M.Clow,G.D.Temperature,accumulation,and ice sheet elevation in constants (ref.51);TEX86.+1.7C (ref.52);ice cores,10%(ref.53);pollen, heopyes 102. microfossil assemblages and MBT/CBT,1.5C.We did not account for 46.Schneider,T.Analysis of incomplete climate data:estimation of mean values and analytical uncertainties in proxy measurements.Chronological errors in the covariance matrices and imputation of missing values.J.Clim.14,853-871 Monte Carlo simulations were temporally autocorrelated but were random in 2001 space (but this does not account for systematic errors among the proxy records 47.Huybers,P.Wunsch,C.A depth-derived Pleistocene age model:uncertainty due to uncertainties in the radiocarbon calibration),whereas temperature errors estimates,sedimentation variability,and nonlinear climate change. Paleoceanography 19,PA1028(2004). were assumed to be random in space and time.We note that our study is concerned 48.Viau,A.E,Gajewski,K.Sawada,M.C.Bunbury,J.Low-and high-frequency with temperature anomalies and is thus sensitive to relative but not absolute climate variability in eastern Beringia during the past 25 000 years.Can.J.Earth temperature errors in a proxy record.See Supplementary Information for more Sc.45,1435-1453(2008). details and examples.Age-model uncertainties for the Antarctic Dome C CO2 49.Rasmussen,S.O.et al Synchronization of the NGRIP,GRIP,and GISP2 ice cores record related to methane synchronization to Greenland were estimated on the across MIS 2 and palaeoclimatic implications.Quat.Sci.Rev.27,18-28(2008). basis of the combined uncertainties associated with Greenland layer counting. 50. Svensson,A etal.A 60000 year Greenland stratigraphic ice core chronology.Clim Past4.47-57(2008). Greenland ice-age/gas-age differences and methane tuning to Antarctica.These 51.Anand,P.Elderfield,H.Conte,M.H.Calibration of Mg/Ca thermometry in uncertainties were used to generate 1,000 realizations of the CO record,which planktonic foraminifera from a sediment trap time series.Paleoceanography 18, together with the 1,000 temperature stack realizations yield the 1,000 temperature- 1050(2003). concentration lead-lag estimates shown in Fig.2b 52.Kim,J.H,Schouten,S.,Hopmans,E C.,Donner,B.Damste,J.S.S.Global Robustness of results.We evaluated how well the proxy sites represent the globe sediment core-top calibration of the TEXB6 paleothermometer in the ocean. Geochim.Cosmochim.Acta 72,1154-1173(2008). by subsampling the twentieth-century instrumental temperature record and our 53.Jouzel,J.et al Magnitude of isotope/temperature scaling for interpretation of transient modelling output of the deglaciation at the 80 proxy sites.Both central Antarctic ice cores.J.Geophys.Res.108,4361 (2003). approaches suggested that the mean of the proxy sites approximates the global 54.von Storch,H.Zwiers,F.W.Statistical Analysis in Climate Research 115 mean fairly well.We recalculated the stack after interpolating the records to (Cambridge Univ.Press,1999). 2012 Macmillan Publishers Limited.All rights reserved

METHODS Age control. All radiocarbon dates were recalibrated using Calib 6.0.1 with the IntCal04 calibration and the reservoir corrections suggested in the original pub￾lications. Age models based on tuning were left unchanged from the original publications. We used the GICC05 timescale for NGRIP and GRIP, the timescale of ref. 13 for EDML and Dome C, and glaciological age models for the Dome F and Vostok ice cores. Proxy temperatures. We converted proxy units to temperature for all alkenone, Mg/Ca and TEX86 records using the calibrations suggested by the original authors for Mg/Ca and TEX86 and the global core-top calibration for alkenone records41. We used the published temperature reconstructions for Antarctic ice-core, pollen, microfossil assemblage andMBT/CBT records and the GISP2 borehole calibration for the Greenland ice cores45. Missing data values near the beginning and end of the ,15% of records not spanning the entire study interval were infilled using the method of regularized expectation maximization46. Stacking. The proxy data were projected onto a 5u 3 5u grid, linearly interpolated to 100-yr resolution and combined as area-weighted averages. We do not otherwise account for spatial biases in the dataset or lack of global coverage. Uncertainty analysis. There are two main sources of uncertainty in the proxy records: age models and temperature calibration. We used a Monte Carlo approach to generate 1,000 realizations of each proxy temperature record after perturbing the records with chronological and temperature errors. These per￾turbed records were then averaged to yield 1,000 realizations of the global and hemispheric temperature stacks. The error bars on the temperature stacks repres￾ent the standard deviations of these 1,000 realizations, which provide an estimate of the propagated uncertainty due to uncertainties in the individual proxy records. A similar approach was applied to the transient model output to develop the modelled temperature stacks and error bars. We developed continuous uncer￾tainty estimates for radiocarbon-based chronologies, taking into account radio￾carbon date errors as well as interpolation uncertainty between dates using a random walk model47. Age-model uncertainties for tuned records, the Dome F and Vostok ice cores, and regional temperature reconstructions for Beringia48were assumed to be 2% (1s). We used the Dome C and EDML ice-core age-model uncertainties13 and GICC05 maximum counting errors49,50 as 2s uncertainties for the NGRIP and GRIP ice cores as suggested in ref. 50. We used the following 1s temperature calibration uncertainties: alkenones, T 5 (Uk0 37 2 0.044 6 0.016)/ (0.033 6 0.001) (ref. 41); Mg/Ca 5 60.02Bexp(60.003AT), where A and B are constants (ref. 51); TEX86, 61.7 uC (ref. 52); ice cores, 610% (ref. 53); pollen, microfossil assemblages and MBT/CBT, 61.5 uC. We did not account for analytical uncertainties in proxy measurements. Chronological errors in the Monte Carlo simulations were temporally autocorrelated but were random in space (but this does not account for systematic errors among the proxy records due to uncertainties in the radiocarbon calibration), whereas temperature errors were assumed to be random in space and time.We note that our study is concerned with temperature anomalies and is thus sensitive to relative but not absolute temperature errors in a proxy record. See Supplementary Information for more details and examples. Age-model uncertainties for the Antarctic Dome C CO2 record related to methane synchronization to Greenland were estimated on the basis of the combined uncertainties associated with Greenland layer counting, Greenland ice-age/gas-age differences and methane tuning to Antarctica. These uncertainties were used to generate 1,000 realizations of the CO2 record, which together with the 1,000 temperature stack realizations yield the 1,000 temperature– concentration lead–lag estimates shown in Fig. 2b. Robustness of results. We evaluated how well the proxy sites represent the globe by subsampling the twentieth-century instrumental temperature record and our transient modelling output of the deglaciation at the 80 proxy sites. Both approaches suggested that the mean of the proxy sites approximates the global mean fairly well. We recalculated the stack after interpolating the records to 500-yr resolution but this did not change the time series or its uncertainty. Differences in areal weighting affect the glacial–interglacial amplitude of the stack but have little impact on its structure. Jackknifing suggests that the stack is not particularly sensitive to the number of records used. A leave-one-out proxy jackknifing approach suggests that the correlation (r 2 5 0.90–0.95) and lead– lag relation (300–600-yr temperature lag) between global temperature and CO2 concentration are not sensitive to proxy type. Statistical infilling of missing data values has negligible impact on the results. Although we here use the IntCal04 radiocarbon calibration, we tested the sensitivity of our results to this choice by recalibrating radiocarbon dates using the IntCal09 calibration. This makes the global stack up to 350 yr older during the Heinrich 1 interval, and shifts the overall phase lag relative to CO2 concentration from 460 6 340 to 350 6 340 yr. We consider the IntCal04 calibration to be more accurate for the reasons discussed in Supplementary Information. Lag correlations suggest that Antarctic temper￾ature led CO2 concentration slightly throughout the deglaciation, whereas global temperature led CO2 concentration at the onset of deglaciation but lagged behind it thereafter. The lead–lag relation between CO2 concentration and the global temperature stack is not significantly changed by detrending the time series to remove the deglacial ramp in each quantity. The significance levels of the correla￾tions between global temperature and CO2 concentration and between Pa/Th and DT were determined by calculating effective sample sizes of these highly auto￾correlated time series (CO2 concentration, 4.3; global temperature, 4.1; Pa/Th, 5.4; DT, 6.2; Atlantic DT, 5.5) using equation 6.26 of ref. 54. See Supplementary Information for more discussion of these tests. Model freshwater forcing. Whereas the forcing from insolation, greenhouse gases and ice sheets during the deglaciation are fairly well constrained, freshwater forcing is comparatively uncertain. Several model freshwater schemes were tested, and the final run was based on the meltwater scenario (Supplementary Fig. 30) that produced North Atlantic climate variability in best agreement with proxy reconstructions. The raw modelled AMOC time series (Fig. 4f, thin lines) were effectively smoothed with a Monte Carlo approach similar to the one used to develop the modelled temperature stacks (Fig. 4e), to facilitate direct comparison of the two. More specifically, the smoothed AMOC time series (Fig. 4f, bold lines) are the means of 1,000 realizations of the raw AMOC time series generated by perturbing them with 300-yr (1s) age-model errors. 45. Cuffey, K. M. & Clow, G. D. Temperature, accumulation, and ice sheet elevation in central Greenland through the last deglacial transition. J. Geophys. Res. 102, 26383–26396 (1997). 46. Schneider, T. Analysis of incomplete climate data: estimation of mean values and covariance matrices and imputation of missing values. J. Clim. 14, 853–871 (2001). 47. Huybers, P. & Wunsch, C. A depth-derived Pleistocene age model: uncertainty estimates, sedimentation variability, and nonlinear climate change. Paleoceanography 19, PA1028 (2004). 48. Viau, A. E., Gajewski, K., Sawada, M. C. & Bunbury, J. Low- and high-frequency climate variability in eastern Beringia during the past 25 000 years. Can. J. Earth Sci. 45, 1435–1453 (2008). 49. Rasmussen, S. O. et al. Synchronization of the NGRIP, GRIP, and GISP2 ice cores across MIS 2 and palaeoclimatic implications. Quat. Sci. Rev. 27, 18–28 (2008). 50. Svensson, A. et al. A 60000 year Greenland stratigraphic ice core chronology. Clim. Past 4, 47–57 (2008). 51. Anand, P., Elderfield, H. & Conte, M. H. Calibration of Mg/Ca thermometry in planktonic foraminifera from a sediment trap time series. Paleoceanography 18, 1050 (2003). 52. Kim, J. H., Schouten, S., Hopmans, E. C., Donner, B. & Damste, J. S. S. Global sediment core-top calibration of the TEX86 paleothermometer in the ocean. Geochim. Cosmochim. Acta 72, 1154–1173 (2008). 53. Jouzel, J. et al. Magnitude of isotope/temperature scaling for interpretation of central Antarctic ice cores. J. Geophys. Res. 108, 4361 (2003). 54. von Storch, H. & Zwiers, F. W. Statistical Analysis in Climate Research 115 (Cambridge Univ. Press, 1999). ARTICLE RESEARCH ©2012 Macmillan Publishers Limited. All rights reserved