High Arctic Holocene temperature record from theAgassiz ice cap and Greenland ice sheet evolutionBenoit S. Lecavaliera,1, David A. Fisherb, Glenn A. Milneb, Bo M. Vintherc, Lev Tarasova, Philippe Huybrechtsd,Denis Lacellee, Brittany Maine, James Zhengf, Jocelyne Bourgeoisg, and Arthur S. Dykeh,i

aDepartment of Physics and Physical Oceanography, Memorial University, St. John’s, Canada, A1B 3X7; bDepartment of Earth and Environmental Sciences,University of Ottawa, Ottawa, Canada, K1N 6N5; cCentre for Ice and Climate, Niels Bohr Institute, University of Copenhagen, Copenhagen, Denmark, 2100;dEarth System Science and Departement Geografie, Vrije Universiteit Brussel, Brussels, Belgium, 1050; eDepartment of Geography, University of Ottawa,Ottawa, Canada, K1N 6N5; fGeological Survey of Canada, Natural Resources Canada, Ottawa, Canada, K1A 0E8; gConsorminex Inc., Gatineau, Canada, J8R3Y3; hDepartment of Earth Sciences, Dalhousie University, Halifax, Canada, B3H 4R2; and iDepartment of Anthropology, McGill University, Montreal,Canada, H3A 2T7

Edited by Jeffrey P. Severinghaus, Scripps Institution of Oceanography, La Jolla, CA, and approved April 18, 2017 (received for review October 2, 2016)

We present a revised and extended high Arctic air temperaturereconstruction from a single proxy that spans the past ∼12,000 y(up to 2009 CE). Our reconstruction from the Agassiz ice cap (Elles-mere Island, Canada) indicates an earlier and warmer Holocenethermal maximum with early Holocene temperatures that are4–5 °C warmer compared with a previous reconstruction, and reg-ularly exceed contemporary values for a period of ∼3,000 y. Ourresults show that air temperatures in this region are now at theirwarmest in the past 6,800–7,800 y, and that the recent rate of tem-perature change is unprecedented over the entire Holocene. Thewarmer early Holocene inferred from the Agassiz ice core leads toan estimated ∼1 km of ice thinning in northwest Greenland duringthe early Holocene using the Camp Century ice core. Ice modelingresults show that this large thinning is consistent with our air tem-perature reconstruction. The modeling results also demonstrate thebroader significance of the enhanced warming, with a retreat of thenorthern ice margin behind its present position in the mid Holoceneand a ∼25% increase in total Greenland ice sheet mass loss (∼1.4 msea-level equivalent) during the last deglaciation, both of which haveimplications for interpreting geodetic measurements of land upliftand gravity changes in northern Greenland.

ice core | temperature reconstruction | Holocene climate | Greenland ice sheet

Instrumented records of temperature and environmental changeextend for a few centuries at most. Although these records

provide evidence of climate warming, the time span covered isrelatively short compared with the centuries to millennia responsetimes of some climate system components (1). In this respect, re-constructions of temperature and environmental changes obtainedfrom climate proxies (e.g., sediment cores, ice cores) play a com-plementary role to the instrumented records by providing a longertemporal context within which to interpret the magnitude and rateof recent changes (2). Furthermore, the relatively large spatial andtemporal variability captured in these reconstructions represents auseful dataset to test models of the climate system (3). Of par-ticular interest are periods during Earth’s history when the climatewas warmer than at present, as these provide information that ispotentially more relevant to changes in the future.In this study, we focus on the reconstruction of past climate

using ice cores from the Agassiz ice cap, located on EllesmereIsland in the Canadian Arctic Archipelago (Fig. 1A). This site isof particular interest as it is located in the high Arctic, andtemperature reconstructions can be compared with those frommore southerly locations to estimate polar amplification of cli-mate in the past (4). Furthermore, it is located proximal to theGreenland ice sheet, and so can be used to better constrain theclimate forcing used to model the past evolution of this ice sheet.In a recent study (5), δ18O measurements in ice from the

Agassiz (81°N) and Renland (70°N) ice caps (Fig. 1A) were usedto estimate temperature records for these locations throughoutthe Holocene. The two time series were remarkably similar,

leading the authors to adopt a spatially homogeneous change inair temperature across the region spanned by these two ice caps.By removing the temperature signal from the δ18O record ofother Greenland ice cores (Fig. 1A), the residual was used toestimate altitude changes of the ice surface through time. Theseso-called thinning curves provide a valuable constraint on modelreconstructions of the Greenland ice sheet (5). A key conclusionof the study was that the current generation of 3D thermo-mechanical ice-sheet models fail to capture the large thinninginferred at sites located closer to the ice margin, particularly innorthwest Greenland (Camp Century drill site; Fig. 1A). How-ever, the veracity of the results in ref. 5 have been brought intoquestion due to the possible influence of the Innuitian ice sheetacross the Canadian Arctic on the altitude correction required toinfer temperature from Agassiz ice during the early Holocene(6). A second issue is that the temperature record estimatedfrom Agassiz ice using two different proxies [ice melt percent (7)and oxygen isotope content (5); see next section] gives incon-sistent results in the early Holocene. Here, we address these is-sues by considering the influence of Innuitian ice sheet thinningon the δ18O temperature reconstruction from Agassiz ice, andapplying the revised reconstruction to force a model of theGreenland ice sheet.

Results and DiscussionReconstructing Holocene Air Temperatures. Previous air temperaturereconstructions inferred from Agassiz ice using observations of the

Significance

Reconstructions of past environmental changes are importantfor placing recent climate change in context and testing climatemodels. Periods of past climates warmer than today provideinsight on how components of the climate system might re-spond in the future. Here, we report on an Arctic climate recordfrom the Agassiz ice cap. Our results show that early Holoceneair temperatures exceed present values by a few degrees Cel-sius, and that industrial era rates of temperature change areunprecedented over the Holocene period (∼12,000 y). We alsodemonstrate that the enhanced warming leads to a large re-sponse of the Greenland ice sheet; providing information onthe ice sheet’s sensitivity to elevated temperatures and thushelping to better estimate its future evolution.

Author contributions: B.S.L., D.A.F., G.A.M., and B.M.V. designed research; B.S.L., D.A.F.,and L.T. performed research; B.S.L., L.T., P.H., J.Z., J.B., and A.S.D. contributed new re-agents/analytic tools; B.S.L., D.A.F., D.L., B.M., and J.Z. analyzed data; and B.S.L., D.A.F.,G.A.M., and D.L. wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.1To whom correspondence should be addressed. Email: b.lecavalier@mun.ca.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1616287114/-/DCSupplemental.

5952–5957 | PNAS | June 6, 2017 | vol. 114 | no. 23 www.pnas.org/cgi/doi/10.1073/pnas.1616287114

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020

melt layers (7) and the δ18O record (5, 8) are inconsistent in theearly Holocene (Fig. 1B). The melt record indicates temperaturespeaking in the early Holocene (∼11 ky) with a steady decline until8 ky (7, 9); whereas, the earlier reconstruction (5) shows air tem-peratures reaching a maximum between 8 and 9 ky. The melt-record reconstruction is a proxy for summer (June, July, August)temperatures, and is derived using a linear transfer function relatingmelt percent to summer air temperatures along with the present-day lapse rate correction at the surface of the ice cap (Methods andFig. S1). This technique can only be used to quantify summer airtemperatures in the range −8 °C (no melt below this temperature)to about −3 °C (100% melt above this temperature). In contrast,the δ18O record is a proxy for mean annual air temperature andspans a much larger temperature range. However, these differencesbetween the two proxies do not reconcile the discrepancies betweenthe earlier δ18O-based temperature reconstructions and the meltreconstruction shown in Fig. 1B.To accurately infer air temperature from ice melt and δ18O

records, the effect of elevation changes must be quantified andremoved (Methods). In the case of the melt record, this correc-tion is applied at the location of the ice core to remove thecontributions from vertical land motion and changes in icethickness (Fig. S2). The elevation correction for the isotope re-cords is more complex, as the correction for vertical land motionand thickness changes of the Innuitian ice sheet is applied at thelocation where the air moisture that precipitates over the ice capreaches a fixed maximum elevation of condensation (5, 6, 10, 11)(Fig. S3). The original δ18O-based reconstruction (5) (Fig. 1B)does not account for this latter effect. When correcting forthickness changes of the Innuitian ice sheet using results from arecent study (12), the δ18O-based temperature reconstruction issignificantly altered (Fig. 1B), with peak temperatures occurringearlier (ca.10.5 ky) and a gradual decline in temperature sincethis time until around 1700 CE. Our reconstruction gives warmerair temperatures during the early Holocene (11.7–8.5 ky) relativeto the original (6), and is now more consistent with that de-termined from the melt record. The offset between maxima ofthe melt and δ18O proxy temperature reconstructions is likely aproduct of the noise levels in the melt record, particularly forearlier times. The A87 melt series has a double peak centeredon 10 and 10.75 ky; whereas, the younger 10-ky peak in theA84 melt series is less pronounced. Therefore, the resulting stackhas a diminished 10-ky peak, which emphasizes the role of noisein a two-series stack. Applying a Gaussian low pass filter (σ =50 y) to the 25-y resolution temperature time series shows a rapidearly Holocene warming with a peak of 6.1 °C at 10 ky (2σ un-certainty 4.3–8.3 °C) followed by a gradual cooling to 1700 CE(Fig. 2A; temperatures are defined relative to the value at 1750 CE).Together, the Agassiz δ18O and melt-layer records point to

substantially higher temperatures during the early Holocenecompared with preindustrial values. Although there are fewquantitative reconstructions of high Arctic air temperatures for

−2−1

0123456789

−12 −10 −8 −6 −4 −2 0Time (kyr before present)

−2−1

0123456789

Tem

pera

ture

ano

mal

y (o C

)

−80˚

−60˚

−60˚

−40˚

−40˚ −20˚ 0˚

60˚60˚60˚

70˚

70˚

80˚

80˚Agassiz

Renland

Camp Century

GRIP/GISP2

A B

Fig. 1. Location map and Agassiz proxy temperature records: (A) Mapshowing the study area with the names and locations of ice core boreholesites mentioned in the text. (B) The 25-y resolution, elevation-correctedAgassiz δ18O temperature reconstruction (dark red) with 2σ uncertainty(light red) and the elevation-corrected Agassiz melt record (green), bothextended to 2009 CE. Ref. 5’s δ18O Agassiz–Renland temperature re-construction is also shown (blue) for comparison. Each record is referencedto its preindustrial temperature value at 1750 CE.

HTM

−2−1

01234567

Tem

pera

ture

ano

mal

y (o C

)

−10

−9

−8

−7

−6

−5

−4

−3

Sum

mer

tem

pera

ture

(o C)

176

178

180

182

184

186

188

Ann

ual a

vera

ge I 8

0o N (W

m−2

)

2

4

6

8

10

12

14

Pol

len

(gra

ins/

L)

8

16

24

32

40

SS

Na

(ppb

)

0

3

6

9

12

15

18

Bow

head

wha

les

(freq

uenc

y %

)

−12 −10 −8 −6 −4 −2 0Time (kyr before present)

A

B

C

D

E

F

Fig. 2. Various records related to high Arctic climate change: (A) Agassiz δ18Otemperature reconstruction (dark red) with its 2σ uncertainty (light red). (B) Agassizmelt record summer (JJA) temperature reconstruction with the potential trend(dashed black) for temperatures below −8 °C and above −3 °C (horizontal graydash; see main text andMethods). (C) Mean annual insolation at 80°N. (D) Agassiztree (purple) and herb (pink) pollen record (15), (E) sea salt sodium in the GISP2(doubled scaling; cyan; ref. 17) and Penny ice cores (green; ref. 16), and (F) numbersof bowhead whale bones from Eastern (gray; 74.3–76.1°N, 83.3–90.5°W; n = 116)and Central (black; 71.5–73.7°N, 89.4–99.0°W; n = 96) Queen Elizabeth Islands (18,19). The gray area denotes the local Holocene thermal maximum, a period whenAgassiz temperatures were regularly above peak contemporary values.

Lecavalier et al. PNAS | June 6, 2017 | vol. 114 | no. 23 | 5953

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020

the early Holocene (13), the occurrence of endostromatolites onDevon Island dated to the early-mid Holocene indicates thattemperatures were 5–8 °C warmer (14), providing evidence of anearly Holocene thermal maximum in this area and supportingour revised δ18O temperature reconstruction.Orbital variations are considered to be the primary driver of a

warm early Holocene in the high Arctic, and this is supported bythe good correlation between peak annual average insolation at80° latitude (Fig. 2C) and temperatures at the Agassiz ice cap. Amaximum in pollen concentrations (Fig. 2D; ref. 15) indicatesthat, as the climate warmed, atmospheric moisture content in-creased alongside a strengthening of meridional heat and mois-ture transport to the Arctic. As a consequence of the warmertemperatures, sea-ice cover was likely at a minimum during theearly Holocene as inferred from the higher concentration of sea-saltsodium found in the Greenland ice sheet and Penny ice cap (Fig.2E; refs. 16, 17). Furthermore, inferences of sea-ice cover in thearchipelago from whale bone remains in raised marine depositssuggest maximum breakup following peak annual insolation andtemperatures at the Agassiz ice cap (Fig. 2F; refs. 18, 19).A shallow ice core was collected in the Agassiz ice cap in 2010,

extending the δ18O time series and thus temperature recon-struction to 2009 (Fig. S4 and Methods). The extended recordindicates that modern air temperatures are ∼4 °C warmer rela-tive to preindustrial values and at their highest in the past∼7000 y (2σ uncertainty 6.8–7.8 ky; Fig. 3A). We calculated ratesof temperature change during the Holocene using a Gaussianfiltered temperature reconstruction and linear regression (Fig.3C). The latter approach was applied to provide a rate that isaccurate for the past 200 y (a contemporary rate is not possibleusing the filtered data due to the edge effect associated with theend of the time series). Rates of temperature change havefluctuated throughout the Holocene, but have generally beenless than ∼1 °C/century. However, the last few centuries haveexperienced the highest rates of temperature change in our

record, exceeding 1.5 °C/century, which is consistent with datafrom nearby weather stations in the Canadian high Arctic wheremean annual air temperatures have increased at a rate of∼0.1 °C/decade since the 1970s (7). Therefore, although airtemperatures were warmer than today in other parts of theHolocene, the rate of climate warming during the industrial erais unprecedented over the past ∼12,000 y.The Agassiz ice cap provides the most northerly paleoclimate

record (∼80°N) of the entire Holocene, and our δ18O temperaturereconstruction provides important information for quantifying thestrength and timing of polar warming. A variety of feedbackmechanisms are responsible for this warming, and there have beensignificant advances in identifying the processes responsible for themodern polar amplification of climate warming (4, 20); however,there remains considerable uncertainty in projections of futureArctic climate change (21) and thus a need for improved observa-tions of past and present climate conditions in the high Arctic. Al-though it is beyond the scope of this study to perform the regional-scale analysis required to rigorously examine polar amplification inthe Holocene, it is of interest to compare the magnitude of earlyHolocene warming from the Agassiz and Renland ice caps. Thecorrected Agassiz early Holocene temperatures warmed by ∼4 °Cmore relative to those inferred at the Renland ice core (easternGreenland; ref. 5) during the 11–9.5 ky interval. Furthermore, peakwarmth at Agassiz and Renland differ by ∼2 ky, with Agassiz ex-periencing an earlier and more distinct Holocene thermal maxi-mum indicating a pronounced reduction in temperaturegradient between these two sites during the early Holocene.

Implications of Agassiz δ18O Temperature Reconstruction for GreenlandIce Sheet Evolution. We apply our δ18O temperature record toconsider its impact on the Holocene evolution of the Greenlandice sheet. As outlined above, elevation histories at Greenland icecore locations (Fig. 1A) were estimated based on the similaritiesbetween the Agassiz and Renland δ18O records (5). Our analysisreveals that these records are different during the early Holo-cene, and so the adoption of a spatially uniform temperaturechange across Greenland (5) is no longer supported. Using ourδ18O temperature record from Agassiz, and assuming that itremains valid regionally, variations in ice elevation at CampCentury were isolated and subsequently corrected for upstreameffects (6), yielding a Camp Century Holocene elevation curve(Fig. 4B; Methods). Our estimated thinning curve at CampCentury displays an elevation reduction during the early Holo-cene that is a factor of 1.6 larger than the original estimate (5)(Fig. S5). As a result, the model–data discrepancies noted pre-viously (5, 6, 22) are enhanced (Fig. S5). The accuracy of theresulting elevation history relies on the accuracy of our as-sumption that the Agassiz climate history reflects climaticchanges occurring at Camp Century. These locations are rela-tively proximal (∼500 km apart), although the possibility of sig-nificant climatic deviations at Camp Century from those onEllesmere Island cannot be disregarded; particularly during theearly Holocene when large changes in ice-surface elevation andsea-ice extent could have produced significant climate variabilityin the Nares Strait region via orographic changes and enhancedatmosphere–ocean interaction as this area became ice free.The majority of Greenland ice sheet models are forced with a

temperature history that is inferred from summit ice cores [e.g.,Greenland Ice Core Project (GRIP), Greenland Ice SheetProject 2 (GISP2)] and then extrapolated across the ice sheetbased on δ18O–T relationships and lapse rates (e.g., ref. 22). Asshown in Fig. 4A, this extrapolated curve greatly underestimatesthe amplitude of early Holocene warmth compared with thatinferred here from the Agassiz ice δ18O record. To address thisissue, we performed a model sensitivity analysis using a revisedtemperature history, based on the Agassiz ice δ18O temperaturetime series (Fig. 4A; Methods), to force the northern sector of a re-cent Greenland ice sheet reconstruction (ref. 22; referred to here-after as Huy3). The revised climate forcing yields rapid thinningacross NorthGreenland, particularly northwest Greenland as presented

−2

−1

0

1

2

3

4

5

6

7

8

9

Tem

pera

ture

ano

mal

y (o C

)

−2.5−2.0−1.5−1.0−0.5

0.00.51.01.52.02.5

Rat

e of

tem

pera

ture

cha

nge

o ( C/c

entu

ry)

−10 −8 −6 −4 −2 0Time (kyr before present)

−2

0

2

4

Tem

pera

ture

(o C)

−0.6 −0.4 −0.2 0.0Time (kyr)

A B

C

Fig. 3. Low-pass-filtered Agassiz δ18O temperature reconstruction: (A) Agassizδ18O temperature reconstruction (dark red) with its 2σ uncertainty (light red)with the Gaussian filtered (σ = 50 y) reconstruction (black). (B) The Agassizδ18O temperature reconstruction over the past 1,000 y. The gray lines denotethe linear regression results for three periods: the cooling leading into thepreindustrial period, the preindustrial era, and industrial era. (C) Rate oftemperature change (black) based on the Gaussian filtered Agassiz tem-perature reconstruction. The gray circles represent the rates of changeobtained from the linear regression segments shown in B. The vertical grayband denotes the local Holocene thermal maximum, a period when Agassiztemperatures were regularly above peak contemporary values.

5954 | www.pnas.org/cgi/doi/10.1073/pnas.1616287114 Lecavalier et al.

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020

in Fig. 4B. As a consequence of this enhanced thinning, it was nec-essary to produce a considerably thicker ice sheet at the last glacialmaximum to match present-day ice extent and thickness in this re-gion. As indicated in Fig. 4B, it was possible to produce a good fit tothe Camp Century thinning curve using our revised temperatureforcing, suggesting that the failure of previous models to capture thissignal (ref. 5; Fig. S5; Methods) reflects inaccuracy in the adoptedclimate forcing. As discussed in Methods, our simulations do notcapture the full buttressing effect of the Innuitian ice sheet on theGreenland ice sheet across Nares Strait. Deglaciation of this areaaround 10 ky before present would also have contributed to the rapidice thinning in the region during the early Holocene (23).The larger deglaciation predicted for North Greenland will

influence predictions of relative sea level in this region, so weinput our revision of the Huy3 ice-sheet reconstruction to a glacialisostatic adjustment (GIA) model (Methods) and performed a data–model comparison to test whether the relative sea-level (RSL)observations support the large thinning suggested by the CampCentury ice core record. Comparison with observations in Northand northeast Greenland suggest that the revised ice model is morecompatible with the majority of the RSL reconstructions comparedwith the Huy3 reconstruction (Figs. S6 and S7; Methods), althoughwe note that the quality of RSL data is low in North Greenland. Innorthwest Greenland, nearest to Camp Century, there are somedata–model discrepancies with both the original and revisedHuy3 reconstruction. At Qeqertat, the models fit the RSL ob-servations only if one takes into consideration uncertainties inthe North American ice complex (see figure S6 from ref. 22). AtSaunders and Thule, the Huy3 model underpredicts the RSLobservations with a mistimed fall in sea level, and the variantHuy3 model over predicts the observations significantly.The accuracy of our revised ice model can also be tested against

Global Positioning System (GPS) estimates of vertical land motion.In a recent study (24), the component of this motion associated

with past ice-sheet changes was isolated by estimating and removingthe signal caused by ice-sheet changes and the corresponding elasticearth deformation during the GPS monitoring period. Table S1provides a comparison of observed and modeled uplift rates thatincludes values for both Huy3 and the revised model. At sites in thenorthwest, the revised model produces an improved fit to the ob-served rates. However, we note that the model rates underpredictthose observed. In particular, we note that the uplift rates at the twoGPS sites nearest to Camp Century are considerably larger thanthose predicted by the two models, suggesting that a greateramount of regional thinning is required, in contrast to that sug-gested by the RSL reconstructions at Saunders and Thule.The above-noted data–model discrepancies with respect to the

RSL and GPS observations have different possible interpretations.For example, the northwest region could have a different earthviscosity structure to that of the Greenland-wide optimum used here(22; determined using the Huy3 ice history). Alternatively, themagnitude and/or timing of ice unloading could be incorrect, in-dicating that extrapolating the Agassiz temperature curve to CampCentury is not applicable; or other processes such as changes in themoisture pathway are influencing the δ18O at Camp Century, thuscomplicating the inferences of elevation changes at this site. De-termining which of these interpretations is correct, and whether allthree data types (Camp Century thinning curve, RSL curves, GPSrates) can be reconciled with a single model parameter set, will re-quire improved observational constraints and a more detailed modelsensitivity analysis that explores parameter uncertainties more fully.Even though we have focused on the north of Greenland in

this analysis, the differences between the original Huy3 modeland the revised version are large enough to be evident in theirrespective volume time series (Fig. S8). Compared with Huy3,the revised North Greenland reconstruction delivers an addi-tional 1.38 m of ice-equivalent sea level to the global oceansduring the most recent deglaciation. Although this is a relativelysmall amount compared with the global ice volume loss (∼130 m), itis 27% of the loss in the Huy3 Greenland model. The larger tem-perature forcing in the early Holocene also results in a modeledretreat of the ice margin interior of its current position by 20–80 km(as early as 8 ky before present in some places) followed by aregrowth to present. The regrowth in North Greenland repre-sents a net drop of 0.18 m of ice-equivalent sea level, similar tothat estimated for the southwest part of the ice sheet (22). Ourrevised model shows that rates of surface thinning reached valuesof 36.7 m/century at Camp Century and the Greenland ice sheetexperienced a peak centennial rate of Holocene mass loss of1,075 gigaton per year (Gt/y) during the period when recon-structed temperatures were greater than those at present (Fig.S9). This rate of model mass loss reflects the “memory” of the icesheet to past air temperature changes, notably the large increase atthe Pleistocene–Holocene transition shortly after 12 ky, as well asthe contemporary response to the peak temperatures during theearly Holocene. Before 10 ky, marine retreat of the ice modelfrom rising sea levels (22) is a significant contributor to the rateof mass loss. After this time, the ice margin was largely landbased and so the mass loss rates can be compared more directly tothose estimated for the ice sheet at present using geodetic methods;for example 142 ± 49 Gt/y for the period 1992–2011 (25).

Concluding RemarksAs demonstrated in the previous section, the δ18O air tempera-ture reconstruction resulting from this analysis has implicationsfor model-derived regional ice-sheet reconstructions. As the icehistory is a necessary input to arrive at the corrections applied tothe Agassiz ice-core data, the ideal approach would involve it-eration to ensure that the resulting temperature reconstructionand regional ice model reconstruction are consistent. Such anapproach was not applied here because it would require a sig-nificantly more complex analysis to model the Greenland andInnuitian ice sheets simultaneously and capture the interactionsbetween them. Of the two corrections applied to the δ18O tem-perature reconstruction—one associated with height changes of

−6

−4

−2

0

2

4

6

8

Tem

pera

ture

ano

mal

y (o C

)−200

0

200

400

600

800

1000

1200

Cam

p C

entu

ry th

inni

ng c

urve

(m)

−12 −10 −8 −6 −4 −2 0Time (kyr before present)

A

B

Fig. 4. Temperature and thinning curves for northwest Greenland: (A) Agassizδ18O temperature reconstruction (dark red) with 2σ uncertainty (light red).The temperature time series at Camp Century inferred from the GRIP ice coreusing a δ18O–temperature relationship and lapse rate correction (dashedblack) and the revised temperature time series based on the Agassiz re-construction (solid black; Methods). (B) Camp Century thinning curve (green)and 2σ uncertainty (light green) compared with model output: Huy3 (dashedblack) and our variant of this model reconstruction (solid black).

Lecavalier et al. PNAS | June 6, 2017 | vol. 114 | no. 23 | 5955

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020

the Innuitian ice sheet and the other with isostatic land motion—the former is potentially the more important in terms of being asignificant feedback on the estimated temperature reconstruc-tion because it is not directly constrained via the results of thisstudy. In contrast, the isostatic response was calibrated to RSLdata from Ellesmere Island and so even if the regional iceloading history is significantly altered when using the revisedtemperature reconstruction, the earth viscosity parameters would bevaried to maintain an isostatic response that best matches the RSLdata. That is, the optimal viscosity model will be different but theisostatic response (and thus land uplift correction) will be similar.Regarding possible changes to the Innuitian ice sheet via revisingthe temperature reconstruction, the impact of this on the resultspresented here is difficult to determine without applying an iterativeapproach. However, we note that a high variance subset of Innuitianice reconstructions was adopted to partially account for not applyingan iterative approach.The consistency between the proxy records (δ18O and melt

records from Agassiz and those shown in Fig. 2), and the good fitbetween the modeled ice thinning and that inferred from theCamp Century ice core, indicate that the feedbacks noted aboveare relatively minor and our primary conclusions are accurate tofirst order. However, this is an aspect of the current analysis thatcould be improved upon in future studies.

MethodsInferring Temperature Records from Ice Core Measurements. The Agassiz δ18Oand melt records have been influenced by a thinning ice sheet and isostaticrebound. Therefore, an altitude correction is required to obtain temperaturereconstructions that are elevation independent. Although the δ18O is sensitiveto elevation changes along the southeast coast of Ellesmere Island, the meltrecord is affected by elevation changes at the borehole site and so separatealtitude corrections are required for each record (Figs. S2 and S3). Previouswork has illustrated that local changes in the altitude and thickness of theAgassiz ice cap do not affect the δ18O composition of the ice (9, 11). The at-mospheric moisture that precipitates onto the Agassiz ice cap primarily passesthrough (and partly originates in) Baffin Bay. It subsequently encounters thesoutheastern shores of Ellesmere Island where it is elevated, and given thereare no further inland features capable of forcing the air masses significantlyhigher, the δ18O composition in the ice is predominantly sensitive to altitudechanges along this shoreline (5, 6, 10). Thus, the elevation correction is notdetermined at the drill site, but rather at the location where the air massinitially encounters a major topographical feature along the eastern coastlineof Ellesmere Island. Furthermore, the position where the correction is esti-mated changes through the Holocene due to topographical changes as theInnuitian ice sheet melted (6). The two processes that dominated altitudechanges of southeastern Ellesmere were: (i) GIA and (ii) thinning of Innuitianice (Figs. S2 and S3). Similar to previous studies (i.e., ref. 6), we assume neg-ligible change in seasonal biases in the δ18O relationships, with snowfall acrossthe Agassiz ice cap occurring throughout the year, although with increasedprecipitation starting in the spring and decreasing in the fall at present (8).

The Agassiz altitude correction was determined using the GIA model pa-rameters from ref. 6 and Innuitian ice sheet thinning from the analysis of ref. 12.Regarding the latter, a high variance subensemble of the Bayesian calibrationspecifically weighted to the Innuitian ice sheet was adopted; the best scoringmodel was used to determine the altitude corrections and the variance from thesubensemble represents the uncertainty. Over the early Holocene, Innuitian icethinned by ∼400 m along southeast Ellesmere and so dominates the altitudecorrection. In contrast, the mid to late Holocene correction is dominated byGIA of ∼100 m (Figs. S2 and S3). Note that the δ18O measurements werealso corrected to account for changes in δ18O content of sea water (26).

The altitude correction for the Agassiz δ18O record was applied throughthe δ18O altitude relation derived from shallow ice cores [−0.62 ± 0.03 ‰ per100 m (27, 28)]. A forward modeling method was used to calibrate a tem-perature–δ18O relationship with the borehole temperature profiles fromGRIP, North GRIP (NGRIP), DYE-3, and Camp Century by numerically solvingthe differential equation for energy conservation in moving ice, yielding adegrees Celsius/δ18O slope of 2.1 ± 0.2 (5, 6, 29). Upon deriving a correctedδ18O record, the temperature–δ18O slope conversion produces the annual airtemperature reconstruction for a fixed elevation (present-day borehole el-evation), as presented in Fig. S3D.

The Agassiz melt record was related to summer temperatures using atransfer function (7). Summer temperatures and melt layers across a number

of shallow Arctic ice cores exhibit a robust linear relationship as shown inFig. S1 (7, 30). This linear transfer function emphasizes that summer (June–August; JJA) temperatures below −8 °C yield no melt fraction in the ice,therefore, melt percent values of zero signify summer temperatures lessthan or equal to −8 °C. Similarly, melt percent values of 100 signify summertemperatures above or equal to −3 °C. The elevation changes for the meltrecord are converted to summer (JJA) temperature using the present-daylapse rate for Ellesmere Island of −0.43 °C per 100 m (7, 30).

The summer temperature to melt percent transfer function is based onpresent-day observations. This transfer function likely correlates with in-solation among other time-varying environmental parameters through theHolocene. However, other processes and feedbacks in the climate systemrender the temporal calibration of the transfer function highly nontrivial,especially because there is a lack of constraints on key environmental pa-rameters in the high Arctic during this period. Therefore, we assume that thepresent-day transfer function and its uncertainties adequately capture thetemperature to melt percent relationship to first order during the Holocene.

The rate of temperature change from the Agassiz annual temperaturereconstruction was determined by discretized differentiation of the Gaussianfiltered record presented in Fig. 3A. This provided multicentennial-scalevariations in the rate of temperature change (Fig. 3). The centennialtrends highlight fluctuations in the rate of temperature change, which arenot strongly influenced by noise within the high-resolution δ18O record.However, the low-pass Gaussian filter truncates the record to 100 y beforepresent where edge effects are negligible. To supplement the truncationnear present, linear regression was conducted on the raw temperature re-construction on 200-y intervals starting at present and going back to 600 ybefore present. This yielded a rate of −0.08, 0.37, and 1.97 °C/century for theperiods before present of 600–400 y, 400–200 y, 200 y to present, re-spectively (Fig. 3).

The error analysis for the Agassiz δ18O temperature reconstruction ac-counts for the uncertainties arising from: thinning of the Innuitian ice sheet(12); moisture pathway to the Agassiz ice cap (6); δ18O altitude relation (27);and the temperature–δ18O relationship (5, 6, 29). The uncertainties in thetemperature reconstruction are nonparametric and for this reason conser-vative error estimates are presented in this study. When dealing withGaussian uncertainties (e.g., δ18O altitude relation based on linear regression),traditional error analysis methods were applied. However, when dealing withnonparametric uncertainties, rather than conducting a Monte Carlo error anal-ysis to estimate confidence intervals, the upper and lower bound error estimateswere used to evaluate the uncertainties. The choice to present conservativeuncertainty estimates in these cases was made to partly compensate for a likelyunderestimation of nontrivial uncertainties such as the temporal evolution of theδ18O–altitude and temperature–δ18O relationships.

The Agassiz 2009 Extension Core Series. Starting in the late 1970s, ice-coredrilling campaigns have been undertaken at the Agassiz Ice Cap. The stud-ies resulted in Holocene records of stable water isotopes, melt layers, solidconductivity, and pollen (8, 9). The Agassiz A84 and A87 isotope records arebased on cores (obtained in those years), which were 100 m apart. In 2010, a16-m core (A09; 80°49’N, 72°53.74’W) was obtained, located between thoseearlier cores, to extend the melt and isotope records to 2009 CE. Eighthundred water samples were analyzed at University of Ottawa for 18O usingan LGR liquid water isotope analyzer. The OA-ICOS liquid water analyzerwas coupled to a CTC LC-PAL autosampler for simultaneous 18O/16O ratiomeasurements of H2O. Following analysis, all measured water samples wereverified for spectral contaminants in the samples using the LGR spectral in-terference contamination identifier software. Analyses were calibrated andnormalized to internal laboratory water standards that were previouslycalibrated relative to Vienna Standard Mean Ocean Water (VSMOW) using aconventional isotope ratio mass spectrometer. Consequently, the results arepresented using the δ-notation (δ18O), which represents the parts perthousand differences for 18O/16O in a sample with respect to VSMOW. An-alytical reproducibility for δ18O is ± 0.3‰.

The age–depth timescale for the A09 ice core is described in ref. 7. TheA09 core includes a period of overlap (∼30 y) with the deep core melt (ref. 7,their figure 2) and δ18O records from Agassiz. The δ18O overlap shows thatthe A09 core is properly aligned with the old deep cores and that there is asmooth transition ensuring a homogeneous extended time series. Fig. S4shows the high-resolution A84, A87, and A09 δ(18O) series.

Greenland Ice-Model Sensitivity Analysis. The North Greenland sensitivityanalysis was conducted using a glaciological model and GIAmodel of relativesea-level change (22). This previous study produced the Huy3 Greenlandreconstruction, which was achieved by simultaneously tuning/calibrating a

5956 | www.pnas.org/cgi/doi/10.1073/pnas.1616287114 Lecavalier et al.

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020

3D thermomechanical ice-sheet model in series with a GIA model. Please seeref. 22 for further details on the models applied and methodology followed.We adopt the Huy3 reconstruction here and revise it by tuning the NorthGreenland climate forcing to reflect our Agassiz temperature reconstruction.

Given that the ice model does not include the adjacent Innuitian ice sheetnor the buttressing effect of potential ice shelves in the Baffin Sea (both ofwhich would result in a thicker last glacial maximum ice sheet in thenorthwest sector), an alternative method was required to enhance past icethickness in this region, as suggested by the Camp Century thinning curve. Tothis end, we decided to vary the input precipitation field because it is poorlyconstrained and the parameterization schemes previously adopted lackedthe necessary degrees of freedom to account for these uncertainties. Thetemperature forcing across North Greenland was parameterized to coincidewith the Agassiz temperature reconstruction. The parameterization of theclimate forcing was conducted in a similar manner to that of the Holocenethermal maximum in ref. 22. A tuning of the climate forcing was achieved byadding linear and/or quadratic empirical equation estimates of the en-hanced temperatures at Agassiz compared with the GRIP temperature de-rived climate fields (Fig. 4A shows an example for the Camp Centurylocation). This revision to the GRIP inferred temperatures was extended priorthe start of the Agassiz record given the large uncertainties in the climateforcing and thus maintains the necessary degrees of freedom.

Taking the variant model reconstruction, we computed RSL histories fordata sites in North Greenland (Fig. S6) to further test the accuracy of theHuy3 model revision. Even though the quality of RSL data from this region isrelatively poor (22), these observations are a primary constraint on modelreconstructions of the ice history. RSL model predictions for the Huy3 modelreconstruction and its variant produced here are compared with observationsin Fig. S7. At a number of RSL sites—Holm Land, Ingeborg Halvo, Herluf-sholm, Ole Chiewitz—both Huy3 and its variant reconstruction remainwithin error of the observations when considering uncertainties in the earthviscosity structure (see figures 12 and S6 from ref. 22). It should be notedthat there is some ambiguity when interpreting the marine limits (i.e.,maximum RSL at a given site) because this quantity represents the highestpoint reached by sea level when it is ice free, and so represents a dual proxy

for ice extent and RSL. Given the sensitivity of this proxy to ice extent, thetime at which maximum RSL occurred can be difficult to constrain and sointroduces ambiguity into the model–data comparison. At the Kronprinssite, the variant reconstruction of Huy3 achieves a perfect fit to the obser-vational constraints by remaining below the upper-limiting and above thelower-limiting RSL dates, including the highest lower-limiting RSL date andmarine limit, which the original Huy3 model does not reach. In contrast atJorgen, no model prediction achieves a fit to the RSL observations. Themodel curves reach the uppermost lower-limiting dates but do not capturethe rapid sea-level fall to reach the oldest upper-limiting date. However, therapid sea-level fall observed at Jorgen is not observed at other neighboringsites, which might be indicative of local effects that are not captured inregional GIA models. At Constable, the variant reconstruction shows amarginally improved fit over the Huy3 chronology; it reaches the marinelimit, remains above the uppermost lower-limiting dates, passes through thesea-level index point, and passes a few meters above the late Holoceneupper-limiting RSL dates. However, there is some conflicting data during theearly to mid Holocene where upper- and lower-limiting dates and sea-levelindex points suggest different sea level histories. The RSL predictions gen-erated by the variant reconstruction achieve a significantly improved fit at anumber of sites—JPKoch, Nyboe, Hall East, Hall West—compared withHuy3 by remaining within the bounds of the limiting dates. The RSL data atsites Lafayette and Humboldt poorly constrain sea level and therefore donot strongly discriminate between the two model reconstructions.

Weapplied the same isostaticmodel to compute vertical landmotion at sites innorthwest Greenland where GPS receivers have been installed. Because theobserved rates of land motion have been corrected for elastic deformation as-sociated with ice loading over the GPS monitoring period (24), the elastic con-tribution was removed from the model output. Comparing observed andmodeled rates in Table S1 indicates that the revised model is an improvement atall sites except those in the northeast (JGBL, KMJP, LEFN, NORD).

ACKNOWLEDGMENTS. This work was funded by the Natural Sciences andEngineering Research Council of Canada. This paper is a contribution to thePAGES/INQUA funded PALSEA2 working group.

1. Stocker TF, et al. (2013) IPCC, 2013: Climate Change 2013: The Physical Science Basis.Contribution of Working Group I to the Fifth Assessment Report of the IntergovernmentalPanel on Climate Change (Cambridge Univ Press, Cambridge).

2. Marcott SA, Shakun JD, Clark PU, Mix AC (2013) A reconstruction of regional andglobal temperature for the past 11,300 years. Science 339:1198–1201.

3. Masson-Delmotte V, et al. (2013) Information from paleoclimate archives. ClimateChange 2013: The Physical Science Basis. Contribution of Working Group I to theFifth Assessment Report of the Intergovernmental Panel on Climate Change (Cam-bridge Univ Press, Cambridge), pp 383–464.

4. Bekryaev RV, Polyakov IV, Alexeev VA (2010) Role of polar amplification in long-termsurface air temperature variations and modern arctic warming. J Clim 23:3888–3906.

5. Vinther BM, et al. (2009) Holocene thinning of the Greenland ice sheet. Nature 461:385–388.

6. Lecavalier BS, et al. (2013) Revised estimates of Greenland ice sheet thinning historiesbased on ice-core records. Quat Sci Rev 63:73–82.

7. Fisher D, et al. (2012) Recent melt rates of Canadian arctic ice caps are the highest infour millennia. Global Planet Change 84–85:3–7.

8. Fisher DA, Koerner RM, Reeh N (1995) Holocene climatic records from Agassiz Ice Cap,Ellesmere Island, NWT, Canada. Holocene 5:19–24.

9. Koerner RM, Fisher DA (1990) A record of Holocene summer climate from a Canadianhigh-Arctic ice core. Nature 343:630–631.

10. Fisher DA (1990) A zonally-averaged stable-isotope model coupled to a regionalvariable-elevation stable-isotope model. Ann Glaciol 14:65–71.

11. Fisher DA (1992) Stable isotope simulations using a regional stable isotope modelcoupled to a zonally averaged global model. Cold Reg Sci Technol 21:61–77.

12. Tarasov L, Dyke AS, Neal RM, Peltier WR (2012) A data-calibrated distribution ofdeglacial chronologies for the North American ice complex from glaciological mod-eling. Earth Planet Sci Lett 315:30–40.

13. Renssen H, Seppä H, Crosta X, Goosse H, Roche DM (2012) Global characterization ofthe Holocene Thermal Maximum. Quat Sci Rev 48:7–19.

14. Lacelle D, Pellerin A, Clark ID, Lauriol B, Fortin D (2009) (Micro)morphological,inorganic-organic isotope geochemisty and microbial populations in endo-stromatolites (cf. fissure calcretes), Haughton impact structure, Devon Island, Canada:The influence of geochemical pathways on the preservation of isotope bioma. EarthPlanet Sci Lett 281:202–214.

15. Bourgeois JC, Koerner RM, Gajewski K, Fisher DA (2000) A Holocene ice-core pollenrecord from Ellesmere Island, Nunavut, Canada. Quat Res 54:275–283.

16. Fisher DA, et al. (1998) Penny ice cap cores, Baffin Island, Canada, and the Wisconsinanfoxe dome connection: Two states of Hudson Bay ice cover. Science 279:692–695.

17. Mayewski Pa, et al. (1997) Major features and forcing of high-latitude northernhemisphere atmospheric circulation using a 110,000-year-long glaciochemical series.J Geophys Res 102:26345–26366.

18. Dyke AS, Hooper J, Savelle JM (1996) A history of sea ice in the Canadian Arctic ar-chipelago based on postglacial remains of the bowhead whale (Balaena mysticetus).Arctic 49:235–255.

19. Dyke AS, Hooper J, Harington CR, Savelle JM (1999) The late Wisconsinan and Ho-locene record of Walrus (Odobenus rosmarus) from North America: A review withnew data from Arctic and Atlantic Canada. Arctic 52:160–181.

20. Pithan F, Mauritsen T (2014) Arctic amplification dominated by temperature feed-backs in contemporary climate models. Nat Geosci 7:2–5.

21. Collins M, et al. (2013) Long-term climate change: Projections, commitments and ir-reversibility. Climate Change 2013: The Physical Science Basis. Contribution ofWorking Group I to the Fifth Assessment Report of the Intergovernmental Panelon Climate Change (Cambridge Univ Press, Cambridge), pp 1029–1136.

22. Lecavalier BS, et al. (2014) A model of Greenland ice sheet deglaciation constrainedby observations of relative sea level and ice extent. Quat Sci Rev 102:54–84.

23. MacGregor JA, et al. (2016) Holocene deceleration of the Greenland Ice Sheet. Science351:590–593.

24. Khan SA, et al. (2016) Geodetic measurements reveal similarities between post-LastGlacial Maximum and present-day mass loss from the Greenland ice sheet. Sci Adv2:e1600931.

25. Shepherd A, et al. (2012) A reconciled estimate of ice-sheet mass balance. Science 338:1183–1189.

26. Waelbroeck C, et al. (2002) Sea-level and deep water temperature changes derivedfrom benthic foraminifera isotopic records. Quat Sci Rev 21:295–305.

27. Dahl-Jensen D, et al.; NEEM community members (2013) Eemian interglacial re-constructed from a Greenland folded ice core. Nature 493:489–494.

28. Johnsen SJ, Dansgaard W, White JWC (1989) The origin of Arctic precipitation underpresent and glacial conditions. Tellus B Chem Phys Meterol 41:452–468.

29. Dahl-Jensen D, et al. (1998) Past temperatures directly from the Greenland ice sheet.Science 282:268–271.

30. Marshall SJ, Sharp MJ, Burgess DO, Anslow FS (2007) Near-surface-temperature lapserates on the Prince of Wales Icefield, Ellesmere Island, Canada: Implications for re-gional downscaling of temperature. Int J Climatol 27:385–398.

31. Tarasov L, Peltier WR (2002) Greenland glacial history and local geodynamic conse-quences. Geophysical Journal International, 150:198–229.

32. Simpson MJ, Milne GA, Huybrechts P, Long AJ (2009) Calibrating a glaciological modelof the Greenland ice sheet from the Last Glacial Maximum to present-day using fieldobservations of relative sea level and ice extent. Quat Sci Rev 28:1631–1657.

Lecavalier et al. PNAS | June 6, 2017 | vol. 114 | no. 23 | 5957

ENVIRONMEN

TAL

SCIENCE

S

Dow

nloa

ded

by g

uest

on

Sep

tem

ber

27, 2

020