A leading hypothesis to explain abrupt climate change during the last glacial cycle calls on fluctuations in the margin of the North American Laurentide Ice Sheet (LIS), which may have routed fresh water between the Gulf of Mexico (GOM) and the North Atlantic, affecting North Atlantic Deep Water variability and regional climate. Paired measurements of δ18O and Mg/Ca of foraminiferal calcite from GOM sediments reveal five episodes of LIS meltwater input from 28 to 45 thousand years ago (ka) that do not match the millennial-scale Dansgaard-Oeschger warmings recorded in Greenland ice. We suggest that summer melting of the LIS may occur during Antarctic warming and likely contributed to sea level variability during marine isotope stage 3.
 Abrupt climate changes during the last glaciation have been linked to variations in Atlantic thermohaline circulation. Numerical models demonstrate that an increased flux of fresh water to sites of deepwater formation decreases the strength of North Atlantic Deep Water (NADW), thereby reducing meridional heat transport and causing cooling/warming in the northern/southern high latitudes [Ganopolski and Rahmstorf, 2001; Knutti et al., 2004]. This bipolar seesaw [Broecker, 1998] has been invoked to explain the antiphased relationship between climate changes in Antarctica and Greenland, where warmings in Antarctica precede those in Greenland by several thousand years [Blunier and Brook, 2001]. Additionally, the climate signature in Antarctica shows gradual temperature changes, while Greenland temperature is characterized by higher-frequency changes, including abrupt warmings that occur in decades, followed by slow coolings (Dansgaard-Oeschger (D/O) cycles).
 The North American Laurentide Ice Sheet (LIS) may have served as a source of fresh water to the North Atlantic during the last deglaciation, when ice sheet retreat led to the diversion of fresh water (meltwater and precipitation) from the Mississippi River drainage to the Hudson and St. Lawrence rivers [Rooth, 1982; Broecker et al., 1988; 1989; Shackleton, 1989; Flower and Kennett, 1990; Clark et al., 2001; Flower et al., 2004]. Meltwater routing has been suggested as a potential control of high-frequency climate variability during intervals of intermediate ice volume, such as during marine isotope stage 3 (MIS 3) [Clark et al., 2001]. However, evidence is needed to assess potential switches in freshwater routing during the millennial-scale D/O cycles, which are characterized by 5°–10°C oscillations in Greenland air temperature [Dansgaard et al., 1993]. Here we test whether D/O warmings correspond to freshwater routing to the Gulf of Mexico (GOM) by reconstructing the δ18O composition of seawater (δ18Osw) using paired measurements of δ18Ocalcite (δ18Oc) and Mg/Ca of GOM foraminifera. Orca Basin (26°56.77′N, 91°20.74′W, Figure 1) in the northern GOM is ideally located to study freshwater input, including LIS meltwater, from the North American continent because of its proximal location to the mouth of the Mississippi River.
2. The δ18O and Mg/Ca Analyses
 Core MD02-2551 was recovered from Orca Basin in July 2002 by the R/V Marion Dufresne as part of the International Marine Past Global Changes Study (IMAGES) program. The core was sampled at 2-cm intervals from 21 to 30 m. All samples were freeze dried prior to wet sieving, and then washed over a 63-μm mesh using deionized water. The ∼60–70 planktonic foraminifera G. ruber (pink variety) were picked from the 250- to 355-μm-size fraction for isotopic and elemental analyses. The foraminifera were sonicated in methanol for 5 s to remove clays, and then weighed to assess down core dissolution effects. Mean G. ruber weights are similar throughout the interval and are comparable to surface-sediment samples [LoDico et al., 2002]. The shells were gently crushed open between two glass plates and carefully homogenized using a razor blade. A ∼50-μg aliquot was removed for stable isotopic analysis, which was performed at the College of Marine Science, University of South Florida, using a ThermoFinnigan Delta Plus XL dual-inlet mass spectrometer with an attached Kiel III carbonate preparation device. The isotopic data (Figure 2) are reported on the VPDB scale calibrated with NBS-19. Standard deviation for the δ18Oc measurements is ±0.04‰, based on measurements of the standard NBS-19 analyzed with MD02-2551 foraminifer samples (n = 105).
 The remaining tests, weighing ∼700 μg, were split into two aliquots that were cleaned separately for Mg/Ca analysis [Barker et al., 2003]. This method involves an initial sonication to remove fine clays, oxidation of organic matter with a buffered peroxide solution, and a dilute acid leach that eliminates any adsorbed contaminants. Samples were dissolved in weak HNO3 to yield calcium concentrations of ∼20 ppm to minimize calcium concentration effects. The Mg/Ca ratios (Figure 2) were analyzed on a Perkin Elmer Optima 4300 dual-view inductively coupled plasma–optical emission spectrometer (ICP-OES). A standard instrument-drift correction technique was routinely used. The analytical precision for Mg/Ca determinations used in this study is <0.6% root-mean standard deviation (1σ), based on an ICP-MS calibrated standard solution. The pooled standard deviation of 70% replicate Mg/Ca analyses is ±2.5% (df = 318), which is equivalent to ∼0.3°C. The δ18O and Mg/Ca data can be accessed at the NOAA/NGDC World Data Center at http://www.ngdc.noaa.gov/paleo/paleo.html.
3. Age Model
 The age model developed for our record (Figure 2) is based on 18 AMS 14C dates (Table 1) determined from monospecific samples (4–10 mg) of pink G. ruber, which were run at the Center for Accelerator Mass Spectrometry, Lawrence Livermore National Laboratory. The 14C ages were corrected for a reservoir age of 400 years and converted to the GISP2 timescale (an approximation of calendar years) using a high-resolution radiocarbon calibration developed on sediment cores from the Cariaco Basin [Hughen et al., 2004]. Inferred minimal changes in upwelling at the Orca Basin site indicate uncertainty in the reservoir correction is much better than 100 years. Age was also constrained by the Laschamp geomagnetic event [Laj et al., 2000], which is recorded as a ∼50 cm minimum in inclination at a depth of ∼27.5 m (C. Kissel et al., personal communication, 2004). A peak in 10Be, which coincides with the Laschamp event in sediment cores from the North Atlantic [Robinson et al., 1995], straddles the δ18O peak of interstadial 10 in the Greenland ice core record [Yiou et al., 1997]. The Laschamp event was therefore assigned a calendar age of 40.9 ka based on the age of the δ 18O peak of interstadial 10 on the Greenland GISP2 timescale [Meese et al., 1997] (Figure 2).
CAMS is Center for Accelerator Mass Spectrometry, Lawrence Livermore National Laboratory; AMS is accelerator mass spectrometry.
Samples not included in the age model because of stratigraphic inconsistencies. The 14C ages at depths of 28.06, 28.46, and 29.88 m are younger than higher depths in the core. We choose not to use the 14C age at 29.20 because it would require very large sedimentation rate changes from 30 cm/kyr to 200 cm/kyr. Although this is possible, we choose, instead, to linearly extrapolate beyond the Laschamp event, and we are conservative with interpretations in our data prior to 41 ka.
 Depth in centimeters was converted to age by applying a weighted curve fit with a 40% smoothing factor and linearly extrapolating beyond the Laschamp event. This function fits a curve to the calibrated 14C age control points, using the locally weighted least squares error method. Because of the uncertainty associated with radiocarbon dates of increasing age, including 14C age plateaus at ∼24 and ∼28 14C ka [Hughen et al., 2004], the weighted smooth fit provides a conservative estimate of depth versus age. Sedimentation rates range from 25 cm/kyr to 325 cm/kyr.
 Total error (1σ) on the age model ranges from 140 calendar years at ∼26 ka to a maximum of 700 calendar years at ∼40 ka. Error was determined by compounding the error on the 14C measurements from this study (Table 1), the error on the 14C measurements from the Cariaco record and the error from the GISP2/Cariaco calibration reported by Hughen et al. . Errors in 14C were converted to calendar years using the Cariaco calibration. Calculating the error prior to 40 ka is difficult because of the uncertainty in the Cariaco calibration. Errors on the layer counting from the GISP2 record were not included in the total error analysis because we do not make conclusions about the absolute age of our events. Rather, we place our records on the GISP2 timescale to compare our results to Greenland air temperature history.
 We have also placed our data on the newly proposed age scale for the Greenland ice cores (SFCP 2004), which is based on 14C dating of foraminifera in core MD95-2042, calibrated by paired 14C and 230Th measurements on corals [Shackleton et al., 2004] (see Auxiliary Material). The conclusions that we report in the paper are the same regardless of which timescale we use for the Greenland ice core record.
4. Gulf of Mexico δ18O of Seawater
 The G. ruber δ18Oc values range from ∼−0.5 to −2.5‰ (Figure 3). This 2‰ variability is not seen in the δ18Oc of N. dutertrei (data not shown), an inferred deep dwelling planktonic foraminifer, suggesting that surface water phenomena are controlling the δ18Oc. The δ18Oc record exhibits four oscillations about a mean value of −1.25‰, from 28 to 45 ka (Figure 3). The δ18Oc values are more negative than the modern core top value of pink G. ruber (−1.7‰) during two of these oscillations (28.7–29.2 ka and 37.3–39.8 ka, Figure 3). Given that sea level was 63–93 m below present from 30 to 45 ka [Siddall et al., 2003], which would result in an enrichment of the foraminifera δ18Oc by ∼0.5–0.75‰ based on the relationship 0.083‰ per 10 m sea level change [Adkins and Schrag, 2001], δ18Oc values ≤−1.7‰ would indicate SSTs of 30°–32°C during MIS 3, which are unreasonably high compared to the modern average summer temperature in the GOM (29°C; June–September) [Levitus, 2004]. A change in δ18Osw associated with salinity variations is therefore required to explain the four negative oscillations recorded in the foraminiferal calcite.
 In order to isolate δ18Osw, we subtract the temperature component from the δ18Oc based on Mg/Ca data [Flower et al., 2004]. The Mg/Ca ratio, a proxy for the temperature of foraminiferal calcification, is ideal for δ18Osw calculations because it is measured on an aliquot of the calcite sample used for δ18Oc. A G. ruber (pink) calibration, based on Atlantic sediment trap data [Anand et al., 2003], was applied to the Mg/Ca measurements to calculate SST (Figure 3). We make the assumption that the effect of riverine input on the Mg/Ca values is minimal based on the large difference in Mg and Ca concentrations in the Mississippi River and the GOM (425 μM Mg versus 53 mM Mg; 870 μM Ca versus 10.3 mM Ca [Briggs and Ficke, 1978]). Despite the lower Mg/Ca ratio of Mississippi River water, oceanic Mg/Ca is not likely to be affected because the concentrations of Mg/Ca are low. A simple box model calculation shows that a 25% dilution of surface seawater (a likely maximum for G. ruber to withstand [Hemleben et al., 1989]) would only decrease Mg/Ca values by <3%, which is within measurement error.
 The Mg-SST component was removed from the δ18Oc using a temperature-δ18O relationship [Bemis et al., 1998] appropriate for G. ruber [Thunnell et al., 1999], resulting in the δ18Osw. The standard deviation for δ18Osw calculations is determined to be ±0.25‰, based on propagating the error through the analytical errors and the combined Mg-SST and SST-δ18O relationships [Beers, 1957]. The variances used for the Mg-SST and SST-δ18O equations are those reported in the literature. Variances for Mg/Ca and δ18O were based on replicate analyses.
 The δ18Osw variations from core MD02-2551 have similarities to the global sea level record from MIS 3 [Siddall et al., 2003] (Figure 3). However sea level fluctuations of <30 m during this interval [Siddall et al., 2003] can explain only 0.25‰ of the >1‰ δ18Osw changes observed in our record, suggesting that changes in evaporation/precipitation (E-P) or freshwater input must be the dominant control on the δ18Osw. We use the sea level record [Siddall et al., 2003] to remove the contribution of global ice volume to the δ18Osw, leaving the GOM δ18Osw residual (δ18OGOM) (Figure 4). This was accomplished by converting sea level height to the δ18O equivalent using the relationship 0.0083‰ per 1-m sea level change [Adkins and Schrag, 2001].
 The δ18OGOM values reflect changes in salinity, which result from a combination of source water variability and/or changes in the volume of water affecting the δ18OGOM signal. The δ18OGOM oscillates by up to 1.5‰, between more fresh versus more saline conditions, about a mean value of 0.45‰ (Figure 4). Major freshwater events, defined as intervals when the δ18OGOM reach values <0.45‰ and persist for >1.5 kyr, occurred from 31.7–34 ka and 37.2–39.8 ka (F2 and F4; Figure 4). The signatures of these two freshwater events are different, however: F2 is defined by a gradual change from more saline to more fresh conditions, while F4 is characterized by an abrupt freshening and an abrupt return to saline conditions. Three minor freshwater events, from 28.3–29.4, 35.0–35.5 and 42.9–43.8, also record values <0.45‰, but persist for <1.5 kyr (F1, F3 and F5; Figure 4).
5. Conversion to Sea Surface Salinity
 Conversion of δ18OGOM estimates to sea surface salinity (SSS) allows us to assess potential sources and magnitudes of freshwater flux to the GOM. SSS can be estimated using a δ18OGOM versus salinity relationship created for the GOM during MIS 3 (Figure 5). This relationship assumes conservative mixing between two end-members: high-salinity GOM waters (δ18Osw = 1.2‰ and S = 36.5 psu) and a low-salinity end-member. The low-salinity end-member is modeled using three different compositions: (1) a −3.5‰ value for GOM precipitation [Bowen and Revenaugh, 2003], and a Laurentide Ice Sheet (LIS) value ranging from (2) −15‰, reflecting the δ18O of source waters that drained from the LIS [Yapp and Epstein, 1977], to (3) −30‰, the average composition of the LIS [Dansgaard and Tauber, 1969]. It should be noted that the more negative the zero salinity intercept, the smaller the changes in the estimated salinity variations (Figure 5). For example, a 1‰ change in δ18OGOM is equivalent to ∼1 psu on the −30‰ LIS mixing line, ∼2 psu on the −15‰ LIS mixing line and ∼8 psu on the −3.5‰ MR mixing line.
 Use of the −3.5‰ end-member would require changes in salinity of up to 10 psu (Figure 6) and a volume of water 3–5 times the largest historical flood [Barry, 1997], or >50X the annual precipitation in the GOM [Ropelewski and Halpert, 1996], lasting for 3 kyr during the largest event. It is possible that the isotopic composition of continental precipitation draining into the Mississippi River was more negative during MIS 3, because of changes in the altitude and/or sources of precipitation. However, a minimal change in the δ18O composition of precipitation during MIS 3 is inferred from model simulations, which show similar δ18O precipitation values between the Last Glacial Maximum and present [Charles et al., 2001]. In addition, midcontinent speleothems, which reflect the changing isotopic composition of meteoric waters, record <0.5‰ variations in δ18O during this interval [Dorale et al., 1998]. We cannot rule out the possibility that increased precipitation over the GOM may reflect an intensification of the North American monsoon system, which is known to bring moisture to the region. However, the amount necessary to create the observed changes in the δ18OGOM record does not support oceanic precipitation as a primary control on this signal. In contrast, meltwater derived from the LIS with a δ18O composition of −15 to −30‰ would require only modest changes in salinity: A −15‰ end-member for the LIS results in a salinity change of up to 3.5 psu, while a −30‰ end-member results in a change in salinity of up to 2 psu (Figure 6). Additionally, the average SSS using a −30‰ end-member is 35.5 ± 1 psu, which is within the modern salinity range in the GOM.
 We recognize that the source of fresh water likely changed through time and may have been a mixture of various sources (i.e., meltwater and precipitation), and therefore the SSS calculations only reflect the end-member scenarios. Regardless, the most conservative estimate for salinity changes indicates a substantial meltwater contribution to δ18Osw in the GOM, particularly when the δ18O composition of GOM waters were most depleted This explanation is supported by recent reconstructions of the LIS during MIS 3, which place the margin of the ice sheet within the MR drainage basin [Dyke et al., 2002].
6. LIS Routing Hypothesis
 The uncertainty in the calibration of 14C to calendar years precludes firm phase comparisons, but there appears to be no consistent relationship between δ18OGOM freshwater input and Greenland interstadials. The LIS routing hypothesis would predict that the nine D/O warmings (IS 4–12) that span 28–45 ka [Grootes et al., 1993] should correspond to freshwater routing to the GOM [Clark et al., 2001], but only five δ18OGOM freshwater events are recorded in the Orca Basin during this interval (Figure 4). There is no age model that we can construct with the 14C dates that would allow the δ18OGOM record from Orca Basin to be on the same timing as the D/O cycles in Greenland. In addition, the Laschamp event coincides with a warming in Greenland (IS 10), but a positive δ18O excursion (more saline) in our record. If each of the D/O warmings corresponds to freshwater routing to the GOM, we would expect to see a negative δ18OGOM excursion in our record during this interval. Although freshwater routed to eastern outlets may have led to NADW reductions and coolings in Greenland, the timing and number of δ18OGOM freshwater events to the GOM suggest that a simple routing hypothesis cannot explain all of the MIS 3 Greenland interstadials. It appears that the D/O warmings cannot be attributed to changes in the strength of NADW associated with southward routing of meltwater by the LIS, which may help explain why it has been difficult to find NADW changes during each of the D/O cycles [Curry et al., 1999; Hagen and Keigwin, 2002; Vautravers et al., 2004]. Additionally, SST in the GOM does not appear to be coupled to Greenland air temperature.
 The δ18OGOM record has similarities to the Antarctic air temperature record [Johnsen et al., 1972], the global sea level record from MIS 3 [Siddall et al., 2003], and to the classic MIS 3 benthic δ18O record off Portugal [Shackleton et al., 2000]. Freshwater events in the GOM have a tendency to coincide with intervals of Antarctic warming. In particular, the largest freshwater event (F4) occurred at the same time as the largest warming in Antarctica (A1 centered at 39 ka; Figure 4) and a 30-m rise in sea level also at 39 ka [Siddall et al., 2003].
 Our δ18OGOM record suggests summer melting on the southern margin of the LIS during Antarctic warming, as also observed during the last deglaciation [Flower et al., 2004]. This provides evidence to support a recent modeling study that suggests that the Northern Hemisphere ice sheets contributed one half of the global sea level rises observed between 35 and 65 ka [Rohling et al., 2004]. Our results are also consistent with a new coupled atmosphere-ocean simulation that predicts that freshwater discharge into the Gulf of Mexico would contribute to Antarctic warming [Knutti et al., 2004]. LIS melting associated with the A1 warming in Antarctica may have provided a positive feedback for Southern Hemisphere warming through changes in the strength of NADW. Similarly, our results indicate that growth/decay cycles of the LIS may have been decoupled from Greenland air temperature history during MIS 3. Our finding underscores recent work suggesting that the LIS (which is influenced by summer melting) does not follow Greenland air temperature (which is influenced by winter temperatures, particularly during stadials) and that seasonality is an important aspect of abrupt climate change [Denton et al., 2005].
 We thank the IMAGES program, Yvon Balut, and Laurent Labeyrie for a productive cruise on the R/V Marion Dufresne in 2002. This work was supported in part by the National Science Foundation under grant OCE-0318361 to B.P.F. and T.M.Q. We thank the IMAGES program also for collection of the core. Ethan Goddard provided technical assistance with data collection. We also thank Luis Garcia-Rubio and Gary Mitchum for assistance with error analysis, Richard Poore and Kelly Kilbourne for comments on the manuscript, and Jyotika Virmani and the USF Paleo lab members for useful discussions. The manuscript was improved through the suggestions of several anonymous reviewers and Larry Peterson, who served as Editor. Radiocarbon analyses were performed under the auspices of the U.S. Department of Energy by the University of California Lawrence Livermore National Laboratory (contract W-7405-Eng-48).