Our understanding of past ocean-climate interactions has been hampered by a paucity of proxy reconstructions of physical ocean circulation parameters. Here we present the first detailed record of deepwater temperature variability that resolves Dansgaard-Oeschger climate fluctuations during the last glaciation. Results indicate a positive coupling between abrupt temperature changes over Greenland and in the deep North Atlantic that was mediated by perturbations to the Atlantic overturning circulation. The occurrence of sharp warming events at the center of the “Heinrich” stadials 4 and 5 further indicates that the response of the ocean interior was less static across these major climate events than has been previously apparent.
 In order to obtain a precise, and therefore predictive, theory regarding the behavior of the overturning circulation, and its possible role in climate change, we must first be able to describe exactly and to generalize the conditions of stability of the Atlantic meridional overturning circulation system [Stocker and Marchal, 2000]. One method for refining our understanding of the fundamental parameters that control the overturning circulation is to compare model outputs for past ocean circulation (i.e., hydrographic) states with constraints set by paleoceanographic proxy data. However, accurately describing and characterizing past hydrographic conditions remains a particularly challenging task; not least because of the long-standing difficulty of reconstructing physical parameters such as the temperature, salinity and/or flow rate of the deep ocean. To date, only a handful of late Pleistocene paleoceanographic studies have proposed to reconstruct past deepwater temperatures [Adkins et al., 2002; Cronin et al., 2000; Cutler et al., 1999; Labeyrie et al., 1987; Martin et al., 2002; Rühlemann et al., 1999; Skinner et al., 2003], and only two of these have proposed to directly measure millennial deep-sea temperature variability [Cronin et al., 2000; Skinner et al., 2003]. Here we present the first record of benthic Mg/Ca that unambiguously resolves millennial changes in the deep northeast Atlantic heat budget across a series of “Dansgaard-Oeschger” climate events during the last glaciation. This record is interpreted in terms of changes in the southward export of relatively warm northern-sourced deep water, and the resulting impact on the balance of “advective” and “diffusive transports” across the North Atlantic water column.
2. Materials and Methods
 Benthic foraminiferal Mg/Ca ratios have been measured in core MD01-2444 (37°33′N, 10°08′W; 2637 m) using the infaunal benthic foraminifer species Globobulimina affinis, which exhibits 1:1 linearly correlated δ18O values (and hence water temperature– δ18O sensitivity) with those of the epibenthic foraminifer species Planulina wuellerstorfii [Shackleton et al., 2000]. After subjecting the samples to a rigorous cleaning procedure, based on that of [Barker et al., 2003], and routine screening for dissolution and contamination artifacts, Mg/Ca ratios were converted to deepwater temperatures using the empirical calibration Mg/Ca = 2.71e0.103T. Parallel isotope measurements performed on the epifaunal benthic foraminifer P. wuellerstorfii (>212 μm) have also provided estimates of past deepwater δ13C.
 Contamination and dissolution artifacts have been ruled out as major controls on the observed Mg/Ca variability in MD01-2444 by reference to iron and aluminum concentrations and by reference to down core planktonic shell weight variations, none of which show a consistent positive correlation with benthic Mg/Ca. Calibration of benthic Mg/Ca to deepwater temperatures (Tdw) is based on combined Holocene and Eemian Mg/Ca measurements and Tdw constraints, and on minimum Mg/Ca ratios observed in the deep Pacific at the Last Glacial Maximum (LGM) being set to the freezing limit [Skinner and Shackleton, 2005; Skinner et al., 2003]. The resulting calibration essentially reproduces a previous core top calibration that made use of living G. affinis specimens [Tachikawa et al., 2003], and bears a temperature sensitivity that is similar to almost all available planktonic foraminiferal Mg/Ca calibrations (exponent ∼ 9%) [Anand et al., 2003].
 The adopted calibration (Figure 1) represents a preliminary attempt to quantify the temperature effect on G. affinis Mg/Ca, though one that is nonetheless robust in its consistency with known modern and past conditions. This work thus adopts the working hypothesis that Mg/Ca in G. affinis in MD01-2444 has remained strongly correlated, with a near constant sensitivity, to deepwater temperature. It has recently been proposed that Mg/Ca in the epifaunal benthic foraminifer Cibicides wuellerstorfii may become strongly controlled by a “carbonate ion effect” under very low temperature and carbonate saturation conditions [Elderfield et al., 2006]. This effect has not yet been explicitly observed in G. affinis, which has a facultative deep infaunal habitat and is therefore likely to have calcified at near-equilibrium carbonate saturation. Were the “carbonate ion effect” assumed to apply to G. affinis it would require deepwater carbonate ion saturation ([CO32−]) at the site of MD01-2444 to have changed repeatedly by up to 40 μmol/kg (according to the relationships determined by Elderfield et al. ). Such [CO32−] variability would probably be detectable, via a partial dissolution effect, in planktonic foraminifer shell weight [Barker et al., 2004; Brown and Elderfield, 1996; Regenberg et al., 2006]. A [CO32−] control on benthic Mg/Ca would thus be supported by positive correlation with a dissolution signal in planktonic shell weights. This is not observed in MD01-2444. On the contrary, a very strong anticorrelation is observed in MD01-2444 between this potential planktonic dissolution indicator and G. affinis Mg/Ca, thus arguing against large changes in carbonate ion as a significant control on the (infaunal benthic) G. affinis Mg/Ca record.
 The absolute uncertainty in the calculated Tdw estimates (because of statistical calibration uncertainty and analytical reproducibility) is estimated at ∼±0.7°C. However, the ostensible “noise” in the Tdw time series, which can be approximated by the average standard deviation of paired adjacent measurements (given the assumption that, as their depth offset tends to zero, adjacent measurements will tend to represent replicates of a single mean value), is closer to ±0.3°C or ∼±0.1 mmol/mol. This essentially provides an estimate of the degree of autocorrelation in the record.
 The results from MD01-2444 have been placed on the modified GRIP calendar age scale of Shackleton et al.  (hereinafter referred to as GRIP-SFCP04) by correlation of Dansgaard-Oeschger temperature fluctuations that are clearly recorded in the δ18O of Greenland ice and in planktonic δ18O from the Iberian Margin [Shackleton et al., 2000]. Age pins have thus been selected at the midpoint of each stadial-interstadial transition, allowing the transferal of GRIP-SFCP04 (ice core) ages to MD01-2444. Note that the GRIP-SFCP04 ages represent ice core ages that have been “calibrated” on the basis of two absolute age assignations, for Greenland interstadial (GIS) 3 (29 kyr B.P.) and for GIS 17 (59 kyr B.P.), with ages in between these points being dictated by previously determined GRIP ice accumulation rates [Shackleton et al., 2004]. This age scale is simply transferred to MD01-2444 by correlation of planktonic foraminiferal δ18O with GRIP δ18Oice. The construction of the chronology adopted here for core MD01-2444 is illustrated in Figure 2. In Figure 2 it is also shown that interpolated ages for successive stadial-interstadial transitions, as defined by the GRIP-SFCP04 age scale and by variously calibrated radiocarbon ages (transferred to MD01-2444 from core MD95-2042 [Shackleton et al., 2004]), as well as by independently dated speleothem records [Burns et al., 2003, 2004; Spötl and Mangini, 2002], are all in close agreement (always to within 2000 years and usually to within 1000 years). This degree of chronostratigraphic control is essential: in a first instance for comparisons with Greenland proxy temperature variability (dependent only on the strength of the correlation of Greenland and local temperature change); and in a second instance for comparisons with timescale characteristics observed in climate models (dependent only on absolute age control).
4. Results and Discussion
 Stable isotope and Mg/Ca results from MD01-2444, shown in Figure 3, indicate a positive coupling between deepwater temperature change in the northeast Atlantic, local deepwater δ13C and Greenland temperatures across Greenland interstadials (GIS) 8 to 12. This time interval comprises the two major ice-rafting periods, Heinrich events 4 and 5. Importantly, the observed coupling between high-latitude deepwater (benthic Mg/Ca) and surface water (planktonic δ18O) temperatures changes character, or breaks down, near the middle of the longer stadial periods that correspond broadly (but not precisely) to Heinrich events 4 and 5 [Bond et al., 1993]. These stadials (wide shaded bars in Figure 3) are hereinafter distinguished as Heinrich stadials, HS4 and HS5. Within these stadials a transient warming of the deep northeast Atlantic is observed, lasting ∼500 years and occurring without an unambiguous indication of a parallel and proportionate increase in local deepwater δ13C (narrow hatched bars in Figure 3). It is notable that both of these transient deepwater warming events occurred in parallel with a weakly expressed warming in the GRIP δ18O proxy temperature record (as well as small excursions in atmospheric CH4 [Blunier and Brook, 2001; Flückiger et al., 2004]). Other indications of mid-HS events have been identified in indicators of Gulf Stream circulation variability [Vautravers et al., 2004], in surface temperature and vegetation records from the Iberian Margin [Sanchez Goni et al., 2002; Vautravers and Shackleton, 2006], and in high-resolution planktonic δ18O records from the northwestern Mediterranean [Sierro et al., 2005]. The existence of such mid-HS events, which is clear only in the best resolved records, strongly suggests that “Heinrich stadials” (formally numbered Greenland stadials associated with major North Atlantic ice rafting) are not “single-event” phenomena that lasted ∼1000 years. This would concur with the recent suggestion that Heinrich event 4 may have lasted only a few hundred years, and may have resulted in only a small contribution to sea level rise [Roche et al., 2004].
 The relationships shown in Figure 3 suggest that northeast Atlantic deepwater temperature and δ13C have remained directly coupled with the Dansgaard-Oeschger climate variability that occurred in the North Atlantic and over Greenland, albeit with seemingly “disproportionate” warming occurring in the middle of the longer “Heinrich stadials.” Four stages in the deepwater temperature evolution across Heinrich stadials may be identified (with shorter “non-Heinrich” stadials only involving the first and last of these changes): (1) initial deepwater cooling with reduced δ13C, into the stadial; (2) a rapid deepwater warming without a clear change in δ13C, after ∼500 years of reduced deepwater temperature and δ13C; (3) a subsequent reversal to colder, less 12C-enriched deep water, again within less than 500 years; and (4) a final abrupt increase in deepwater temperature and δ13C when interstadial conditions resumed.
 In this oceanographic context, two main competing controls on ocean interior temperature change may be identified: (1) the turbulent diffusive “erosion” of vertical temperature gradients, generally between warm shallow water and cold deep water; and (2) the “lateral” advection of cold deep water, possibly of variable temperature and/or high-latitude origin. At present, the core site of MD01-2444 is bathed in recirculated Northeast Atlantic Deep Water (NEADW), consisting primarily of a mixture of Labrador Seawater (LSW), Denmark Strait Overflow Water (DSOW) and Iceland Scotland Overflow Water (ISOW). Deep water below the core site consists of a lower fringe of NEADW that obtains an increasing component of southern-sourced Lower Deep Water (LDW) with depth, until ∼4000 m where LDW dominates [van Aken, 2000a]. At intermediate depths above the core site, Mediterranean Outflow Water (MOW) flows northward, with a “core depth” of ∼1000 m [van Aken, 2000b]. There are therefore three possible bases for interpreting deepwater temperature change at the site of MD01-2444: (1) vertical migration, or indeed vertical diffusive “erosion,” of the overlying “mixing boundary” between MOW and “glacial NEADW” (hereinafter referred to as “glacial northern sourced deep water” (GNDW)); (2) changes in the character of GNDW, due to changing source conditions; and (3) vertical migration of the lower “boundary” between GNDW and “glacial LDW” (hereinafter referred to as “glacial southern sourced deep water” (GSDW)). Note that the first of these possibilities, in contrast to the others, may involve the dominance of vertical turbulent diffusive mixing over the lateral advection of northern- or southern-sourced deep water.
 In order to determine which of these hydrographic changes has been responsible for the deepwater temperature variability observed between GIS 7 to 13 we refer to the record of benthic δ13C. In principle, this analysis might be aided by a comparison of deepwater temperature with calculated deepwater δ18O (δ18Odw, derived from parallel benthic Mg/Ca and δ18O measurements), as a secondary conservative hydrographic tracer [Skinner et al., 2003]. However, in this context, changes in deepwater δ18O will have incorporated a time-transient glacioeustatic component [Chappell, 2002; Siddall et al., 2003], the timing (if not also the magnitude) of which remains unknown with sufficient precision. As a result, an analysis of the deepwater δ18O record that might be reconstructed in parallel with the deepwater temperature results presented here is sufficiently complex to require a separate and dedicated treatment, which we present elsewhere (L. C. Skinner et al., Phasing of millennial events and northeast Atlantic deep-water temperature change since ∼50 ka BP, submitted to Geophysical Monograph Series, AGU, 2006).
 Benthic δ13C is a nonconservative proxy for the conditions and mode of deepwater formation (the extent of ocean-atmosphere CO2 equilibration, and the “preformed nutrient” content of the deep water) as well as its subsequent overturning history (the accumulation of “remineralized” organic carbon). In the modern ocean, the δ13C of dissolved inorganic carbon (DIC), benthic δ13C, and the temperature of deep water (>2000 m) are positively correlated, primarily due to the fact that the warmest source of water to the deep ocean (North Atlantic Deep Water) is efficiently stripped of 12C-enriched particulate organic carbon while the coldest source of deep water (Antarctic Bottom Water) is not [Kroopnick, 1985; Levitus and Boyer, 1994]. This positive correlation is illustrated in Figure 4, where the range of variability of core top and Holocene (“modern”) benthic δ13C measurements (E. Michel, personal communication, 2006) is compared with the corresponding range of variability of modern deepwater temperatures. Also shown in Figure 4 are glacial deepwater “end-member” estimates for the North Atlantic, South Atlantic and Pacific (from Duplessy et al. ) as well as the deepwater temperature and δ13C values recorded in MD01-2444 between GIS 7 and 13. In order to facilitate a visual comparison of the modern Tdw and δ13C data with the paleoceanographic estimates, the former have been adjusted for approximate “whole ocean” glacial-interglacial Tdw and δ13C changes of −2.0°C [Duplessy et al., 2002] and −0.32‰ [Matsumoto and Lynch-Stieglitz, 1999; Shackleton et al., 1992] respectively. Also shown in Figure 4 is an estimate of the temperature-δ13C signature of glacial Mediterranean Outlfow Water (MOW) [Cacho et al., in 2006].
 What is clear from the comparisons shown in Figure 4 is that, with the possible exception of the “mid-HS warming events,” deepwater temperature and δ13C variability recorded in MD01-2444 essentially falls along the estimated mixing line between glacial northern- and southern-sourced deepwater end-members, which in turn mimic the modern mixing relationship. What this implies is that Tdw and δ13C changes recorded in MD01-2444 (shown in Figure 3) most likely reflect changes in the relative proportions of warmer GNDW versus colder GSDW bathing the core site in the past. Correlations between high-latitude surface cooling, midlatitude deepwater cooling, and reduced deepwater δ13C in Figure 3 would therefore reflect the impact of changes in the meridional overturning circulation on the efficiency of northward heat transport at the surface, and the resulting balance of southward versus northward heat transport (or cold water transport) at depth.
 During the “mid-HS warming events,” however, a different mechanism has been at play. As suggested in Figure 4, it is possible that these events involved a degree of mixing between GSDW (advected northward into the North Atlantic because of a reduction in the supply of GNDW) with overlying MOW. It is important to note that the observation of very low δ13C, yet relatively warm deepwater temperatures during these mid-HS events strongly suggests that any mixing with MOW must have involved a much colder and low-δ13C deepwater mass than appears to have bathed the core site at other times between GIS 7 and 13. In addition, it clearly precludes a complete “incursion” of MOW at the core site, which would require much more pronounced changes in Tdw and δ13C. The situation during mid-HS events might therefore reflect a pronounced incursion of GSDW in parallel with: (1) an elimination of the northern deepwater end-member (GNDW); (2) a brief but intense deepening of the MOW core on the Iberian Margin, leading to a slightly enhanced influence at even greater depth [Voelker et al., 2006; Zahn and Sarnthien, 1987]; and/or (3) an increase in the efficiency of vertical turbulent diffusive mixing (versus lateral cold water advection), bringing heat and 13C from an overlying (and very much warmer than 13C-enriched) MOW.
 None of these mechanisms is exclusive, and all may have operated in unison. What is essential, however, is that whatever mechanism was responsible for the mid-HS events, it must have been transient in nature, lasting only ∼ 250 years and resulting in an equally rapid return to prior conditions. If the mid-HS warming was due to vertical turbulent diffusive processes (bringing heat from the overlying MOW), this would require an approximate vertical diffusivity of at least ∼ 10−5 m2 s−1 (assuming Kv ≅ , for a vertical temperature gradient ∼ 0.01°C m−1, Δz ∼ 1000 m, ΔT ∼ 1°C and Δτ ∼ 250 years), which is in line with most “mean ocean” estimates [Munk and Wunsch, 1998; Naveira Garabato et al., 2004]. In this case, the subsequent reversal to colder, low-δ13C conditions would require either an exhaustion of the dominance of turbulent diffusive mixing over lateral advection of cold GSDW (possibly via some sort of localized “diffusive-advective oscillation” [Adkins et al., 2005; Schulz et al., 2002; Timmermann and Goosse, 2004; Winton, 1993]), or an end of the influence of MOW as source of heat that could be mixed downward by turbulent diffusive processes (possibly via a rapid shoaling of MOW). The latter possibility is supported by recent observations indicating very brief pulses of low δ18O measured in planktonic foraminifera from the western Mediterranean, precisely in time with the mid-HS deepwater warming events observed here [Sierro et al., 2005]. If these pulses of low planktonic δ18O were to indicate brief incursions of very low salinity Atlantic surface waters into the western Mediterranean, they could very well have triggered intensified pulses (and perhaps a deepening) of MOW on the Iberian Margin by accentuating the density gradient between the surface inflow and deep outflow through the Straits of Gibraltar. In any event, the fact that the mid-HS events are not only recorded in Iberian Margin deepwater temperature and western Mediterranean planktonic δ18O, but are also reflected in indicators of Gulf Stream warmth/intensity [Vautravers et al., 2004], atmospheric methane [Blunier and Brook, 2001; Flückiger et al., 2004] and Greenland temperatures (see Figure 3), argues strongly in favor of an atmospheric role in their generation and communication. A complete explanation for these mid-HS events remains lacking, but it is nonetheless clear from the results presented here that the response of the ocean interior across Heinrich stadials 4 and 5 was far less static than has been previously apparent from deepwater proxy data.
5. Concluding Remarks
 Just as the distribution and character of variability of sea surface temperatures provides a robust constraint on numerical model simulations of the past ocean circulation [CLIMAP Project Members, 1976], so too does the distribution and character of variability of ocean interior properties. This is particularly true of ocean interior temperatures, which directly influence density distributions and therefore dynamics in the ocean (both at steady state and in response to perturbations to its salinity structure for example). When a cooling of the deep Atlantic is well expressed in numerical model simulations of “Heinrich-type” overturning perturbations it tends to arise because of a reversed circulation scheme that brings very cold southern-sourced deep water into shallower depths in the North Atlantic, precisely as suggested by the data presented here from MD01-2444. However, models that reproduce such a response require two things: (1) a significant temperature gradient between colder southern-sourced water and warmer northern-sourced water, analogous to today; and (2) a “reversed” circulation scheme that brings southern-sourced water into the North Atlantic.
 The first of the above requirements is at odds with the commonly adopted assumption that the glacial ocean (and in particular the ocean during MIS 3) was of essentially homogenous temperature distribution (contrast Adkins et al.  and Duplessy et al. ), and both are far from being consistently simulated by numerical models of overturning circulation change. This point is illustrated in Figure 5, which shows water column temperature changes in the northeast Atlantic across a Heinrich event simulation, based on a number of numerical models of varying degrees of complexity. A “notional average” of these simulations indicates a depth-dependent response, with surface water cooling, intermediate warming (due to advective and/or diffusive processes) and deep to abyssal cooling in the North Atlantic. However, the variability among model simulations shown in Figure 5 is overwhelming, and depends on individual model design, parameterizations and initial/boundary conditions. Although the “survey” depicted in Figure 5 is far from exhaustive, it nevertheless illustrates the range of rather fundamental responses obtained by numerical models of meltwater-induced overturning circulation change, and underlines the sensitivity of model outcomes to the “initial state” of the ocean interior: an initial state that crucially remains essentially unknown for conditions other than modern.
 Convergence among model simulations of overturning circulation collapse is essential to our confidence in model explanations of past abrupt changes [Ganopolski and Rahmstorf, 2001; Manabe and Stouffer, 1995], as well as model predictions of future overturning stability [Joos et al., 1999; Manabe and Stouffer, 1994]. Arguably, only if a body of paleoceanographic data may be built up so as to set constraints on the ocean interior temperature response to overturning perturbations, for example from a range of water depths or across zonal and meridional profiles, is such convergence likely to be achieved. Here we have presented a first step in this direction, revealing a tight link between surface and deepwater temperature variability across Dansgaard-Oeschger climate events, and indicating a clear submillennial component of deepwater variability during “Heinrich stadials.”
 The authors would like to acknowledge the significant contribution of Nick Shackleton to this work through discussions, contribution of data, and motivation. Nick was not able to read the final draft of this manuscript and consequently could not give his formal consent for authorship and publication in Paleoceanography. We are also grateful for the expert assistance of Mike Hall, James Rolfe, and Mervyn Greaves at the University of Cambridge. This work was made possible by NERC research grant NER/B/S/2003/00815, by laboratory support from the Gary Comer Foundation, and by a Research Fellowship held by L. C. S. at Christ's College. L.C.S. is grateful for many insightful scientific discussions and exchanges of information with Michael Schulz, Matthias Prange, Andre Paul, Martin Butzin, Uta Krebs, Didier Roche, Elisabeth Michel, and Manuel Gloor. The authors are grateful to Laurent Labeyrie and Jess Adkins for their helpful comments, which helped to significantly improve upon an earlier version of this manuscript. L.C.S. would also like to acknowledge fruitful discussions with Axel Timmermann during a visit to the IPRC/SOEST, enabled by Honolulu University.