Core MD09-3257 (04∘14.7′ S, 36∘21.2′ W, 2344 m) was recovered in 2009 from the northern Brazilian margin during the R/V Marion Dufresne cruise MD173/RETRO3 at approximately the same position as core GeoB3910-2 (04∘14.7′ S, 36∘20.7′ W, 2362 m) (Arz et al., 2001; Jaeschke et al., 2007). The improved recovery of deep-sea sediments with little or no deformation of sediment layers was achieved thanks to the systematic use of the CINEMA software (Bourillet et al., 2007; Woerther and Bourillet, 2005). This software computes the amplitude and duration of the aramid cable elastic recoil, as well as the piston displacement throughout the coring phase, accounting for the length of the cable (water depth) and total weight of the coring system.

At present, the northern Brazilian margin is bathed by southward flowing upper North Atlantic Deep Water (NADW) at these depths (Lux et al., 2001; Schott et al., 2003; Rhein et al., 2015) (Fig. 1). The southward advection of dense waters formed at higher northern latitudes is channeled through the western boundary current (Rhein et al., 2015), meaning that our sediment cores are ideally located to detect changes in the transport of northern-sourced waters above a depth of 2500 m.

Elemental composition was measured by employing nondestructive, profiling X-ray fluorescence (XRF) spectrometry. The measurements were made using an AVAATECH XRF core scanner at the Bjerkness Centre for Climate Research, Bergen (Norway) at intervals of 0.5 mm on core MD09-3257, and using a CORTEX XRF scanner at the Bremen Integrated Ocean Drilling Program core repository at intervals of 0.4 cm on core GeoB3910-2 (Jaeschke et al., 2007). This automated scanning method allows for a rapid qualitative determination of the geochemical composition of the sediment at very high resolution (Croudace and Rothwell, 2015).

Ti / Ca records from core MD09-3257 and core GeoB3910-2 exhibit marked peaks corresponding to increased terrigenous input due to enhanced precipitation and runoff from the continent (Arz et al., 1998; Jaeschke et al., 2007) (Fig. 2). These precipitation events are also recorded in South American speleothems, and have been shown to correspond to North Atlantic cold stadial periods (Cheng et al., 2013). The core GeoB3910-2 radiocarbon (14C) age model shows that increases in sedimentary Ti / Ca are indeed synchronous with decreases in South American speleothem δ18O (Burckel et al., 2015). Based on the observed synchronicity, composite age models of core MD09-3257 and GeoB3910-2 have been developed using 14C dating for the past 35 kyr, combined with the alignment of sediment Ti / Ca increases with decreases in speleothem δ18O for the older portion of the cores (Vazquez Riveiros et al., 2018); thus, speleothem ages beyond 35 ka could be transferred to the marine cores. The chronology of core GeoB3910-2 is based on 17 monospecific radiocarbon dates between 0 and 31 ka (Burckel et al., 2015; Jaeschke et al., 2007). The Ti / Ca record of core GeoB3910-2 was aligned to that of core MD09-3257 in order to transfer the radiocarbon dates of GeoB3910-2 over the interval from 12 to 36 ka to this nearby core. In addition, five monospecific radiocarbon dates over 1–21 ka were directly obtained from core MD09-3257. Speleothem tie points were used to derive the chronology of this core over the period from 38 to 48 ka (Tables S1 and S2 in the Supplement) (Vazquez Riveiros et al., 2018). All radiocarbon dates were converted to calendar dates using the OxCal 4.2 software, the IntCal13 calibration curve (Reimer et al., 2013), and a surface water reservoir age of 550±50 yr over 0–18 ka (Key et al., 2004), and of 750±250 yr over 18–31 ka (Freeman et al., 2016). The final age models of cores GeoB3910-2 and MD09-3257 were obtained using a P_Sequence depositional model (Bronk Ramsey, 2008), i.e., a Bayesian algorithm producing posterior probability distributions for each core depth (Tables S1 and S2) (Vazquez Riveiros et al., 2018).

In the present study, the GeoB3910-2 age scale for the 32–50 ka interval was further adjusted by precise alignment of GeoB3910-2 XRF to the MD09-3257 XRF signal (Fig. S1 in the Supplement), which produced a composite record from these two nearby cores. Given that both XRF signals are virtually identical and measured at very high resolution (sampling step ≤0.5 cm), the mean relative dating uncertainty between the two cores is extremely small (less than 105 yr; Fig. S1).

Here, we use XRF ln(Ti / Ca) rather than Ti / Ca because log-ratios provide a unique measure of sediment composition, in contrast to simple ratios, which are asymmetric (i.e., conclusions based on evaluation of A/B cannot be directly translated into equivalent statements about B/A) and suffer from statistical intractability (Weltje and Tjallingii, 2008). We adopt the same terminology as Burckel et al. (2015) and define the larger ln(Ti / Ca) peaks as precipitation events PE0 to PE5, with PE0 occurring during the Younger Dryas, and PE1 to PE5 occurring during Heinrich stadials 1 to 5 (Fig. 2). What we refer to as Heinrich stadials are strictly those stadials characterized by the occurrence of iceberg discharges in the mid- to high-latitude North Atlantic. We refer to smaller ln(Ti / Ca) peaks corresponding to D–O stadials using the Greenland stadial numbering system, as defined by Rasmussen et al. (2014).

Epifaunal benthic foraminifers of the Cibicides wuellerstorfi species were handpicked in the > 150 µm size fraction (Vazquez Riveiros et al., 2018). Core MD09-3257 C. wuellerstorfi 13C / 12C (δ13C, expressed in ‰ versus Vienna Pee Dee Belemnite, VPDB) was measured at the LSCE on Finnigan DELTAplus and Elementar Isoprime mass spectrometers on samples of one to three specimens. VPDB is defined with respect to the NBS-19 calcite standard (δ18O = -2.20 and δ13C =+1.95 ‰). The mean external reproducibility (1σ) of carbonate standards is ±0.03 ‰ for δ13C; measured NBS-18 δ18O is -23.27±0.10 and δ13C is -5.01±0.03 ‰ VPDB. Core GeoB3910-2 C. wuellerstorfi gδ13C was measured at the University of Bremen, Germany, on a Finnigan MAT 252 mass spectrometer on samples of one to five specimens (Heil, 2006), with a mean external reproducibility (1σ) for carbonate standards of ±0.05 ‰ for δ13C. A composite high-resolution benthic isotopic record was generated by combining isotopic data from the upper 294 cm of core MD09-3257 (covering the last 32 kyr) with isotopic data from the interval from 246 to 451 cm in core GeoB3910-2 for the older part of the record (Vazquez Riveiros et al., 2018).

The 13C/12C isotopic ratio (δ13C, expressed in per mil –  ‰ – versus VPDB) of the epifaunal benthic foraminifer C. wuellerstorfi has been shown to record the δ13C of bottom-water dissolved inorganic carbon (DIC) with minor isotopic fractionation (Duplessy et al., 1984; Zahn et al., 1986; Schmittner et al., 2017). A water mass' initial DIC isotopic concentration is governed by surface productivity in its formation region (i.e., the preferential consumption of 12C by primary productivity, which increases dissolved δ13C), as well as temperature dependent air–sea exchanges (Lynch-Stieglitz et al., 1995). DIC δ13C subsequently decreases as deep water ages, due to the progressive remineralization at depth of relatively 13C-depleted biogenic material. As a result, DIC δ13C largely follows water mass structure and circulation in the modern ocean, and C. wuellerstorfi δ13C (δ13CCw hereafter) has been used to trace water masses and as a proxy of bottom-water ventilation (Duplessy et al., 1988 and numerous subsequent studies). A recent study further highlighted that DIC δ13C more faithfully follows water oxygen content than phosphate content (Eide et al., 2017), lending strong support to the use of δ13CCw as a proxy for bottom-water ventilation – the term ventilation here refers to the transmission of oxygen-rich, atmosphere-equilibrated water to the ocean interior.

New sedimentary (231Paxs,0/230Thxs,0) measurements (excess activity ratio at the time of deposition, Pa / Th hereafter) were produced in core MD09-3257 in order to extend the Pa / Th record of Burckel et al. (2015) and to cover the entire time interval from 10 to 43 ka (Table S3). The excess activity corresponds to the fraction of each radioisotope produced in the water column by uranium (U) decay and is transferred to the sediment by adsorption onto particles sinking in the water column. 230Th and 231Pa excess activities are calculated from bulk sediment measurement by correcting for the contribution of the detrital and authigenic fractions (François et al., 2004; Henderson and Anderson, 2003) using a detrital (238U / 232Th) value of 0.5±0.1 (2σ) (Missiaen et al., 2018). These excess activities are then further corrected for radioactive decay since the time of sediment deposition. Bulk sediment measurements were performed by isotopic dilution mass spectrometry on the LSCE MC-ICP-MS (NeptunePlus, Thermo Fisher), following a method derived from Guihou et al. (2010). Error bars (2 standard deviations) on Pa / Th measurements were computed by Monte Carlo runs (Missiaen et al., 2018), accounting for the uncertainties in Pa, Th and U measurements, as well as those of the detrital (238U / 232Th) value, spike calibrations and dating.

Sedimentary Pa / Th can be used to reconstruct changes in the renewal rate of water masses overlying the core site. This tracer has been successfully used to reconstruct past changes in deep Atlantic circulation intensity (Burckel et al., 2015 and references therein). 231Pa and 230Th are produced at a constant Pa / Th activity ratio of 0.093 by dissolved uranium, which is homogeneously distributed in oceans. 230Th is much more particle reactive than 231Pa, as reflected by their respective residence time in the ocean (30–40 yr for 230Th and 100–200 yr for 231Pa, François, 2007). Therefore, 230Th is rapidly removed from the water column to the underlying sediment, while 231Pa can be advected by oceanic currents. Thus, when averaged over an entire ocean basin, high (low) flow rates result in high (low) 231Pa export and a low (high) sedimentary Pa / Th ratio. In contrast to δ13CCw, which records the DIC δ13C of bottom waters at the core site, sedimentary Pa / Th does not reflect the flow rate at the seabed but that of a water layer of a few hundreds to more than 1000 m above the seafloor (Thomas et al., 2006).

Several potential caveats of the proxy were tested. In particular, 231Pa has a higher affinity for opal than for the other types of particles (Chase et al., 2002), meaning that high opal fluxes can result in high sedimentary Pa / Th values even in the presence of lateral advection. Similarly, areas of very high vertical particle flux, such as the Atlantic off western Africa, are characterized by high Pa / Th values (Yu et al., 1996; François, 2007; Lippold et al., 2012). Recent studies have shown that the caveats that may apply to this proxy in some areas do not apply to the western tropical Atlantic region. More specifically, a study including core top material from the western tropical Atlantic margin and using a 2-D model (Luo et al., 2010) showed that the measured Pa / Th vertical profile is consistent with a dominant role of the overturning circulation, rather than particle scavenging; this demonstrates that Pa / Th can be used to record changes in water mass overturning rates in that region (Lippold et al., 2011). However, because there are large increases in terrigenous material deposition on the northeastern Brazilian margin during the last glacial, we carefully evaluated/assessed if increased terrigenous deposition may have impacted Pa / Th values.

The 230Th-normalized 232Th flux, hereafter simply referred to as the 232Th flux, is indicative of the vertical terrigenous flux to the core site. As 232Th is a trace element that is mostly contained in the continental crust (Taylor and McLennan, 1985), it is commonly used as a geochemical tracer for material of detrital origin (e.g., Anderson et al., 2006). Besides the main precipitation events PE0 to PE4, there is no significant correlation between the Pa / Th ratio and the 232Th flux (r=0.21, p=0.07) (Figs. S2 and S3). In contrast, because the correlation between Pa / Th and the 232Th flux becomes significant (r=0.57, p≪0.001) when including the main precipitation events (Fig. S3), the high Pa / Th values observed during PE0 to PE4 could be partly caused by increased terrigenous flux and should be interpreted with caution (empty symbols in Fig. 2). Note that a possible terrigenous influence during the main precipitation events does not preclude that the high Pa / Th values during these periods reflect an almost halted oceanic circulation above the core site. Indeed, Pa scavenging by boundary scavenging can be intensified in times of reduced overturning circulation due to boundary scavenging becoming the main control on sedimentary Pa / Th.

Another source of possible biases in Pa / Th results from variations in opal flux (Chase et al., 2002). However, the northern Brazilian margin is known for its low siliceous primary production (Arz et al., 1998). This is confirmed by 230Th-normalized opal flux measurements in MD09-3257, which are below 0.06 g cm-2 kyr-1 (Fig. S3). Moreover, outside of precipitation events PE0 to PE4, there is no correlation between Pa / Th and opal flux (Fig. S3). In conclusion, we may consider that outside of the main precipitation events, our Pa / Th record can be interpreted in terms of changes in the strength of overturning circulation above the MD09-3257 coring site.

Assuming that there is a constant phase shift between two time series over their entire length, one can perform a simple cross-correlation analysis and compute how the correlation coefficient between the two time series varies as a function of the time lag between the two series (e.g., Davis, 1986).

We normalized (i.e., subtracted the mean and divided by the standard deviation) and resampled the time series Pa / Th, δ13CCw and ln(Ti / Ca) to a common age scale using scenarios with constant time steps varying between 50 and 500 yr. We then used the R function cor.test (R package stats version 3.2.2) for correlation between paired series (R script in Supplement) to compute the Spearman correlation coefficient between all pairs of the three time series, after having shifted one with respect to the other by increments of the time step.

Another approach consists of classical spectral analysis methods that examine the coherence and phase between two time series in frequency space, such as Fourier transforms. Fourier transforms involve decomposing a signal into infinite-length oscillatory functions (such as sine waves). As such, these methods also rely on the assumption that the decomposition of each signal into characteristic frequencies is valid over its entire length, i.e., that the underlying processes are stationary in time.

In contrast, wavelet analysis can be used to decompose a time series into “time–frequency” space, rather than frequency space, that is, to determine both the dominant modes of variability and how these modes vary in time (Torrence and Compo, 1998). To do so, the wavelet transform decomposes the signal into a sum of small wave functions of finite length that are highly localized in time. Thus, wavelet transform can describe changes in frequencies along the studied time series and are particularly relevant for dealing with climatic signals, since they are in essence not stationary in time, but in constant evolution in response to external forcing (i.e., insolation changes), and as a result of internal climate variability.

Given two times series X and Y, with wavelet transforms WX and WY, the cross-wavelet spectrum is defined as WXY=WXWY*, where WY* is the complex conjugate of WY (Torrence and Compo, 1998). Similarly to Fourier coherency, which is used to identify frequency bands in which two time series are related, the wavelet coherency was developed to identify both frequency bands and time intervals over which the two time series are related. The wavelet coherence between two time series is defined as the square of the smoothed cross-wavelet spectrum normalized by the smoothed individual wavelet power spectra (Torrence and Webster, 1999). This definition resembles that of a traditional correlation coefficient, i.e., wavelet coherence ranges between 0 and 1, and may be viewed as a localized correlation coefficient in time–frequency space (Grinsted et al., 2004).

Analogous to Fourier cross-spectral analysis, the phase difference between two time series can also be computed using a cross-wavelet spectrum. The complex argument arg(WXY) can be interpreted as the local relative phase between X and Y in time–frequency space (Grinsted et al., 2004).

In the present study we use the software developed by Grinsted et al. (2004) to compute the cross-wavelet spectrum, coherence and relative phase between our time series that were normalized and resampled as previously described. To test for the persistence of regions of high cross-wavelet coherence, we ran all cross-wavelet analyses 1000 times for each dataset pair (i.e., a Monte Carlo approach). For each of the 1000 runs, each time data point was randomly sampled, whereby a Gaussian distribution of each data point's value (based on the measurement uncertainty) was used to weight the random sampling. Mean and standard deviation values for the coherence and phase direction were calculated using the 1000 runs.

The Pa / Th record of core MD09-3257 now covers the entire 10–43 ka time interval, encompassing the Younger Dryas (YD) and the last four Heinrich stadials (Fig. 2). We have increased its temporal resolution over the time interval from 31 to 38 ka comprising Dansgaard–Oeschger (D–O) events 5 to 8, with respect to the rest of the study period, in order to examine Atlantic circulation dynamics during D–O events.

Pa / Th data exhibit systematic increases in conjunction with stadials, even if Pa / Th data points that are potentially biased towards elevated values by increased terrigenous input (empty symbols) are discarded. Thus, Pa / Th data indicate that the renewal rate of the water mass overlying the site decreased during stadials. More specifically, transport of the overlying water mass decreased not only during the YD and Heinrich stadials, but also during practically all D–O stadials. Among D–O stadials, the Pa / Th increase is well marked for GS-7, GS-8 and GS-11, but the signal is too noisy to provide a clear picture for GS-6. This noisy Pa / Th signal is very likely due to sediment reworking, given that the δ13CCw record is also noisy over this section of the core. Also, it is noteworthy that no precipitation event is recorded in MD09-3257 or GeoB3910-2 during GS-10 (Fig. 2). There is no clear decrease in the well-dated El Condor (Cheng et al., 2013) speleothem δ18O records associated with GS-10 either, which is in contrast with the other Greenland stadials (Burckel et al., 2015). It would seem that there was no apparent increase in precipitation during GS-10 over tropical South America, in contrast with all other GS of the past 40 kyr. Overall, longer stadials seem to be associated with larger increases in Pa / Th than shorter stadials.

The δ13CCw composite record varies in concert with Pa / Th, with high values indicating the presence of well-ventilated waters during the Holocene and interstadials, and low values indicating a marked reduction in water ventilation during stadials at ∼2350 m in the western equatorial Atlantic (Vazquez Riveiros et al., 2018).

Pa / Th, δ13CCw and Ti / Ca are recorded in the same core or in two cores from the same location, which could be precisely aligned through high resolution XRF signals. This situation provides ideal conditions to examine the relative phasing of one proxy with respect to another. Pa / Th and Ti / Ca are recorded in the same core, so their relative phasing can consequently be examined with the smallest possible relative dating uncertainty, whereby the only remaining source of uncertainty is bioturbation. The situation is practically the same when examining δ13CCw versus Ti / Ca or δ13CCw versus Pa / Th. Apart from the unavoidable uncertainty introduced by bioturbation, the relative dating uncertainty between the δ13CCw composite record and any MD09-3257 record is null over 0–32 ka, and amounts to 102 yr on average over the 32–50 ka time interval (Fig. S1).

In what follows, we assess the relative phasing between Pa / Th, δ13CCw and Ti / Ca, using all Pa / Th data points (including Pa / Th values susceptible to being partially impacted by large particle fluxes or boundary scavenging resulting from slower overturning circulation) in order to have sufficient data to examine periodicities ranging from 1000 to 6000 yr. In doing so, we assume that changes in particle fluxes may affect the amplitude of the Pa / Th changes, rather than the timing of these changes. In the following text, we show that excluding the Pa / Th values susceptible to being partially impacted by large particle fluxes does not change our conclusions concerning D–O periodicities (i.e., 1000 to 3000 yr).

Oceanographic studies have shown that the southward transport of northern-sourced waters in the equatorial Atlantic mainly takes place between a depth of ∼1300 and 4000 m in a ∼100 km wide Deep Western Boundary Current (DWBC) (Lux et al., 2001; Rhein et al., 2015). Using hydrographic, geochemical and direct velocity measurements acquired in 1993 to inverse an ocean circulation model, Lux et al. (2001) estimated that the volumetric flow of upper NADW occupying water depths between ∼1300 and 2300 m at 4.5∘ S within the DWBC is 11.2 Sv (1 Sv = 106 m3 s-1). This estimate is in good agreement with the 10.9 Sv estimated by Schott et al. (2003) based on data from 13 shipboard current-profiling sections taken during the World Ocean Circulation Experiment period (1990–2002).

Our data show that outside of the main precipitation events, the total vertical particle flux did not vary much (remaining within 25.1±3.6 g m-2 yr-1, 1σ) (Fig. S2). The Pa / Th values of these interstadials are similar or slightly higher than those of the late Holocene (Fig. 2), suggesting that the transport of the water mass overlying the MD09-3257 core site was also ∼10 Sv during these interstadials.

It is more difficult to translate the observed increases in Pa / Th during stadials into quantified decreases in water mass transport. However, our new data bring additional observational constraints on the Atlantic circulation changes associated with last glacial millennial climate changes. Two recent studies have indicated that decreases in northern-sourced deep water flow took place during each stadial. On the one hand, increases in Pa / Th during each stadial of the last glacial have been observed at a very deep western North Atlantic site located at ∼42∘ N and a depth of 4500 m (Henry et al., 2016). On the other hand, reconstructions of water corrosiveness in a South Atlantic core located at ∼44∘ S and a depth of 3800 m indicate the absence of northern-sourced deep water at that site during stadials, whereas nearly all interstadials of the last 60 kyr are characterized by incursions of northern-sourced deep water into the deep South Atlantic (Gottschalk et al., 2015). Together with these independent results, our results indicate that decreases in both the flow rate and extension of northern-sourced deep waters during stadials were not limited to very dense waters circulating at 3800 m or deeper, but also affected water mass transport above 2350 m in the western equatorial Atlantic.