Climate of the Past Towards orbital dating of the EPICA Dome C ice core using δ O 2 / N 2

Based on a composite of several measurement series performed on ice samples stored at −25C or −50C, we present and discuss the first δO2/N2 record of trapped air from the EPICA Dome C (EDC) ice core covering the period between 300 and 800 ka (thousands of years before present). The samples stored at −25C show clear gas loss affecting the precision and mean level of the δO2/N2 record. Two different gas loss corrections are proposed to account for this effect, without altering the spectral properties of the original datasets. Although processes at play remain to be fully understood, previous studies have proposed a link between surface insolation, ice grain properties at close-off, and δO2/N2 in air bubbles, from which orbitally tuned chronologies of the Vostok and Dome Fuji ice core records have been derived over the last four climatic cycles. Here, we show that limitations caused by data quality and resolution, data filtering, and uncertainties in the orbital tuning target limit the precision of this tuning method for EDC. Moreover, our extended record includes two periods of low eccentricity. During these intervals (around 400 ka and 750 ka), the matching betweenδO2/N2 and the different insolation curves is ambiguous because some local insolation maxima cannot be identified in theδO2/N2 record (and vice versa). Recognizing these limitations, we restrict the use of our δO2/N2 record to show that the EDC3 age scale is generally correct within its published uncertainty (6 kyr) over the 300–800 ka period.


Introduction
While ice core records offer a wealth of paleoclimatic and paleoenvironmental information, uncertainties associated with ice core dating limit their contribution to the understanding of past climate dynamics. Absolute age scales have been constructed for Greenland ice cores thanks to layer counting in sites offering sufficient accumulation rates (GRIP, GISP2, NorthGRIP; Rasmussen et al., 2006;Svensson et al., 2006Svensson et al., , 2008, allowing the construction of the GICC05 Greenland age scale currently spanning the past 60 ka (i.e. thousand of years before present, present being year 1950 AD in our study). While layer counting is not possible for deep Antarctic ice cores obtained in low accumulation areas, the transfer of the GICC05 age scale (using gas synchronization methods; e.g. Blunier et al., 2007) to Antarctic records allows researchers to partly circumvent this difficulty for the past 60 ka. Absolute time markers are generally lacking for these long Antarctic records, now extending up to 800 ka, with the exception of promising studies using Ar/Ar and U/Th dating tools (Dunbar et al., 2008;Aciego et al., 2010) and the links between 10 Be peaks and well dated magnetic events . As a result, dating of the deepest part of these Antarctic cores is largely based on various approaches combining an ice flow model with orbital tuning. Classically, orbital tuning assumes that northern hemisphere summer insolation drives large climate transitions (e.g. Milankovitch, 1941), and has long been used for dating paleoclimatic records, especially marine ones (e.g. Martinson et al., 1987).
As for Antarctic ice cores, two different orbital dating approaches, initially developed by Bender et al. (1994) and Bender (2002), are now commonly used. First, long records of δ 18 O of atmospheric O 2 (δ 18 O atm ) have revealed that this parameter is highly correlated with insolation variations in the precession band with a lag of about 5-6 kyr (thousands of years) (Bender et al., 1994;Petit et al., 1999;Dreyfus et al., 2007). Studies have linked variations in precession to δ 18 O atm through changes in low latitude water cycle and biospheric productivity (Bender et al., 1994;Malaizé et al., 1999;Wang et al., 2008;Severinghaus et al., 2009;Landais et al., 2007Landais et al., , 2010. The significant time delay between changes in precession and changes in δ 18 O atm has been attributed to a combination of the 1000-2000 year residence time of O 2 in the atmosphere (Bender et al., 1994;Hoffmann et al., 2004) and to the numerous and complex processes linking the isotopic composition of seawater to atmospheric oxygen via the dynamic response of the tropical water cycle to precession forcing and the associated variations in terrestrial and oceanic biospheres (Landais et al., 2010, and references therein). This superposition of processes also suggests that lags may vary with time (Jouzel et al., 2002;Leuenberger, 1997). As a consequence, the δ 18 O atm record from long ice cores can be used to constrain ice core chronologies (e.g. Petit et al., 1999;Shackleton, 2000), but with a large associated uncertainty (6 kyr) (Petit et al., 1999;Dreyfus et al., 2007). In parallel, the link between precession, low latitude hydrology, and atmospheric methane concentration (Chappellaz et al., 1993) has been used to propose an orbital age scale for Vostok (Ruddimann et al., 2003). However, past methane variations exhibit a strong impact from obliquity (Loulergue et al., 2008) and a weaker correlation with precession than δ 18 O atm (Schmidt et al., 2004;Landais et al., 2010), hence limiting this approach.
Second, Bender (2002) has proposed that the elemental ratio δO 2 /N 2 in the trapped air could be used as a new orbital tuning tool. Indeed, δO 2 /N 2 measurements in the firn near the pore close-off depth (about 100 m below the icesheet surface, i.e. where unconsolidated snow is compressed to the density of ice) have revealed that the trapping process is associated with a relative loss of O 2 with respect to N 2 (Battle et al., 1996;Severinghaus and Battle, 2006;Huber et al., 2006). Between 160 and 400 ka, the δO 2 /N 2 record of the Vostok ice core displays variations similar to those of the local 21 December insolation (78 • S). From these two observations, Bender (2002) formulated the hypothesis that local Antarctic summer insolation influences near-surface snow metamorphism and that this signature is preserved during the firnification process down to the pore close-off depth, where it modulates the loss of O 2 . From this hypothesis, he proposed the use of δO 2 /N 2 for dating purposes.
Despite a limited understanding of the physical mechanisms linking local 21 December insolation and δO 2 /N 2 variations in polar ice cores, this approach has been used by Kawamura et al. (2007) and Suwa and Bender (2008a) to propose an orbital dating of the Dome F and Vostok ice cores back to 360 and 400 ka, respectively. The validity of the link with local summer insolation has been supported by a similar correspondence observed in the Greenland GISP2 ice core (Suwa and Bender, 2008b). Using their high quality δO 2 /N 2 record on the Dome F ice core and comparison with radiometric dating obtained on speleothem records,  estimated the dating uncertainty to be as low as 0.8-2.9 kyr. Moreover, it was suggested that, combined with an inverse glaciological modeling approach, the dating uncertainty could be pinched down to 1 kyr .
Up to now, the oldest ice core climatic and greenhouse gases records have been obtained from the EPICA Dome C (EDC) ice core that covers the last 800 ka Lüthi et al., 2008;Loulergue et al., 2008). The state-ofthe-art dating of the EDC ice core (EDC3 chronology) has been described in . It is based on ice flow modeling using an inverse method constrained by available age markers. These age markers include reference horizons such as volcanic horizons (Mt Berlin eruption, 92.5 ka; Dunbar et al., 2008) and peaks in 10 Be flux (i.e. Laschamp event, 41.2 ka; Yiou et al., 1997;Raisbeck et al., 2008). Other tie points have been introduced based on the comparison of the ice core records with records of other well dated archives: as an example, the abrupt methane increase at Termination 2 was assumed to be synchronous (within 2 kyr) with the abrupt δ 18 O of calcite (speleothem) shift recorded in Chinese (Yuan et al., 2004) and Levantine (Bar-Mathews et al., 2003) regions at around 130.1 ka. For the last 42 ka, the EDC3 age scale was synchronized with the layer-counted Greenland GICC05 chronology (Svensson et al., 2008). For ice older than the last interglacial period, tie points are exclusively derived from orbital tuning. In addition to 37 δ 18 O atm tie points used between 400 and 800 ka, additional orbital information was derived from local insolation changes imprinted in the record of total air content in polar ice. Raynaud and colleagues (2007) indeed showed that the majority of the variance in total air content in the EDC ice core can be explained by the variations of an integrated summer insolation parameter (i.e. summation of the daily insolation over a certain threshold for a given latitude) that has a dominant obliquity component. This marker was therefore suggested as another tool for orbital dating of ice core records. Moreover, the study by Lipenkov et al. (2011) shows that the two air content and δO 2 /N 2 local insolation proxies lead to orbital timescales that agree to within less than 1 kyr on average. Ten such "air content" tie points have been used between 71 and 431 ka for EDC3, assuming a 4 kyr uncertainty to account for the scatter in the raw data and the uncertainty due to the choice of the integrated summer insolation target (threshold value for daily insolation). The overall uncertainty attached to the EDC3 time-scale is estimated at 6 kyr from 130 ka down to the bottom of the record. As a result, the uncertainty on event durations can reach 40 % between 400 and 800 ka (i.e. over the period mainly constrained by δ 18 O atm orbital tuning) .
In this article, we present the first records of δO 2 /N 2 measured on the EDC ice core between 800 and 300 ka. This record is of special interest since (1) no δO 2 /N 2 data were used for constraining the EDC3 age scale, and (2) no δO 2 /N 2 data have been published to date prior to 410 ka. Our record includes two periods of low eccentricity, centered at around 400 ka and around 750 ka, characterized by a minimum influence of precession variations on insolation (Loutre and Berger, 2003). This contrasts with the relatively large eccentricity context for the time interval between 50 and 360 ka, where previous δO 2 /N 2 records were obtained. Our δO 2 /N 2 record was therefore used for examining (1) the feasibility of orbital dating (in particular with δO 2 /N 2 ) at times of low eccentricity, and (2) the validity of the EDC3 age scale between 300 and 800 ka with respect to the δO 2 /N 2 constraints.
We first present the analytical methods used to perform the measurements, as well as discuss the effects of gas loss associated with ice storage on the integrity of the δO 2 /N 2 record and the necessary corrections. The spectral properties of the resulting composite EDC δO 2 /N 2 curve are then analyzed with respect to the orbital forcing. The local insolation influence on this EDC signal is compared with previous studies on the Vostok and Dome F ice cores (Bender, 2002;Kawamura et al., 2007;Suwa and Bender, 2008a). Finally, the uncertainties and limitations attached to the use of δO 2 /N 2 as a dating tool for the EDC ice core between 300 and 800 ka are discussed.

Technique used at LSCE for obtaining the EDC δO 2 /N 2 record
The technique used at the Laboratoire des Sciences du Climat et de l'Environnement (LSCE) for extracting the air trapped in ice core samples is based on melting and refreezing ice samples as first developed by Sowers et al. (1989) and detailed in Landais et al. (2003). For each depth, two adjacent ice samples covering the same depth interval were cut from the ice core. Approximately 3-5 mm of the outer ice was shaved off each face to yield two 10 g ice samples. The samples were placed in pre-cooled glass flasks with glass/metal transitions to Conflat flange tops. The flasks were connected to a vacuum manifold using gold-plated copper gaskets. The manifold is equipped with a Pfeiffer-Balzar turbo molecular pump, two pressure gauges (Baratron, Pirani), manual Nupro valves and 6 ports. Typically, we processed 6 samples per day with this method in two batches of three samples. Following a leak test, the ambient air surrounding the ice samples was evacuated using the turbo-molecular pump for 35-40 min while the ice was kept frozen by immersing the flask in a −20 • C ethanol bath. The flask was then isolated using a manual Nupro valve, and the ice was allowed to melt at room temperature. Once the samples were completely melted, we began refreezing the first sample. Since only one sample can be cryogenically transferred at a time, the samples were refrozen sequentially. Refreezing was accomplished using a 10 cm long copper cold finger with flat top plate placed in contact with the bottom of the sample flask. Only the bottom 3 cm of the cold finger were initially immersed in liquid nitrogen. Heat transfer was facilitated by adding alcohol to the top plate in contact with the flask bottom. This arrangement allowed the melt water to refreeze slowly from the bottom, minimizing the capture of dissolved gases. Cracking of the ice signals refreezing was complete, at which time we completely immersed the cold finger in liquid N 2 . In order to remove residual water vapor from the headspace, we heated the metal flange connection with a heat gun for 2.5 min, then maintained the cold finger maximally immersed in liquid N 2 for an additional 10 min, a procedure that ensures that the sample flask is never in direct contact with liquid N 2 . The headspace gases were then cryogenically transferred into a quarter inch steel tube plunged into liquid He for 6 min. The gases in the stainless steel tube were allowed to come to room temperature and equilibrate for a minimum of 40 min before being introduced into the mass spectrometer for isotopic and elemental analysis using a dual inlet system. The measurements of δO 2 /N 2 on the EDC ice core were performed on two different mass spectrometers. The Series 1 (Table 1, Fig. 1) was obtained on a 4-collector Finnigan MAT 252. On this mass spectrometer, the masses (m/z 32 (O 2 ) and 28 (N 2 ) could not be measured simultaneously. so peak jumping (jumping from one mass to the other) was used to measure δO 2 /N 2 . The measurement Series 2, 3 and 4 (Table 1, Fig. 1) were measured on a 10-collector Thermo Delta V Plus that permitted simultaneous acquisition of m/z 32 and 28. A careful inter-comparison of the performances of the two mass spectrometers using air standard and firn air on the two instruments showed no significant offsets (Dreyfus, 2008 Fig. 1). We indicate the number of the series as referenced in the text, the number of depth levels studied, the mean depth interval, the mass spectrometer on which the δO 2 /N 2 measurements were carried out, the year when the analyses were performed, the pooled standard deviation, and the numbers of depth levels rejected associated with each series. Series 1 and 2 were obtained from ice stored at −25 • C, while Series 3 and 4 were obtained from ice stored at −50 • C. ice core plotted on the EDC3 ice core age scale  and associated uncertainty (1 σ , bottom panel). The EDC δO 2 /N 2 record is composed of four distinct measurement series (top panel). Samples measured in Series 1 and 2 were stored at −25 • C whereas samples measured in Series 3 and 4 were stored at −50 • C. Characteristics of the different series are given in Table 1: Series 1 in red, Series 2 in green, Series 3 in blue, Series 4 in purple.

Series
The method for δO 2 /N 2 measurements at LSCE is similar to the one used for obtaining the Vostok δO 2 /N 2 record (Sowers et al., 1989;Bender, 2002). However, it significantly differs from the one used for the Dome F δO 2 /N 2 record  which requires a much larger ice sample (∼200 g instead of 10 g) and is based on a gas extraction method with no refreezing . No inter-calibration of these different methods has yet been conducted.

EDC raw data
The measurements have been performed on clathrate ice below 2400 m depth well below the bubble -clathrate transition zone where positive δO 2 /N 2 values have been observed in other records (Bender, 2002). This effect is due to O 2 being more easily dissolved in ice as gas hydrate than N 2 , such that after coring, N 2 (from bubbles) is preferentially lost relative to O 2 (from clathrates) in this zone (Bender, 2002;Ikeda-Fukazawa et al., 2005). The complete record is a composite of four different series of measurements from ice with different storage histories (Table 1, Fig. 1). All data have been corrected for gravitational fractionation as follows: For all series, each sample value corresponds to the average of at least two replicate samples analyzed at each depth level. We then calculated the pooled standard deviation as: where n i and σ i are the sample size and the standard deviation of the i-th sample, respectively. Over the whole set of measurements, σ p is equal to 1.5 ‰.
As expected from the inter-comparison between the two mass spectrometers, no shift between the mean values appears between the 1st and the 2nd series. The oldest values (700-800 ka) obtained in Series 2 are associated with a large scatter of the δO 2 /N 2 data between neighboring and replicate samples. For two of these depth levels, δO 2 /N 2 reaches extremely low values (lower than −40 ‰) with an associated standard deviation of 10 ‰ (Fig. 1). These ice samples are likely affected by significant gas loss after ice coring that favors the loss of the smaller molecule O 2 with respect to the larger molecule of N 2 (Huber et al., 2006).
On the contrary, samples from Series 3 (stored at −50 • C after ice core drilling instead of at −25 • C) are associated with a very low pooled standard deviation (0.32 ‰) and less scattering than Series 1 and 2. This shows that high precision Clim. Past, 8, 191-203, 2012 www.clim-past.net/8/191/2012/ δO 2 /N 2 measurements on EDC ice are possible with our experimental set-up if the ice is stored at a very low temperature. Moreover, samples from Series 4, also stored at −50 • C but measured one year later than the samples from Series 3, show a significantly lower scattering than the measurements performed over the same depth range from Series 2 from ice stored at −25 • C. This again confirms the quality of δO 2 /N 2 record from EDC ice samples stored at −50 • C. Still, the uncertainty associated with Series 4 is larger than the one associated with Series 3 despite the fact that the samples were stored under the same conditions. This may be related to the ice history, with warmer temperatures encountered near Dome C bedrock (above −10 • C), to the increased fragmentation of the deepest ice core sections (short fractured cores of ∼20 cm extracted using ethanol as a drilling liquid), or to a change in ice crystal structure at high depths (Pol et al., 2010;Durand et al., 2010). Bender et al. (1995) showed anomalously low O 2 /N 2 and Ar/N 2 ratios in air extracted from ice cores compared to atmospheric air. This effect was initially attributed to gas loss during coring and storage, with the smallest molecules (O 2 , Ar) being more easily lost than the larger ones (N 2 ). Ikeda-Fukazawa et al. (2005) observed a drift in the O 2 /N 2 ratio correlated with the storage duration. Two mechanisms were proposed to explain this size dependent effect: diffusion through the ice lattice by breaking of hydrogen bonds (Ikeda-Fukazawa et al., 2005) or diffusion through small channels in the ice with a threshold dimension of 3.6Å (i.e. molecules with a diameter larger than 3.6Å, like N 2 , will not escape from the bubbles) (Huber et al., 2006). For the Dome F ice, Kawamura et al. (2007) found that δO 2 /N 2 decreased by 6.6 ‰ per year of storage at −25 • C, and used this relationship to apply a gas loss correction. While the exact storage temperature histories of our samples are less well known, we observe significant shifts in δO 2 /N 2 levels between samples stored 1-2 years at −25 • C (Series 1 and 2) and samples stored at EDC (−50 • C) and maintained at this temperature during transport and storage (Series 3 and 4) as depicted above.

Corrections
In order to remove outliers, we excluded all the measurements at depth levels where the δO 2 /N 2 standard deviation associated with replicates is larger than 3 ‰ (Fig. 1). This rejects less than 16 % of the data (mainly over the deepest part of Series 2, see details in Table 1) and results in a pooled standard deviation of 0.9 ‰. This is very comparable to the pooled standard deviation obtained on the δO 2 /N 2 records of  Fig. 2. Two gas loss corrections for the EDC δO 2 /N 2 record. We propose two gas loss corrections since no accurate storage history documentation is available: "gas correction 1" (curve 1 = red) shifts all Series 1 and 2 data by +6.43 ‰ (the average offset between Series 1 and 2 when compared with Series 3), resulting in a variance of 13 ‰ after outlier rejection; "gas correction 2" (curve 2 = blue) seeks to homogenize Series 1 and 2 with Series 3 and 4 where they overlap.
the Vostok and Dome F ice cores after gas loss correction and removal of 15 to 20 % of outliers (Bender, 2002;Kawamura et al., 2007;Suwa and Bender, 2008a). As discussed above, the δO 2 /N 2 measurements performed on ice stored at −50 • C are not appreciably affected by gas loss so that we keep Series 3 and 4 without any correction. Then, we propose a correction that accounts for a systematic bias in the measurements when ice is stored at −25 • C instead of −50 • C. It is based on the following observations: (a) there is no obvious shift between Series 1 (measured in [2004][2005] and 2 (measured in 2006-2007); (b) after an homogenization of Series 1, 2 and 3 through a linear interpolation every 1 kyr between 380 and 480 ka, we found an average offset between Series 1 and 2 on the one hand and Series 3 on the other hand of 6.43 ‰. We thus decided to shift all the δO 2 /N 2 values of Series 1 and 2 by adding 6.43 ‰. We call this "gas loss correction 1". This correction is subject to discussion. In particular, it leads to a significant decrease in δO 2 /N 2 with time: the mean δO 2 /N 2 level before 480 ka is less depleted than the mean δO 2 /N 2 level after 380 ka (Fig. 2); the variance of the whole record after the rejection of outliers and "gas loss correction 1" is about 13 ‰. We therefore propose an alternate correction, "gas correction 2". This second correction aims to homogenize (1) the mean level of δO 2 /N 2 between Series 2 and Series 4 around 700-750 ka, (1) the mean level of δO 2 /N 2 between Series 1, 2 and Series 3 around 430-480 ka and (3) the mean level of δO 2 /N 2 between Series 1 and Series 3 around 380-430 ka. In order to fulfill requirements (1) and (2), a simple solution is to add 2.5 ‰ to the δO 2 /N 2 of Series 1 and 2 between 430 and 700 ka. Then, in order to fulfill requirement (3) one) is to apply the following correction to Series 1 and 2 between 300 and 430 ka: δO 2 N 2 corr = δO 2 N 2 uncorr + 2.5 − 0.035 × (t − 500) (3) where δO 2 /N 2 corr is the corrected δO 2 /N 2 , δO 2 /N 2 uncorr is the original δO 2 /N 2 measurements from Series 1 and 2 and t is the age of the δO 2 /N 2 data point in ka. This second correction is different from the first one, being larger for the recent time period (300-430 ka) than for the oldest period (430-700 ka). The choice of the correction will thus have an impact on the magnitude of the long-term trend of δO 2 /N 2 . The final variance of the δO 2 /N 2 record after outlier rejection and "gas loss correction 2" is less than 9 ‰, which is comparable with the variance of the Dome F and Vostok δO 2 /N 2 records.

Reconstructed curve from EDC δO 2 /N 2
We now construct a composite EDC δO 2 /N 2 record over the period 300-800 ka as follows: -when Series 1 and 2 overlap with Series 3 and 4, we only keep the measurements from Series 3 and 4.
-we use the two gas loss corrections (outlier correction and gas loss correction) described in the previous paragraph for Series 1 and 2 on the remaining periods.
Because we have two alternative "gas loss corrections", we produce two different composite curves, hereafter curves 1 and 2 (Fig. 2). This provides a means of estimating the effect of our subjective gas loss corrections on the final δO 2 /N 2 record. It should be noted that with such corrections, we do not correct the scattering of δO 2 /N 2 data probably due to small-scale gas loss observed in Series 1 and 2. This scattering is especially visible on the corrected curves between 550 and 600 ka. An obvious difference between the two composite curves is the temporal evolution of δO 2 /N 2 with time. While curve 1 shows a long-term decrease of δO 2 /N 2 of −0.014 ‰ per kyr, its value is only of −0.008 ‰ per kyr for curve 2 (Fig. 2). The general long-term evolution of δO 2 /N 2 with time is robust with respect to gas loss correction independent of our empirically derived gas loss corrections. This evolution is due to the fact that the average δO 2 /N 2 is higher in Series 4 than in Series 3, and neither of these series have been corrected since they should not be significantly affected by gas loss since they were stored at −50 • C.
For Vostok, a decrease of −0.013 ‰ per kyr was observed between 150 and 400 ka (Bender, 2002;Suwa and Bender, 2008a), while Dome F data  show a smaller negative temporal trend (decrease of −0.006 ‰ per kyr). Given that various gas loss corrections are applied to these different data sets, we cannot assess the origin of this trend, i.e. natural long-term δO 2 /N 2 variability, a gas loss effect, or a pore close-off effect.   Fig. 3. Comparison between the two composite EDC δO 2 /N 2 curves and existing δO 2 /N 2 records from Dome F  and Vostok (Bender et al., 2002;Suwa and Bender, 2008a). Because of the different time periods covered by the different ice cores, the comparison is limited to the period 300-340 ka for Dome F and EDC and 300-410 ka for Vostok and EDC. Figure 3 compares the two composite curves obtained from our EDC δO 2 /N 2 data and the previous δO 2 /N 2 records from Dome F  and Vostok (Bender et al., 2002;Suwa and Bender, 2008a). Because of the different time periods covered by the different ice cores, the comparison is restricted to the period 300-340 ka for Dome F and EDC and 300-410 ka for Vostok and EDC. A broad agreement is found with the exception of two features. First, the large peak observed around 400 ka on the EDC record appears as a double peak in the Vostok record. We have confidence in the quality and resolution of our measurements over this period because they were performed on ice stored at −50 • C with a pooled standard deviation of 0.32 ‰ and a mean age resolution of 1.5 kyr. Second, the Dome F data are on average less depleted than the Vostok and EDC data. This could be due to differences in the gas loss effect resulting from different stress experienced by each core after coring. However, we have applied our empirical gas loss corrections to the EDC data over the period 300-340 ka, so a future comparison on samples minimally affected by gas loss would be more useful to evaluating such an effect.

Spectral analysis
The initial δO 2 /N 2 data set on the EDC3 age scale over the period 300 to 800 ka is associated with a minimum, average, and maximum sampling step of 1, 2.5 and 6 kyr (the latter only in one extreme case), respectively. The data are first interpolated at a constant time step. The Multi-Taper Method, producing a spectrum in amplitude, is then used to identify the major spectral components of the δO 2 /N 2 record. We have checked that the results are robust with respect to the spectral analysis method used, as well as with respect to the resampling. Indeed, the same spectral components were obtained using the Blackman Tukey or Classical FFT (Fast Fourier transform) periodogram methods. Reinterpolating the data at steps of 2 or 3 kyr does not yield significantly different results. These analyses were performed with Analyseries software (Paillard et al., 1996).
For the two corrected curves, we observe the same significant frequency peaks (Fig. 4a). Two large peaks coincide with the frequencies of precession (periods of 23 kyr and 19 kyr). A spectral peak at 41 kyr is associated with obliquity (less prominent with the Blackman-Tukey method). A secondary peak is identified at 28 kyr but its detection depends in the spectral analysis method.
Our results can be compared to spectral analysis of the δO 2 /N 2 records of Vostok over the period 150-400 ka (Bender, 2002) and of Dome F between 82 and 360 ka . They all show the same patterns, a large peak corresponding to a period of 23 kyr and a smaller one corresponding to a period of 41 kyr. We note that neither Vostok nor Dome F δO 2 /N 2 records exhibit a shoulder at 19 kyr.
As already demonstrated in previous studies, the δO 2 /N 2 power spectrum resembles that of local insolation, more precisely the insolation received the 21 December at the Dome C site, with the dominance of precession and obliquity (Fig. 4b). It should be noted that the 19-kyr peak, corresponding to precession frequency, is present both in the spectrum of δO 2 /N 2 from our record between 300 and 800 ka and in local 21 December or monthly mean December insolation spectra over the same period. In contrast, it is less clearly imprinted in both the spectrum of δO 2 /N 2 and in the summer insolation spectrum over the last 400 ka (Bender, 2002;Suwa and Bender, 2008a;Kawamura et al., 2007).

Impact of data filtering
Filtering the δO 2 /N 2 record reveals the strong correlation with orbital forcing. Following previous studies Suwa and Bender, 2008a), we performed a data re-sampling with a step of 1 kyr, and the resampled series were band pass filtered based on a fast Fourrier transform (FFT) using a piecewise linear window with sharp slopes at the edges (<10 −5 kyr −1 ) (Analyseries software; Paillard et al., 1996). In order to study the link between δO 2 /N 2 and local summer insolation, it is essential that the precession and obliquity frequencies be preserved. Therefore, we chose two different ranges of filtering frequencies corresponding to the following periods: 15-100 kyr and 15-60 kyr (Fig. 5). We applied these filters to both composite curves, and found results similar to the digital filter of Kawamura et al. (2007) (Fig. 5). Like the digital filter described in  based on finite duration impulse response, these filters do not significantly affect the position of the peaks of the local insolation curves (position of peaks are not affected by more than 0.15 kyr). However, applying these filters does slightly influence the timing of the EDC δO 2 /N 2 extrema (Fig. 5), which can have important consequences when matching δO 2 /N 2 record with insolation curves.
The time-delays between the filtered δO 2 /N 2 records using both different filtering ranges have been quantified using the cross-wavelet transform technique (Mallat, 1998;Torrence and Compo, 1999) as follows. The cross-wavelet spectrum of two series, computed from the continuous wavelet transform of each series, provides an estimate of the local phase difference between the two series for each point of the timefrequency space. Its integration over a frequency interval allows the computation of the instantaneous time lag between the two series in the corresponding frequency band (Mélice and Servain, 2003). With this method, the time-delays between the filtered δO 2 /N 2 obtained with the different filtering www.clim-past.net/8/191/2012/ Clim. Past, 8, 191-203, 2012 ranges is of a few centuries (Fig. 5), with the exception of the period 380-450 ka where the shifts can reach 1 kyr. There are two main causes for the time shifts detected by the different filtering methods: the resolution of the data and the ratio between signal and noise. Tests with resolution ranging from 1 to 5 kyr have shown that peaks can be shifted by a maximum of ±300 yr over the period 300 to 450 ka. Monte-Carlo tests on white noise added to our composite δO 2 /N 2 curve have produced peak displacements of up to ±1.3 kyr (2 σ ), in particular over the period from 300 to 400 ka. Shifts in peak position from filtering are largest over the period 340-360 ka (small signal and poor resolution), 360-450 ka (small signal) and between 610 and 680 ka (poor resolution). In the following, we consider the δO 2 /N 2 record filtered between 15 and 100 kyr. Because of the time shift in the δO 2 /N 2 extrema discussed above, their position is associated with an uncertainty of 1 kyr between 380 and 450 ka, and 0.5 kyr elsewhere.
In addition to the two effects quantified above, it should be noted that the short-term variations in δO 2 /N 2 due to smallscale gas loss variations can have an effect on the final, filtered curve if the resolution is too low. We should thus consider cautiously the δO 2 /N 2 profile in the low resolution period where the isolated data points may be more affected by this gas loss effect than by the insolation signal. To get a qualitative sense of the validity of the filtered curve, we compare it directly with the original data (Fig. 5). This comparison reveals two periods, i.e. between 610 and 680 ka and between 340 and 360 ka. During these intervals, relatively sparse δO 2 /N 2 coverage and small δO 2 /N 2 variations make the identification and correlation of extrema with the filtered curve ambiguous or impossible. This prevents us from using these data for assessing the quality of the EDC3 chronology.

Testing EDC3 using δO 2 /N 2 and local insolation
The previous studies using δO 2 /N 2 in Antarctica (Bender, 2002;Kawamura et al., 2007;Suwa and Bender, 2008a) have compared the δO 2 /N 2 records over the last 400 ka with 21 December insolation or monthly (December) mean insolation at the latitude of each ice core, which display only minor differences. Based on the assumption that the phase lag between δO 2 /N 2 and 21 December insolation is nil as suggested by the Vostok data (Bender, 2002), the new dating of Vostok and Dome F were constructed by matching the peaks of the filtered δO 2 /N 2 curve and of either December (Suwa and Bender, 2008a) or 21 December insolation . In the northern hemisphere, for the Greenland GISP2 ice core, the dating was based on the local summer insolation (June) (Suwa and Bender, 2008b).
To explore the link between our EDC δO 2 /N 2 and local insolation, we have used the 21 December insolation at 75 • S computed with Analyseries software (Paillard et al., 1996) using astronomical solution inputs from Laskar (2004). Effect of band pass filtering of the δO 2 /N 2 record. Bottom panel: composite δO 2 /N 2 data (curve 2, dark grey). Resampled (1 kyr) and filtered δO 2 /N 2 signal with "gas loss correction 2" (i.e. curve 2) for frequencies corresponding to 15-100 kyr (blue) and 15-60 kyr (red). The effect of the Kaiser window filter described in Kawamura et al. (2007) is displayed in green. Top panel: time delay between the filtered curves at 15-100 kyr and at 15-60 kyr (black). The green curve indicates the time delay between the filtered curve at 15-100 kyr and the one using the Kaiser window filter of Kawamura et al. (2007). The grey rectangles indicate the periods when the resolution of the δO 2 /N 2 signal is too low (>3 kyr), limiting the validity of the filtering.
Values were computed with a time step of 1 kyr and then filtered to preserve only the periodicities between 15 and 100 kyr (Fig. 6). The comparison between our δO 2 /N 2 filtered data (curves 1 and 2) and the insolation target was first explored through a correlation calculation enabling a relative temporal shifting between the series (cross correlation function of Analyseries;Paillard, 1996). Considering the whole record, we got a maximum R 2 = 0.51 between our filtered δO 2 /N 2 curve and a mean temporal shift of 2 kyr between the δO 2 /N 2 and insolation curves. To further study this temporal shift, we completed this correlation analysis with a time delay analysis using again the cross wavelet transform technique (Fig. 6). We note several features of this phase analysis. First, we confirmed that the average time delay is 2 kyr between δO 2 /N 2 and insolation 21 December. Second, we observed that the time delays exhibit four maxima at ∼450, ∼550, ∼650 and ∼750 ka. Note that this time delay analysis does not depend on the gas loss corrections.
The variations of the time delays between the EDC δO 2 /N 2 curve and the insolation curve could have several origins/implications: 1. The target curve for δO 2 /N 2 record over 300-800 ka at Dome C should be insolation at another date than 21 December. This is suggested by the significant lag observed at Dome C between the maximum of summer insolation 21 December and the maximum of surface temperature (see Appendix). -Time delay between δO 2 /N 2 (curve 2) and local summer insolation curves (21 December insolation at 75 • S). The original series were band pass filtered (between 15 and 100 kyr) before the computation of the time delay.

Clim. Past
-δO 2 /N 2 curve (curve 2) from the EDC ice core (light blue: raw data; dark blue: after 1 kyr re-sampling and band pass filtering between 15 and 100 kyr). Note the reversed vertical axes.
The time periods highlighted in yellow correspond to significant changes in the time delay of δO 2 /N 2 vs. insolation. Time periods with low eccentricity are highlighted in grey.

2.
A systematic bias of the EDC3 age scale toward ages that are too old over the 300-800 ka period.
If the problem is in the EDC3 age scale, the correlation and time delays between δO 2 /N 2 and insolation variations enable an independent estimate of the EDC3 age scale uncertainty. Our data suggest that the EDC3 dating is correct within ±2.5 kyr (i.e. 2 kyr due to the average time delay plus 0.5 kyr due to uncertainty in the filtering) over the period 300-800 ka, with the exception of four intervals marked by larger uncertainties. During the periods 390-460 ka, around 550 ka, around 650 ka (we note a lower confidence over this period, as shown in Fig. 5) and around 750 ka, the EDC3 age scale uncertainty could be as high as ±5 kyr (i.e. 4 kyr due to the time delay plus up to 1 kyr due to uncertainty in the filtering method). Such a large uncertainty is still within the uncertainty range of the published EDC3 timescale (±6 kyr) .
We now examine the opportunity to improve the EDC3 age scale by tuning our δO 2 /N 2 record on the insolation curve on 21 December, as has already been done for the Vostok and Dome F ice cores with peak-to-peak correspondence. Such systematic peak-to-peak correspondence is sometimes difficult to identify, in particular during periods with low eccentricity. A first example can be seen over the low eccentricity period between 390 and 460 ka, when the insolation curves display two small peaks or shoulders at 405 and 424 ka (Fig. 6). Neither of these secondary peaks is clearly identifiable in our filtered δO 2 /N 2 signal nor in the original δO 2 /N 2 record. This mismatch cannot be attributed to a deficient quality of our δO 2 /N 2 record since the measurements were performed on ice stored at −50 • C. Another reason may be that small variations of insolation during this period do not have significant impact on the processes controlling δO 2 /N 2 . Whatever the causes of such mismatch, the time difference between the two insolation minima on each side of the small peaks reaches up to 20 kyr (for the 405 ka peak). This difference will lead to a tuning uncertainty of up to ±10 kyr over this period, if we match the peaks of the δO 2 /N 2 record with the mid-peaks of the insolation target curve as is classically done. A second example is the minimum of insolation at 750 ka: here the number of peaks is similar between the δO 2 /N 2 and the insolation curves. Still, an unambiguous identification is difficult, as shown by a time delay of 4-5 kyr. In turn, caution should be taken, at least for the EDC record, when tuning δO 2 /N 2 variations to the summer insolation curve over periods of low eccentricity for which high precision (i.e. ice stored at −50 • C) and high resolution δO 2 /N 2 data are needed in any case.

Conclusion and perspectives
We have presented the first record of δO 2 /N 2 over the EDC ice core covering the period between 306 and 796 ka. Many samples were stored at −25 • C for 1 year or more before their analysis, such that raw δO 2 /N 2 measurements are strongly affected by gas loss fractionation. Using high precision δO 2 /N 2 measurements performed on similar depths on EDC samples carefully kept frozen at −50 • C, we were able to propose two gas loss corrections to build composite δO 2 /N 2 curves. Using one or another gas loss correction has no significant influence on the orbital chronology issue. However, the band pass filtering method on our δO 2 /N 2 record can lead to an uncertainty of the order of 1 kyr.
The frequency spectrum of EDC δO 2 /N 2 composite curves and of local insolation of 21 December 75 • S are very similar over the period 300-800 ka, as previously observed for other ice core δO 2 /N 2 records over the 0-400 ka period. Following previous studies performed on the Vostok and Dome F ice cores over the last 400 ka, we have explored the added value of the δO 2 /N 2 signal to test the EDC3 age scale over the period 300-800 ka. In our case, the time correspondence of δO 2 /N 2 with 21 December insolation is not so obvious because there is a mean time delay of 2 kyr between our filtered δO 2 /N 2 record and the 21 December local insolation. Moreover, we have shown that for low eccentricity time periods, it remains a challenge to identify unambiguously peak-to-peak correspondence between δO 2 /N 2 and insolation. These two effects result in a large uncertainty (more than 10 kyr locally) in the determination of a new chronology, which prevents us from using the current δO 2 /N 2 record to produce a new EDC age scale. Even if we call for cautiousness in the use of δO 2 /N 2 as an unambiguous dating tool and if this uncertainty prevents us from using our δO 2 /N 2 constraints for building a new EDC age scale, we can still use the comparison between δO 2 /N 2 and the local 21 December insolation to test the current EDC3 chronology. First, we show that over the major part of the 300-800 ka period, EDC3 is correct within the published uncertainty (6 kyr). We identify, however, several specific periods where the shift between δO 2 /N 2 record and the local 21 December insolation signal shows strong variations. These anomalies are observed during periods of low eccentricity and suggest that δO 2 /N 2 cannot be used as dating constraints during minima of eccentricity, or that the EDC3 age scale should be revised over the following periods: 360-450 ka and 720-760 ka.
In order to improve the dating of the oldest Antarctic records (EDC and Dome F), it would be valuable to produce high-accuracy records of total air content, δO 2 /N 2 and δ 18 O atm over the period 300-800 ka with a focus on the period 350 to 390 ka to improve our constraint on the length of MIS 11. The consistency of the various records in two ice cores would allow us to establish common and more accurate age scales. Our long δO 2 /N 2 record further reveals that the variance of the signal is preserved back to 800 ka even close to bedrock, and that, if stored at −50 • C, deep and old ice can provide accurate δO 2 /N 2 records. This has strong implications for the IPICS (International Partnerships in Ice Core Sciences) oldest ice challenge, with the target to obtain Antarctic ice cores spanning more than one million years and to date them.

Uncomplete understanding of the link between δO 2 /N 2 and local insolation
In this section we explore the possibility that δO 2 /N 2 in the EDC ice core is not solely dependent on 21 December or December insolation (which have very similar spectral properties).
A clear mechanism linking 21 December insolation and δO 2 /N 2 in ice core is not understood. A link has been suggested through seasonal maximum in surface temperature based on the strong link evidenced between 21 December Fig. A1. Evolution of the daily insolation (black), air temperature (dotted green), snow temperature at 10 cm depth (blue) and temperature at 50 cm depth (pink) at the Dome C station between the year 2006 and 2008. The temperature measurements of the snow at Dome C are part of a more complete system which records temperatures every hour at 40 levels from the surface down to 21 m since November 2006. Temperatures were measured with 100 ohm Platinum Resistance Temperature (PRT) detectors (IEC751 1/10 DIN). A more detailed description of the system and an open data access is available on the OSUG website (http://www.obs.ujf-grenoble.fr). The air temperature sensor is housed in a naturally aspirated, multi-plate radiation shield (Young 41003), and the measurement was performed at 1 m height. The lag between the maximum of snow temperatures measured at 10 cm and 50 cm is due to the diffusion time of the annual wave temperature between these two levels. The obvious decrease of the diurnal amplitude variations of T 10 cm during the observed period is due to the snow accumulation at the surface. This accumulation implies also an increase of the lag between air and snow temperature measurements.
insolation and the seasonal maximum of Dome F surface temperature (with no lag) . However, the timing is different at Dome C with a lag of 15-20 days between the maximum of insolation (21 December) and the surface temperature maximum (Fig. A1).
Recent progresses have been done to improve our understanding of the mechanisms linking surface temperature, snow metamorphism and δO 2 /N 2 . On the one hand, Hütterli et al. (2010) developed a simple model based on the concept that the evolution of temperature gradient metamorphism affects the snow structure in response to local insolation and suggested that significant shifts by several kyr can exist between snow metamorphism and 21 December local insolation. On the other hand, Fujita et al. (2009) measured physical properties of the Dome F firn and studied the density layers evolution during firnification. Based on these results, they proposed a model linking firn properties with conditions for the gas transport processes Clim. Past, 8, 191-203, 2012 www.clim-past.net/8/191/2012/ near the bottom of firn. This model explains how stronger insolation can lead to bulk ice with a lower δO 2 /N 2 . To conclude, the assumption that δO 2 /N 2 is systematically linked with the 21 December local insolation whatever the climatic conditions at the site and the orbital context deserves further scrutiny.