In this study we use data from six individual records obtained in Princess Elizabeth Land (Fig. 1, Table 1).

The designation “105 km” (67.433∘ S and 93.383∘ E, time interval 1757–1987) refers to a 727 m ice core drilled in 1988 about 105 km inland from Mirny Station by specialists from the St Petersburg Mining Institute. The isotopic content was measured in the late 1980s at the Laboratoire des Sciences du Climat et de l'Environnement (LSCE) with a resolution of 1 m. In 2013, the upper 109 m of the core were remeasured at the Climate and Environmental Research Laboratory (CERL), with a depth resolution of 5 cm. This core is the only one in which the accumulation rate allows the annual layers to be preserved in the snow thickness, so the core was dated by layer counting. The initial dating was then adjusted using the reference horizon of the 1816 Tambora volcanic eruption, identified from electrical conductivity measurements (ECMs) (Vladimirova and Ekaykin, 2014). As a result, a record of the annual accumulation rate is available.

The designation “400 km” (69.95∘ S and 95.617∘ E, 1254–1987) refers to an ice core drilled in 1988 at the 400th kilometre from Mirny Station, down to 150 m depth. Isotopic measurements were performed at LSCE on 1 m samples. The core was dated according to the simple Nye depth–age model, taking into account the average accumulation rate at the drilling site (Lipenkov et al., 1998) and the density profile of the core. The uncertainty of the dating, estimated with the Nye model, mainly comes from the error of the accumulation rate estimate and is evaluated at about 10 %. As a result, no record of the annual accumulation rate is available.

The designation “VRS 2013” (78.467∘ S and 106.84∘ E, 1654–2010) refers to a stack of 15 individual isotopic records from snow pits and shallow cores recovered in the vicinity of Vostok Station (Ekaykin et al., 2014). The data on temporal variability of snow accumulation rate is also available for this site.

The designation “NVFL-1” (77.11∘ S and 95.072∘ E, 1711–1944) refers to an 18.3 m firn core drilled from the bottom of a 2.5 m snow pit in 2008 close to Dome B. The chronology was established using the firn density data and the 1816 Tambora volcano ECM peak as a reference horizon.

The designation “NVFL-3” (76.405∘ S and 102.167∘ E, 1978–2009) refers to a 3.1 m snow pit dug in 2010 in the northern part of subglacial Lake Vostok. It is dated based on snow stratigraphy and identification of the 1993 Pinatubo volcano peak in SO42- vertical profile. Chemical measurements were performed at the Limnological Institute of the Russian Academy of Sciences, Irkutsk, Russia.

The designation “PV-10” (72.805∘ S and 79.934∘ E, 1976–2009) refers to a 7.55 m firn core drilled in 2010 about 400 km inland from Progress Station. It was dated using firn density data and takes into account the ECM peak associated with the 1993 deposition from the Pinatubo eruption.

We estimated the dating uncertainty by comparing age calculated using only firn density data and average snow accumulation rate for a given site with the age of the reference age markers and came to the conclusion that the age errors do not exceed 10 %. For the reference years (1816 and 1993, where we have absolute dating), the error approaches 0. The largest error is expected for the 400 km series, where we do not have reference age markers. However, if we use the prominent 1840 cold event (see Sect. 3.3) observed in all records as a marker, then we may estimate a relative dating error for this series as < 6 %.

We also use the accumulation data from the site “200 km” (Fig. 1), spanning the period 1640–1987, as published in Ekaykin et al. (2000). The accumulation values from sites “150 km” and 400 km are corrected both for layer thinning with depth and for the advection of ice from upstream of the glacier to account for the spatial gradient of the snow accumulation rate.

Figure 2 displays the individual δD time series from all six sites. Differences between mean values reflect well-known differences in isotopic distillation along a gradient of inland elevation (e.g. Masson-Delmotte et al., 2008). In order to investigate temporal variations only, we calculate normalized values for each series using the interval 1757–1944 as a reference period. The short series (NVFL-3 and PV-10) are normalized over the 1978–2009 period, and then the mean and variance of the normalized values are adjusted to those of the long series for the corresponding period of time, in order to avoid an overestimated contribution of the short records in the stacked series.

We then apply a rectangular-shaped low-pass filter to cut off the variability with periodicities shorter than 27 years (i.e. frequencies > 0.037). (All spectral analyses and filtering are performed with the use of Analyseries software; Paillard et al., 1996.) We decide to do this because one single record in inland Antarctica cannot provide reliable climatic information on a short-term timescale, due to a very low signal-to-nose ratio (Ekaykin et al., 2014) and non-temperature effects on isotopes in precipitation including post-depositional alterations. Moreover, the latter study (Ekaykin et al., 2014) also highlighted multi-decadal climatic variability in this sector of central Antarctica, with a period of 30–50 years.

The normalized and filtered time series are displayed in Fig. 3. Despite some common features, this comparison shows significant discrepancies between individual records. One reason for the mismatches may lie in age scale uncertainties. However, this hypothesis is ruled out by the comparison of individual series around 1816 and 1993 (the dates of firn layers containing Tambora and Pinatubo volcanic eruption debris, denoted by vertical dashed lines in Fig. 3), when the relative dating error tends to 0: observed discrepancies do not arise from chronological uncertainties alone. Alternatively, this mismatch may arise from a significant level of noise even in the filtered series and factors other than the local temperature that control the isotopic composition of precipitation.

In order to isolate the climatic signal from the noise, we constructed a stacked climatic record for the PEL region, hereafter named PEL2016 (grey line in Fig. 3). For a given year, the value of this record consists of the average of the values of individual records available for this year.

A number of research stations have been established in the PEL area, as indicated in Fig. 1. Unfortunately, most of them have very short (if any) meteorological records. Relatively long records are available only for five stations: the Australian station Davis (1957–1964 and 1969–2015), the Chinese station Zhongshan (1989–2015), and the Russian stations Progress (1989, 1991 and 2003–2015), Mirny (1956–2015) and Vostok (1958–2015 with gaps in 1962, 1994, 1996 and 2003). The monthly data were downloaded from https://legacy.bas.ac.uk/met/READER/ (Turner et al., 2004), and then the annual means were calculated.

The correlation between Progress, Zhongshan and Davis annual mean temperature datasets, located very close to each other, is 0.96–0.98 (note that only statistically significant correlation coefficients with a confidence level > 95 % are reported in the paper, unless otherwise mentioned). Hereafter, we only use data from the station with the longest record (Davis).

We also use data from the automatic weather station (AWS) LGB59 located at the slope of the Antarctic ice sheet inland from Progress Station (Fig. 1), available for the period 1994–1999, as well as surface air temperature data from Casey and Mawson.

In order to investigate possible relationships between PEL climate multi-decadal variations and large-scale modes of variability, we use data on the indices of the Antarctic Oscillation (AAO), the Interdecadal Pacific Oscillation (IPO) and the Indian Ocean Dipole (IOD).

The AAO index, also known as the SAM (Southern Annular Mode), is defined as a mean latitudinal difference of sea level pressure at 40 and 65∘ S and is considered as a prevailing mode of atmospheric circulation in the Southern Hemisphere (SH) representing about 35 % of the extratropical SH climate variability (Marshall, 2003). The monthly AAO index is available from NOAA (http://www.cpc.ncep.noaa.gov/products/precip/CWlink/daily_ao_index/aao/monthly.aao.index.b79.current.ascii.table; since 1979) and the British Antarctic Survey (http://www.antarctica.ac.uk/met/gjma/sam.html; since 1957, although data for the 1957–1978 period is considered to be less robust).

IPO is defined as a sea surface temperature (SST) anomaly over the Pacific Ocean. The positive phase of IPO is characterized by a relatively warm central and eastern tropical Pacific and a relatively cold north-western and south-western Pacific (Henley et al., 2015; Dong and Dai, 2015). The IPO index is closely related to PDO (Pacific Decadal Oscillation), but PDO better characterizes the Northern Pacific, while IPO is better applicable to the whole Pacific region. We used IPO data because in the previous study we found a teleconnection between the climate variability in the central Antarctic and the tropical Pacific (Ekaykin et al., 2014). The data on the IPO index since 1870 are available here: http://www.esrl.noaa.gov/psd/data/timeseries/IPOTPI/.

The IOD is characterized by the Dipole Mode index (DMI), which is defined as the SST gradient between the western equatorial Indian Ocean (50–70∘ E and 10∘ S–10∘ N) and the south-eastern equatorial Indian Ocean (90–110∘ E and 10∘ S–0∘ N). Thus, IOD is an analogue of SOI (Southern Oscillation index), but for the Indian Ocean. The data on the DMI index since 1870 can be found at http://www.jamstec.go.jp/frsgc/research/d1/iod/iod/dipole_mode_index.html.

Here, we first consider the variability of surface air temperature recorded at the meteorological stations in Princess Elizabeth Land to assess whether the studied sector is characterized by uniform climate variability and to provide a reference regional temperature record for comparison with the δD stacked record.

Correlation coefficients between annual mean surface air temperature data at Vostok, Mirny and Davis vary between 0.6 and 0.9 (Table 2). Correlation coefficients between the automatic weather station LGB59 (located between Davis and Vostok, Fig. 1) and these three stations vary between 0.86 and 0.96. Despite the short record at LGB59, they are also significant at the 95 % confidence level. These results demonstrate that the region encompassed between these three stations has experienced similar climatic variability. This is further confirmed by a cluster analysis of surface air temperature data from 12 Antarctic stations (see Fig. S1 in the Supplement), showing that Vostok, Mirny, Casey, Mawson and Davis data form a single cluster in terms of climatic variability.

Interestingly, the correlation coefficient between Mirny and Vostok data is significantly weaker in 1958–1976 (R=0.53) than in 1976–2015 (R=0.74). This suggests that, before the so-called “1976 climate shift” (Giese et al., 2002) Vostok experienced a higher influence from the Pacific sector of the Southern Ocean (Ekaykin et al., 2014) not encompassed at Mirny. Indeed, the correlation coefficient between temperature data from Vostok and McMurdo Station (located in the Pacific sector) was higher before the 1976 shift (R=0.46) than after 1976 (R=0.35).

During the whole period of instrumental observations, the strongest relationships observed for temperature at Vostok were with temperature data at Mirny and Mawson coastal stations from the Indian Ocean sector and more precisely the sector between the Davis Sea and the Cooperation Sea.

As a result, Fig. 4a shows the average temperature anomaly from Vostok, Mirny and Davis stations. Hereafter, we use this stacked temperature record as an estimate of the temperature anomaly for the whole PEL sector.

We then compare the low-frequency variations in these various temperature records, using the 27-year low-pass filter (Fig. S2). Both Vostok and Mirny demonstrate a quasiperiodical variability with a period of about 30 years and maxima in the late 1970s and the late 2000s and demonstrate a very high similarity at low frequency. While Davis data have the same periodicity, their maxima are shifted to the early 1970s and early 2000s. If we consider other Antarctic stations, we see complex behaviour of air temperature in different sectors of Antarctica: most stations also show a 30-year cycle, but with a significant phase shift relative to the PEL region.

In the Indian Ocean sector, temperature peaks appear increasingly delayed when moving from west to east. For example, the first maximum occurred late in the 1960s at Mawson, early in the 1970s at Davis, in the second half of the 1970s at Mirny and late in the 1970s at Casey. This feature may reflect a low-frequency component of the Antarctic Circumpolar Wave (Carril and Navarra, 2001).

With respect to multi-decadal trends, contrasted patterns emerge: some stations (Esperanza, Novolazarevskaya, Davis, Vostok, Mirny, McMurdo) display a warming trend, while a cooling trend emerges at Halley or Dumont d'Urville (Fig. S2).

This comparison of instrumental temperature records highlights different patterns of multi-decadal variability across different sectors of Antarctica, which is important for interpreting paleoclimate records and for combining various proxy records for temperature reconstructions (Jones et al., 2016). Our analysis nevertheless demonstrates coherency within Princess Elizabeth Land, where we will use the stacked temperature record from Vostok, Mirny and Davis as a reference regional signal (hereafter named PEL temperature anomaly) for the calibration of δD records.

Here we compare the PEL temperature anomaly with indices that characterize climatic variability in the Southern Hemisphere. First, as expected, a very strong negative relationship with the AAO index (r=-0.68) is observed in 1979–2015 (Fig. 4b). The Antarctic Oscillation is the predominant mode of climatic variability in Antarctica: a strong AAO index reflects a larger pressure gradient between low and high latitudes, associated with a more zonal circulation around Antarctica and colder conditions in East Antarctica. We note that no correlation between PEL and AAO is identified prior to 1979, which could be an artefact due to poor estimates of AAO before 1979, when few instrumental records were assimilated in atmospheric reanalyses.

The correlation coefficient of PEL temperature anomaly with the IPO index is weak (Fig. 4c), but the residuals of the PEL temperature regression with AAO are negatively correlated with the IPO index (r=-0.47).

A multiple linear regression approach leads to the conclusion that combined variations in AAO and IPO explain 59 % of the temperature variance on an inter-annual scale. While such teleconnection between Pacific and central Antarctic climate have previously been reported from Vostok data (Ekaykin et al., 2014), the underlying mechanism is not known. Finally, no significant correlation was identified between PEL temperature and the IOD index (Fig. 4d).

However, different results emerge when considering the low-pass-filtered time series. On multi-decadal timescales, a strong positive correlation (r=0.8, significant with a 0.06 confidence level) relates PEL temperature and the AAO (Fig. 4a and b), and a very strong positive correlation appears between PEL temperature and the IOD index (r=0.93, p<0.05). We suggest that the Indian Ocean Dipole affects the Antarctic climate through a modulation of cyclonic activity. This is indirectly confirmed by a negative correlation (r=-0.56) between the IOD index and the pressure anomaly at Mirny and Davis (not shown). The positive relationship between AAO and temperature in the low-frequency band could then be an “induced correlation” caused by a very strong positive correlation between AAO and IOD (r=0.8–0.9) on these timescales.

The stacked δD record (built from low-pass-filtered individual records) is now compared with the filtered PEL temperature composite (Fig. 4a). We observe a positive correlation with r=0.66. Although the length of the series is 52 years, the number of degrees of freedom is only 4 due to the 27-year filtering. The uncertainty of the correlation is ±0.4, so it is statistically insignificant (p=0.17).

This invokes a discussion of the factors that may disturb the correlation between the local air temperature and the stable water isotopic composition of precipitation in Antarctica (Jouzel et al., 2003).

Firstly, isotopic composition of precipitation is not a function of local air temperature but of the temperature difference between the evaporation area and the condensation site, which defines the degree of heavy water molecule distillation from an air mass. The study of the moisture origin for this sector of Antarctica (Sodemann and Stohl, 2009) demonstrates that different parts of the PEL differ in their moisture origin. Coastal areas receive moisture from higher latitudes (46–52∘ S) and from more western longitudes (0–40∘ E) than inland areas (34–42∘ S and 40–90∘ E). It means that even if our sector is climatically uniform, as was shown above, the temporal variability of the precipitation isotopic content may differ in the different parts of the sector due to varying moisture origin.

Secondly, we should define which temperature is actually recorded in the isotopic composition of precipitation. For central Antarctica, where much (or most) of the precipitation is “diamond dust” from a clear sky (Ekaykin, 2003), the effective condensation temperature is conventionally considered equal to the temperature on the top of the inversion layer. But this is definitely not true for the coastal areas, where most precipitation falls from the clouds. Thus, the difference between near-surface and condensation temperature may vary in space and time.

Thirdly, the precipitation seasonality is another factor that may change the relationship between the air temperature and stable isotope content in precipitation. At Vostok the amount of precipitation is evenly distributed throughout the year (Ekaykin, 2003), so the snow isotopic content corresponds well to the mean annual air temperature, but we do not have robust information either about the other parts of the PEL or about the seasonality changes in the past.

Yet we believe that the main factor affecting the isotope–temperature relationship is “stratigraphic noise”. Indeed, even when we study the ice cores obtained at a short distance from one another (Ekaykin et al., 2014), the correlation between the individual isotopic records is still small, though the climatic conditions are the same.

This is why we argue that constructing a stacked isotopic record is an optimal way to reduce the amount of noise in the series and to highlight the variability that is common for the whole studied region, provided that the region is climatically uniform.

Despite the statistically insignificant correlation coefficient, we assume that the stacked δD record is a proxy of surface air temperature in the PEL region (or following Steig et al., 2013, a proxy that “covaries with atmospheric circulation in a manner similar to temperature”). Thus, we estimate the calibration coefficient between these two parameters as a ratio of the standard deviation of the δD composite record to the standard deviation of the PEL low-pass-filtered temperature record, which allows us to assign a temperature scale to the isotopic record. The apparent isotope–temperature gradient, obtained as a standard deviation of isotopic values divided by standard deviations of temperature values, is 13.8 ± 2.5 ‰ ∘C-1 (the uncertainty is due to different standard deviations of isotopic values in individual records). This approach implicitly suggests a perfect correlation between the compared series. If we correct the apparent slope by the observed correlation coefficient, 0.66, it becomes 9 ± 5.4 ‰ ∘C-1. The latter value is still higher than the corresponding slopes observed in other regions of Antarctica (see a review in Stenni et al., 2016) but corresponds nicely to an isotope–condensation temperature slope predicted by a simple isotope model (Salamatin et al., 2004). Actually, low apparent isotope–temperature slopes obtained based on ice core data may be due to a significant amount of noise in the isotopic records, while in our case we removed the noise to a considerable extent by filtering and constructing the stacked record.

The temperature reconstruction is displayed in Fig. 5b as a temperature anomaly relative to the 1980–2009 period. We also show the instrumentally obtained air temperature anomaly in Fig. 5b on the same temperature scale.

Following Ekaykin et al. (2014), who reported a closer relationship between Vostok isotopic data and summer temperature than with the annual mean temperature, we performed additional analyses of relationships between our stacked isotope record and other temperature time series (e.g. monthly or seasonal temperature anomalies), but this does not improve the isotope–temperature correlation.

Despite discrepancies in the individual isotopic records (Fig. 3), a common signal identified in the stacked record leads to several conclusions about PEL climate variability over the past 350 years. During this time interval, regional surface air temperature shows a long-term increasing trend and an overall warming by about 1 ± 0.6 ∘C. Superimposed on this multi-centennial trend, quasiperiodical variability occurs with periods of 30–40 and about 60 years. A colder period is identified in 1750–1860 – i.e. approximately at the same time interval as the Little Ice Age reported in other regions (PAGES 2k Consortium, 2013).

A remarkably cold phase is observed during the 1840s, during which PEL temperature could fall 1.2 ± 0.7 ∘C below that of the present day (defined as the average value of the last 30 years). As seen in Fig. 3, this event is a robust feature, observed in all four individual records available for this time interval. This minimum was also identified in an Antarctic temperature stack record (Schneider et al., 2006) – see Fig. 5d – as well as in an ice core drilled in the Ross Sea sector (Rhodes et al., 2012) and in the isotope record from Ferrigno (coastal Ellsworth Land) (Thomas et al., 2013).

Further studies are needed to understand whether such remarkably cold conditions arise from internal variability or are driven by the response of regional climate to an external perturbation. A possible candidate could be a response to volcanic forcing (Sigl et al., 2015). A moderate event is associated with the eruption of Cosigüina in 1835. According to the inventory of volcanic events recorded in the Vostok firn cores (Osipov et al., 2014), an unknown volcano erupted in 1840; however, the amount of deposited sulfate was about 15 % of that of Tambora, so it is not expected to have a major effect on the climate system. So far, the influence of volcanic forcing on Antarctic climate and the response time remain poorly known. By contrast, recent studies have stressed the delayed response of the North Atlantic Oscillation (Ortega et al., 2015) to major volcanic eruptions as well as their role as pacemakers of bidecadal variability in the North Atlantic (Swingedouw et al., 2015).

The period before 1700 is probably the coldest part of the record, but this is not a robust result as the two records spanning this time interval show somewhat different behaviours (Fig. 3). However, another stack of five East Antarctic cores from PAGES (Past Global Changes) 2k (Fig. 5e) also highlights that the 1690s could have been the coldest decade of the last 350 years. We also compare the PEL2016 record with other Antarctic temperature reconstructions. Schneider et al. (2006) used high-resolution isotopic records from five Antarctic sites (a stack of Law Dome records, Siple Station, a stack of Dronning Maud Land records and two ITASE (International Trans Antarctic Scientific Expedition) sites from West Antarctica). Although this record is not significantly correlated with PEL2016 (r=0.36), we note some common features in both records (warming in the 1820s and 1890s, cold events in the 1840s and 1900s, etc.).

We also investigate the similarities between PEL2016 and the filtered stack normalized isotopic East Antarctic record based on five East Antarctic ice cores (Fig. 5e; data are available in the supplement of PAGES 2k Consortium, 2013). The correlation with PEL2016 is weak (r=0.13) and insignificant, and so is the correlation with the stack from Schneider et al. (2006) (r=0.36).

The main difference between our PEL2016 record and the other isotopic stacked records for the whole of Antarctica (Fig. 5d) and for East Antarctica (Fig. 5e) appears for long-term trends, with a long-term increase in PEL2016 but no similar feature in the other reconstructions. We suggest that contrasted regional long-term trends may disappear in continental-scale reconstructions (see Fig. S2).

Finally, we compare our PEL2016 record with an IOD time series since 1870, also processed with a low-pass filter. The strong correlation coefficient (r=0.79) confirms the close relationship between multi-decadal variations in surface air temperature in this sector of Antarctica and IOD. The Indian Ocean Dipole oscillation appears as the predominant climatic mode affecting multi-decadal climate variability in this part of East Antarctica. While the exact mechanisms underlying this relationship are not known, the IOD is expected to affect the inland Antarctic climate by modulating the cyclonic activity that brings heat and moisture to the Antarctic continent.

We now investigate the low-pass-filtered values of the snow accumulation rate, available at the 105 km, 200 km and Vostok sites (the latter is a stack curve from three deep snow pits), normalized over the period 1952–1981 (Fig. 6). All of them exhibit a negative trend, more prominent for the 200 km series. This result contradicts the stacked Antarctic snow accumulation rate record (Frezzotti et al., 2013) showing an overall increase in the accumulation rate during the last 200 years. Our finding is also not supported by the accurate assessment of average accumulation rate change between successive reference horizons at Vostok, showing a slight but significant increase in snow accumulation rate since 1816 (Ekaykin et al., 2004). Our results, moreover, stress the fact that, during the past few centuries, opposite long-term trends may have occurred in temperature and accumulation. This is counter-intuitive with respect to atmospheric thermodynamics and to the expected covariation of heat and moisture advection towards inland Antarctica. A similar divergence of the centennial trends of snow isotopic composition and accumulation rate was observed by Divine et al. (2009) at the coastal sites of Dronning Maud Land but not at the inland sites (Altnau et al., 2015).

Processes other than snowfall deposition may, however, affect the ice core records. In the vicinity of 105 km, large “transversal” snow dunes have recently been evidenced (Vladimirova and Ekaykin, 2014). Such features may lead to a strong non-climatic variability in the snow accumulation rate at a given point, due to dune propagation effects. Blowing snow events may also have a significant influence on mass balance in the coastal zone of Antarctica (Scarchilli et al., 2010), potentially introducing additional post-deposition noise.

As a result, we are not confident that the datasets reported in Fig. 6 can be interpreted in terms of climate (snowfall) variations. Further work is needed to distinguish the large-scale climate effect (snowfall deposition) from the non-climatic effects potentially associated with post-deposition (wind erosion, dune propagation, etc.).