Climate of the Past Changes in atmospheric variability in a glacial climate and the impacts on proxy data : a model intercomparison

Using four different climate models, we investigate sea level pressure variability in the extratropical North Atlantic in the preindustrial climate (1750 AD) and at the Last Glacial Maximum (LGM, 21 kyrs before present) in order to understand how changes in atmospheric circulation can affect signals recorded in climate proxies. In general, the models exhibit a significant reduction in interannual variance of sea level pressure at the LGM compared to pre-industrial simulations and this reduction is concentrated in winter. For the preindustrial climate, all models feature a similar leading mode of sea level pressure variability that resembles the leading mode of variability in the instrumental record: the North Atlantic Oscillation (NAO). In contrast, the leading mode of sea level pressure variability at the LGM is model dependent, but in each model different from that in the preindustrial climate. In each model, the leading (NAO-like) mode of variability explains a smaller fraction of the variance and also less absolute variance at the LGM than in the preindustrial climate. The models show that the relationship between atmospheric variability and surface climate (temperature and precipitation) variability change in different climates. Results are model-specific, but indicate that proxy signals at the LGM may be misinterpreted if changes in the spatial pattern and seasonality of surface climate variability are not taken into account. Correspondence to: F. S. R. Pausata (francesco.pausata@bjerknes.uib.no)


Introduction
Much of our knowledge about past climates comes from only a few locations, most notably Greenland and Antarctica, because it is difficult to obtain good quality, high resolution proxy records of long duration. It follows that data from single locations have been used to infer climate changes back in time at regional, hemispheric and even global spatial scales (e.g. Dansgaard et al. [1993]; Jouzel et al. [1994]; Shackleton [2001]). For example, the Greenland ice cores' oxygen isotope records have been used to reconstruct temperature as far back as the last interglacial (e.g. Dansgaard et al. [1993]) using the modern climate temperature-isotope relationship. There is awareness of the potential pitfalls of assuming stationarity in the relationship between climate and signal captured by proxies, but in many cases there has been little investigation of the causes of non-stationarity and therefore few proposed solutions. The climate-proxy relationship cannot, however, be assumed to be stationary on climate change time scales. For instance, it has been suggested that changes in the position of the centers of action of the leading modes of climate variability (such as the North Atlantic Oscillation, NAO), shown in several studies (e.g. Christoph et al. [2000]; Raible et al. [2006]), have led to a change in the signal recorded by proxies (Hutterli et al. [2005]).
Climate models can be a useful tool for assessing how internal atmospheric variability may be altered by external forcings, and how these changes may affect what the proxy data record. For example, model simulations suggest that persistent positive anomalies in the NAO index in the 1980s-1990s are linked to increases in greenhouse gas concentrations (Shindell et al. [1999]; Miller et al. [2006]). Past climates offer a wider range of climate states to explore, in addition to the possibility of comparing model simulations with proxy-based observations when and where these are available. Previous studies have shown that during the mid-Holocene (6000 kyrs before present, 6 ka) warm interval, the atmosphere supports variability that has NAO-like characteristics similar to the pre-industrial (PI, 1750 A.D.) period (Gladstone et al. [2005]). Last Glacial Maximum (LGM, 21 ka) simulations permit an exploration of the dominant patterns and seasonality of climate variability during an interval when the atmospheric circulation was substantially perturbed by the presence of large land-based ice sheets and by lower greenhouse gas concentrations. Simulations of the LGM cold climate exhibit substantial differences in both the mean state and variability of the extratropical circulation compared to PI 2 simulations. These differences include: 1) a southward shift of the Pacific and Atlantic storm tracks (Laîné et al. [2008]); 2) a shift (Justino and Peltier [2005]; Peltier and Solheim [2002]) and weakening (Otto-Bliesner et al. [2006]) of the NAO's main centers of action; and 3) a decrease in interannual jet variability and storminess in the Atlantic sector (Li and Battisti [2008]). It is difficult, however, to evaluate how robust the changes in atmospheric variability are and how the relationship between changes in the atmospheric flow patterns and proxy signals might be expected to vary, because each of the aforementioned studies (with the exception of Laîné et al. [2008]) was performed using a single model. We present a model intercomparison of sea level pressure (SLP) variability in the extratropical Northern Hemisphere (20°-90°N) in two fundamentally different climate states, the PI and the LGM.
The aim of this paper is to document how SLP variability and its leading mode might have changed in a glacial climate, and how this change could affect proxy data. Our study attempts to elucidate issues associated with the assumption of a stationary relationship between climate and proxy signals.
The goals are to better understand the spatial scale represented by proxy records and the influence of changed SLP variability on those proxy records. Finally, we try to identify the locations that are able to detect a substantial amount of large-scale variability in both climate states -preferred proxy locations where a straightforward comparison of the glacial and modern states might be possible.
This work is structured as follows: Section 2 gives a description of the coupled models used and the boundary conditions for the PI and LGM climates; Section 3 presents the changes in the magnitude and spatial pattern of SLP variability in the Northern Hemisphere (NH) and the distribution of this variability over the seasonal cycle; Section 4 discusses the influence of atmospheric circulation changes on the signal recorded in proxies. Conclusions are presented in Sect. 5. The horizontal resolution in the atmosphere varies slightly between models, but has a nominal grid spacing of 300 km or T42 (Table 1). Boundary conditions for the two climate states (PI and LGM) follow the protocol established by PMIP2. In the PI simulations, the orbital configuration is set to 1950 A.D. values, the greenhouse gases correspond to 1750 A.D. and vegetation is prescribed to a static model-dependent present day distribution. In the LGM simulations, the orbital configuration is set to 21 ka, greenhouse gas concentrations are lower and result in a 2.8 Wm −2 decrease in radiative forcing (Braconnot et al. [2007]), the static vegetation is as in the PI simulations and the ice sheets are prescribed according to the ICE-5G reconstruction (Peltier [2004]).

Data and Methods
For each model's equilibrium simulation of the LGM and PI climates, 100 years of monthly postspinup SLP, temperature and precipitation data from 20°-90°N are analyzed. The results presented here are based on monthly anomalies from the seasonal cycle. The variability in the resultant time series is concentrated at interannual time scales and is hereafter referred to as interannual variability.
Standard Empirical Orthogonal Function (EOF)/Principal Component (PC) analysis has been used to assess the leading mode of SLP variability in the North Atlantic. All differences discussed in this study are significant at the 1% confidence level, unless otherwise noted. The Atlantic sector is defined as 20°-90°N, 120°W-45°E using the Rocky Mountains and the Urals as boundaries, but the results presented here are not strongly sensitive to the particular definition of the sector.

Variability in the Atmospheric Circulation
This section is divided into three parts. The first describes differences in interannual Northern Hemisphere SLP variability and the leading patterns of North Atlantic SLP variability between LGM and PI simulations. The second describes the distribution of interannual SLP variability and the leading mode of SLP variability over the seasonal cycle. The last discusses reasons for the differences in SLP variability, not only between the LGM and PI simulations from a given model, but also between models for simulations of a given climate. Both the ERA-40 (Uppala et al. [2005]) and NCEP-NCAR (Kistler

Spatial Distribution of SLP variance
In three out of four models, the interannual variability of NH SLP is reduced in the LGM simulations compared to the PI simulations (Fig. 1). The differences in standard deviation (σ) are largest in high latitudes, especially along Greenland's east coast, over the northeastern Pacific Ocean along the coast of Alaska, and over the Barents Sea ( Fig. 1   There is an LGM decrease of between 18 and 28% in the interannual SLP standard deviation associated with the leading mode ( λ 1 σ 2 N A ) for each model except MIROC3.2. For the three consistent models, the LGM decrease results from combined reductions in both interannual SLP variance (σ 2 N A ) 8 and the fraction of variance explained by the leading mode of SLP variability (see Fig. 2 and Table   2). There is no change in the interannual SLP standard deviation explained by the leading mode in MIROC3.2, due to compensating changes in SLP variance and the fraction of variance explained by the leading mode.
The leading mode of North Atlantic LGM SLP variability is qualitatively similar to that in the PI. The spatial pattern in each simulation is an opposing dipole of SLP anomalies that straddle the simulation's climatological-mean low in SLP. Both the centers of action and the related SLP gradient associated with the leading mode are weaker in the LGM simulations from two models (CCSM3 and IPSL) and are comparable to the PI leading mode in the other two models (HadCM3M2 and MIROC3.2). There is no model-to-model agreement on the absolute location of the centers of action, but each model simulates a shift southward/southeastward of the EOF1 pattern at the LGM relative to the PI. In two models (CCSM3 and IPSL), the southern lobe moves southeastward towards the Mediterranean Sea, qualitatively similar to what was seen in the studies of Justino and Peltier [2005] and Peltier and Solheim [2002]. Table 2: LGM-PI changes in interannual variability of SLP in the Northern Hemisphere (σ N H of SLP) and in the North Atlantic (σ N A of SLP); fraction of variance explained by the leading EOF (λ 1 ); amount of raw variability explained by the leading EOF in standard deviation units ( λ 1 σ 2 N A ); and the amplitude of the seasonal cycle of Northern Hemisphere SLP variability (seasonal cycle of σ N H ). Standard deviation changes that are not significant at 1% confidence level are in red.

Seasonal Cycle of Variability
Reductions in interannual SLP variability within the LGM simulations relative to the PI simulations are not evenly distributed throughout the year. Figure  The seasonal cycle of interannual SLP variability is altered in the LGM relative to the PI simulations as a result of the greater wintertime reductions. (Fig. 3, Table 2). Simulations of the LGM climate exhibit not only less interannual variability within each month, but also less of a change in interannual variability across the seasonal cycle. In other words, interannual summer variability is more similar to interannual winter variability during the LGM. Note that the weaker seasonal cycle is also seen further aloft in flow-related fields such as 500-hPa geopotential height (not shown), suggesting that these changes in atmospheric variability are dynamically driven.

Discussion
The LGM simulations show a significant reduction in interannual SLP variance and a weakening of the leading mode of SLP variability relative to the PI simulations. These changes in atmospheric variability must be due to differences in radiative forcing, land/ocean geometry, land ice, and their combined influence on surface properties such as sea surface temperature and sea ice distribution.  Table 3.
First, we address the differences between the LGM and PI climates. LGM (LGMbc+LGMccsmSST) experiments to two sensitivity experiments (PIbc+LGMccsmSST and LGMbc+PIccsmSST) where we mix PI/LGM ocean and PI/LGM external forcing (Table 3). LGMbc+LGMhadSST LGM LGM HadCM3M2 The sensitivity experiment with LGM external forcing (LGMbc+PIccsmSST) shows a mean SLP field that resembles the mean state of the full LGM simulation (LGMbc+LGMccsmSST), even when 12 forced by PI SST/sea ice (Fig. 5). The leading mode of SLP variability in each LGM external forcing experiment is also quite similar to the full LGM simulation (Fig. 6), but the pattern of internannual SLP variability is different in each experiment (Fig. 5). The experiments using PI external forcing (PIbc+LGMccsmSST) have mean SLP distributions that resemble those of the full PI simulation, even when forced by LGM ocean forcing (SST and sea ice) (Fig. 5). The leading mode of SLP variability and its explained variance are also comparable to the full PI simulation (Fig. 6), while the pattern of interannual SLP variance is different in all three experiments (Fig. 5). In all the sensitivity experiments the leading mode variance is not consistently affected by the SST and sea ice (not shown).
13 Figure 5: The mean (contours: 4 hPa interval from 1000 to 1040 hPa; higher values omitted for clarity; bold contour denotes 1016 hPa) and standard deviation (colored shading: hPa) of monthly SLP averaged over all months in the sensitivity experiments. Numbers show the SLP standard deviation area-averaged over the Northern Hemisphere (σ N H in bold) and over the North Atlantic (σ N A in italic). There are a number of ways in which changes in the external forcing can affect atmospheric variability. We focus our discussion on ice sheets and greenhouse gases as these exhibit larger LGM-PI differences than insolation (see Figure 1 in Otto-Bliesner et al. [2006]). The large Laurentide ice sheet covering North America creates an upstream-blocking situation that may be related to a stronger, but less variable, Atlantic jet at the LGM relative to the PI climate (Li and Battisti [2008]; Donohoe and Battisti [2009]). The reduced variance associated with the leading mode of SLP variability (Fig. 4) is broadly consistent with this change in upper-level jet variability; it could also be linked to the lower greenhouse gas concentrations at the LGM, much as the recent increase in NAO variance is thought be linked to external factors such as increases in greenhouse gas concentrations and/or changes in surface properties (Feldstein [2002]). Our results suggest that surface properties (SST and sea ice), are not important for determining the leading mode of SLP variability, but have some influence on the magnitude and pattern of interannual SLP variability. These findings are qualitatively consistent with the study of Kushnir et al. [2002], in which it is demonstrated that atmospheric variability is more affected by internal atmospheric processes than by the extratropical ocean. Other studies suggest that sea ice anomalies can affect atmospheric variability, particularly the phase and amplitude of the NAO (Deser et al. [2000]; Seierstad and Bader [2008]).
Second, we address the fact that while the PMIP2 coupled models show a consistent response to PI forcings, this is not the case for LGM forcings (Figs. 1 and 2). A possible reason is that the PMIP2 models produce similar SST and sea ice distributions in the PI climate but not at the LGM.
A simple test is to impose the ocean forcing (SST/sea ice) produced by one of the other models on CAM3. We choose the HadCM3M2 ocean forcing because it is most dissimilar to those from CCSM3 (experiment LGMbc+LGMhadSST). The resulting SLP field is more like that of the CCSM3-based simulation (LGMbc+LGMccsmSST) than that of the fully coupled HadCM3M2 simulation, both in terms of interannual variability ( Fig. 5 and Fig. 1) and its leading mode of SLP variability ( Fig. 6 and   Fig. 2). This suggests that the atmosphere model CAM3 produces similar SLP variability, regardless of the exact ocean forcing used, consistently with the results of the aforementioned sensitivity studies (PIbc+LGMccsmSST and LGMbc+PIccsmSST). A likely explanation is that the atmosphere models respond differently to the same external forcing because of differences in the PI zonal mean flows simulated by the two coupled models, or perhaps because of differences in physics internal to the two atmospheric models (e.g., differences in the parametrization of gravity wave drag). We can not rule out, however, that the atmosphere model used in the HadCM3M2 is more sensitive to the prescription of the ocean forcing than is the atmospheric model used in the CCSM3, and the differences in the LGM mean state and variability simulated by the HadCM3M2 and the CCSM3 are symptomatic of the differences in the SST and sea ice simulated in these two coupled models.
In summary, our findings suggest that the differences between LGM and PI simulations in CCSM3 are due to the external forcing, with the ocean forcing playing a minor role. Reduced interannual variability at the LGM relative to the PI is a consistent change observed in several of the models, and hence considered a robust model result. The exact amount and spatial characteristics of this variability appear to be model-specific.

16
Changes in the mean and variability of atmospheric circulation, in the leading modes, or in the seasonality of any of these components are interesting from a dynamical standpoint, but they could also have a demonstrable impact on the signal recorded in climate proxies.
The reconstruction of past climate from proxies is based on the idea that natural archives record variations in temperature, precipitation, or some combination of these and other environmental conditions. For simplicity, variability in surface temperature and precipitation are referred to as "surface climate variability". Reconstructions of surface climate variability over recent centuries have been performed using archives such as tree rings (e.g. Glueck and Stockton [2001]), ice cores (e.g. Appenzeller et al. [1998]) and pollen in lake sediments (e.g. Voigt et al. [2008]). A common goal in selecting proxy sites is to find locations where local variability represents larger spatial scales. For example, in the present climate, surface temperature and precipitation at many locations in the Atlantic sector are coherently coupled to the NAO (e.g. Hurrell [1995]), such that any site able to capture the leading mode of SLP variability (NAO) will also capture dynamically linked aspects of regional climate variability.
There have been studies attempting to reconstruct regional climate variability in different climate states from a limited number of locations (e.g. Allen et al. [1999]; Bakke et al. [2005]; Bahr et al. [2006]). Unfortunately, the same geographic site may record a qualitatively different mixture of mean and variance contributions in different climates. For example, the pattern of atmospheric variability could change, resulting in a center of action shifting towards or away from a proxy site. Alternatively, a change in seasonality (i.e., in how variability is distributed throughout the year) could affect proxies that record signals preferentially at certain times of the year.
To help assess the impact of such changes on the structure of surface climate variability in different climate states, we construct coherence maps for temperature and precipitation from the simulation of PI and LGM climate (Figs. 7-8 panels e to h). In these maps, higher values indicate that the variability at that location has higher coherance with variability throughout the North Atlantic. In the PI simulations, the models show a similar and coherent pattern of variability for both temperature and precipitation (Fig. 7-8 panels e and g). In contrast, the pattern of variability during the LGM is model dependent, but in each model different from that in the PI simulations ( Fig. 7-8 panels f and   h). Figure 7: PI and LGM correlations between North Atlantic winter surface air temperature (November to April) and PC1 (NAO-like index) for CCSM3 (a, b) and HadCM3M2 (c, d). An indicator of temperature coherence in the sector for CCSM3 (e, f) and HadCM3M2 (g,h): the value at each point is the absolute value of the area-averaged correlation between temperature at that point and the rest of the North Atlantic basin. The results from the IPSL model are similar to CCSM3 and the results from MIROC3.2 are similar to HadCM3M2. Only the winter months are included, as this is when the NAO-like signal is strongest. When including all months the result is the same, but with slightly weaker correlation patterns. Markers indicate the locations used in Table 5.
Figure 8: Same as Fig. 7 Hurrell [1995] in observations. In the LGM simulations, the models do not agree about the link between the leading mode of SLP variability and surface climate variability.
Two behaviors emerge from the model analysis: 1) for CCSM3, surface climate variability in the North  Table 2), which could be the reason for the disagreement between the models. Figures 7 and 8 show that because of changes in the link between the pattern of leading mode of SLP variability and surface climate, a location might be able to record the leading mode but not reflect regional surface climate variability at the LGM. As one example, regional variability may be disconnected from the leading mode of SLP, as seen for Summit, Greenland in the CCSM3 (compare panels b and f in Figs. 7 and 8). Another possibility is that because of the southward shift of the leading mode of SLP variability at the LGM, Greenland might be situated at the flank of the dominant North Atlantic atmospheric variability in a glacial climate. An example of this is that surface climate at the Summit of Greenland in HadCM3M2 reflects a combination of both regional and leading mode variability in the PI (panels c and g in Fig. 7, Table 4), but not in the LGM (panels d and h in Fig.   7).
Finally, changes in the seasonality of surface climate variability might cause an altered signal recorded by proxies (Krinner and Werner [2003]). Figure 9 shows how the magnitude of the seasonal cycle varies for two ice core locations in Greenland during the LGM compared to PI: the seasonal cycle of temperature is enhanced at the station NASA-U in the LGM simulations, whereas it is comparable at Summit; the seasonal cycle of precipitation is substantially modified at both locations. Neglecting this change in seasonality might cause a bias in LGM temperature estimates based on water isotopes, since the δ 18 O signal recorded at the LGM would have a different seasonal imprint than during the PI 20 (Steig et al. [1994]; Krinner and Werner [2003]). CCSM PI (25.0mm) CCSM LGM (9.5mm) HAD PI (19.0mm) HAD LGM (6.0mm) Figure 9: PI and LGM seasonal cycle for temperature (upper panels) and precipitation (lower panels) in CCSM3 and HadCM3M2 for two locations in Greenland (NASA-U (left) and Summit (right)). The annual mean has been subtracted to facilitate comparison between climate states and models.
Our study shows how assuming modern climate relationships for past climates can produce erroneous interpretations of paleoclimate records. Modelers must also be cautious when interpreting simulations of the LGM, given that the models are not able to depict a consistent spatial pattern of 21 surface climate or SLP variability in this climate state. In a few areas where the models do agree, it is possible to infer that in both climates a substantial amount of regional variability of either temperature or precipitation can be reliably reproduced, for example in southern Norway (Table 4 and Figs. 7 and 8) or in Labrador (Fig. 10, Table 4).

Conclusions
In this paper, we analyze surface climate variability in the extratropical North Atlantic using LGM and PI simulations from four climate models. We describe how changes in atmospheric variability may affect signals recorded by proxies. The main findings are: • The interannual variability of Northern Hemisphere sea level pressure (SLP) is significantly reduced at the LGM.
• The seasonal cycle of sea level pressure variability is decreased during the LGM. The reduction 22 is more prominent and significant in the winter months, when the variability is highest in both the PI and LGM climates.
• An NAO-like pattern is the leading mode of SLP variability in each LGM simulation examined, though it represents less interannual variability and the centers of action are weaker.
• The ice sheets and greenhouse gases largely determine the mean circulation and the amplitude of the leading mode of variability, while the SST/sea ice help to determine the amount of SLP variability. Different atmospheric models respond differently to the same ice sheet and greenhouse gas forcings, so simulated differences in the pattern and amplitude of the leading mode of SLP variability (NAO-like) appear to be sensitive to different model's physics and/or parameterizations.
• The relationship between atmospheric variability and surface climate variability is different during the LGM. Therefore, caution is necessary when interpreting proxy records using the modern relationship as an analog.

APPENDIX A
In order to compare the PI simulations with modern observations, the same analyses have been performed using the ERA-40 reanalysis and are shown in Fig. A1. 24