The loss of Arctic sea ice has wide-ranging impacts, some of which are readily apparent and some of which remain obscure. For example, recent observational studies suggest that terrestrial snow cover may be affected by decreasing sea ice. Here, we examine a possible causal link between Arctic sea ice and Siberian snow cover during the past 3 decades using a suite of experiments with the National Center for Atmospheric Research Community Atmospheric Model version 3. The experiments were designed to isolate the influence of surface conditions within the Arctic Ocean from other forcing agents such as low-latitude sea surface temperatures and direct radiative effects of increasing greenhouse gases. Only those experiments that include the observed evolution of Arctic sea ice and sea surface temperatures result in increased snow depth over Siberia, while those that maintain climatological values for Arctic Ocean conditions result in no snow signal over Siberia. In the former, Siberian precipitation and air temperature both increase, but because surface air temperatures remain below freezing during most months, the snowpack thickens over this region. These results suggest that Arctic Ocean surface forcing is necessary and sufficient to induce a Siberian snow signal, and that other forcings in combination can modulate the strength and geographic extent of the response.
 Changes in the Northern Hemisphere high-latitude environment, including a rapid decline in summer Arctic sea ice cover, are under scrutiny for their impact on global climate [Hassol, 2004; Stroeve et al., 2005, 2007; Comiso et al., 2008]. For example, one recent study examines changes in precipitation, evapotranspiration, and river discharge over the terrestrial pan-Arctic region through the analyses of observations and coupled GCM simulations [Rawlins et al., 2010]. They found positive trends in precipitation, evapotranspiration, and river discharge both in observations and in model simulations, suggesting an intensification of the freshwater cycle.
 A study based on 110 stations from the Former Soviet Union (FSU) showed that the mean cold season snow depth was affected by warming in only a few areas during the second half of 20th century [Fallot et al., 1997]. A more recent study of 820 Russian stations with snow measurements between 1966 and 2007 revealed that snow cover duration decreased over the northern regions of European Russia, and increased over central Siberia (Yakutia) and the Far East [Bulygina et al., 2009].
 Warming due to increasing greenhouse gases leads to the loss of seasonal snowpacks, which causes a threat to the water supply of one-sixth of the world's population [Barnett et al., 2005]. General circulation model simulations suggest that in response to increasing greenhouse gases, snow water equivalent (SWE) will decrease over most of the midlatitudinal regions of the Northern Hemisphere, but will increase over the Canadian Arctic and Siberia [Barry et al., 2007; Räisänen, 2008]. A recent study based on 11 high-resolution regional model simulations collectively suggest an expected loss of snow water equivalent over Northern Europe during the 21st century; though there are some exceptions (e.g., the mountains of Northern Sweden) [Räisänen and Eklund, 2011].
 Here, we investigate a possible causal linkage between Siberian snow and sea ice using a suite of modeling experiments, the objective of which is to isolate the impact of changing sea ice from other forcings. Details of the experiments are discussed in section 2, results are shown in section 3 and conclusions are discussed in section 4.
2. Experiments and Methodology
 We use the National Center for Atmospheric Research (NCAR) Community Atmospheric Model version 3 (CAM3) coupled to the Community Land Model (CLM), which are the atmospheric and terrestrial components of the Community Climate System model version 3 (CCSM3), respectively. These models are described in detail by Collins et al.  and references therein. The strengths and weaknesses of CAM3 in simulating the mean state and interannual variability are reported by Hurrell et al. .
 We analyze results from a suite of five CAM3/CLM experiments, which have been specifically designed to isolate the impacts of Arctic Sea Surface Temperature (SST) and sea ice changes from other forcings (discussed below). These experiments have been performed at T42 horizontal resolution (approximately 2.8° of latitude and 2.8° of longitude). Each of the 5 experiments consists of five ensemble members each of which begins from different atmospheric initial conditions. We present results for the average of five ensemble members.
 For these experiments we group boundary forcings into two categories: radiative and surface. Radiative boundary forcings include atmospheric chemical composition (greenhouse gases, tropospheric and stratospheric ozone, sulfate and volcanic aerosols) and their associated radiative impacts as well as solar insolation variations. Surface boundary forcings include global SST and sea ice concentration variations. The differences between the experiments are whether or not they include the time-evolving observed values for these forcings. When the time-varying radiative boundary forcings are not included, they are maintained at 1990 levels to approximate conditions during the late 20th century. When the time-varying surface boundary forcings are not included, the mean seasonal cycle is repeated during each year of the experiment. TheHurrell et al.  data set has been used for the SST and sea ice concentrations.
Table 1summarizes the forcings included in each of the five experiments denoted as “GHG+SST+ICE,” “SST+ICE,” “ICE,” “GHG” and “Tropical SST.” The “GHG+SST+ICE” experiment is forced by all observed, time-evolving boundary conditions, including both radiative and surface. “SST+ICE” is forced by global time-evolving surface boundary conditions and constant radiative forcing. “ICE” is forced by time-varying Arctic sea ice concentrations, with climatological values for the other surface forcing components with the exception of SSTs within grid boxes partially covered by ice. In these regions, we have incorporated the local, direct effect of ice loss on SST (e.g., when there is less ice, there is more incident solar radiation at the sea surface which heats the ocean mixed layer) by including time-varying SSTs in those grid boxes and months when the sea ice concentration is less than the climatological value by at least 10% (seeScreen et al. [2012b] for additional details). “GHG” is forced by the observed evolution of radiative forcings [Meehl et al., 2006], with climatological values for the mean seasonal cycle of SSTs and sea ice. “Tropical SST” is forced by the observed evolution of SSTs in the tropics (20°N–20°S), with climatological values elsewhere. More details of these experiments are provided by Deser and Phillips  and also can be found online at http://www.cesm.ucar.edu/working_groups/Climate/experiments/ccsm3.0/.
Table 1. The Forcing Characteristics of the CAM3 Experiments
Radiative Forcings (Greenhouse Gases, Tropospheric and Stratospheric Ozone, Sulfate and Volcanic Aerosols, and Solar Insolation)
Extra Tropical Surface Forcings (SSTs and Sea Ice Poleward 20°)
Tropical Surface Forcings (SSTs 20°N–20°S)
GHG + SST + ICE
SST + ICE
fixed at 1990 level
climatological seasonal cycle
climatological seasonal cycle
fixed at 1990 level
climatological seasonal cycle
fixed at 1990 level
time-varying for sea ice and local SST (see text); climatological seasonal cycle for SSTs elsewhere.
climatological seasonal cycle
 The temporal domain of our analyses is restricted to October through March, 1979/1980–2007/2008, which coincides with the period of regular satellite observations of sea ice from passive microwave observations. This helps us to understand how the historic simulations can be compared with the observed variability. The spatial domain of our analysis is limited to the Arctic Ocean and Eurasia north of 45°N. We present results at bimonthly time scale, i.e., October–November (ON), December–January (DJ) and February–March (FM).
 The variables analyzed include monthly mean snow depth (‘SNOWDP’ in meters), lowest model level water vapor mixing ratio (‘QBOT’ in gm/kg), surface air temperature (‘SAT’ in °C), precipitation (‘PRECIP’ in mm/d) and snowfall (‘SNF’ in mm/d). Linear regression analysis has been employed between the September sea ice extent time series (SIE) (available from NSIDC's Sea Ice Index, http://nsidc.org/data/seaice_index) and each of the variables. Because the SIE has a negative trend, the resulting regression coefficients are displayed with reversed sign so that positive values represent increases in the regressed variable as SIE decreases. Similar results are obtained with linear trend analysis (not shown). We assess the 5% statistical significance of the regression coefficients following a Student's t test [Helsel and Hirsch, 1992]. Although we present regression coefficients at each grid cell, we discuss only those features that show significant values over large geographical areas.
 As context for the regression analysis, Figure 1 shows climatological October–November snow cover conditions from the “GHG + SST + ICE” model experiment and from observations. Snow cover extent (fraction of land covered by snow) is shown for the model as well as from observations [Armstrong and Brodzik, 2005]. EASE grid (Equal Area Scalable Earth Grid) snow extent data [Armstrong and Brodzik, 2005] based on NOAA-NESDIS snow charts has been obtained from the National Snow and Ice Data Center (NSIDC). Please note thatFigures 1a and 1b have different spatial resolution. A visual comparison indicates that the spatial patterns match well, with larger values over high latitudes and over mountainous regions (compare Figures 1a and 1b).
 Regression analyses between September SIE and SNOWDP in the following ON, DJ and FM show that loss of September SIE is associated with an increase in SNOWDP over central and eastern Siberia (over the northern Yenisey and Lena river basins) in the “GHG + SST + ICE,” “SST + ICE” and “ICE” experiments (Figure 2). Significant increases in SNOWDP are observed during DJ and FM in the “GHG + SST + ICE” and “SST + ICE” experiments and during ON, DJ and FM in “ICE” experiment. In the two experiments without time-varying Arctic surface boundary forcings (“GHG” and “Tropical SST”), no significant signal in SNOWDP emerges over this region (not shown). Results from individual ensemble members (not shown) exhibit similar Siberian snow depth signals, though the geographical extent and statistical significance of this association vary slightly among them.
 Another noticeable pattern is found over western Eurasia, where retreating September SIE is followed by significantly decreasing SNOWDP during October through March over the region extending from Scandinavia to the Ob river basin in both “GHG + SST + ICE” and “SST + ICE” experiments (Figure 2). This signal is not apparent in the “ICE” experiment. The “GHG” and “Tropical SST” experiments show no coherent and consistent signal except over central Russia in “Tropical SST” (not shown).
 To understand the snow depth signals, we examine regression maps between September SIE and subsequent SAT (Figure 3), QBOT (Figure 4), PRECIP (Figure 5) and SNF (Figure 6). Reduced September SIE is followed by statistically significant warming over the Pacific sector of the Arctic Ocean and over the adjacent Eurasian landmass including Siberia in all three experiments, mainly during ON (Figures 3a–3c). The strongest response is found over the East Siberian, Chukchi and Laptev Seas which coincides with the area of recent observed summer ice loss. In the “ICE” experiment, this signal does not persist into DJ and FM. In other experiments, the ON signal is more extensive than “ICE” experiment, and persists in some form through March (Figure 3). Neither of the experiments that exclude time-varying Arctic surface boundary forcing exhibit significant Siberian warming (not shown).
 Response of QBOT to the loss of September Arctic SIE during ON, DJ and FM is similar to the responses of SAT (compare Figures 3 and 4). Figures 4a–4c confirm an increase in moisture over the Arctic Ocean and over Siberia mainly during ON in all three experiments, which is intuitive from significant warming as seen in Figures 3a–3c over the same region. The increased QBOT in the atmosphere provides a mechanism for increased precipitation and snowfall in this region associated with changing sea ice conditions.
 Therefore, we investigate whether this model shows the expected precipitation changes. In fact, a decrease in September SIE is associated with a significant increase in precipitation over the Arctic Ocean and adjacent Siberia during ON in the experiments with time-varying sea ice forcing (Figures 5a–5c). The experiments without time-varying sea ice forcing show no such signal (not shown). There is some persistence of the significantly increased precipitation signal into DJ in the “GHG + SST + ICE” and “SST + ICE” experiments, but not in the “ICE” experiment.
 There are other regions with significant precipitation responses that are unique to each of the five experiments. For example, in “GHG + SST + ICE,” precipitation increases over the western North Pacific Ocean and over eastern Eurasia during DJ and FM associated with retreating summer SIE (Figures 5d and 5g). In “SST + ICE,” we find a FM precipitation signal that is mainly centered over the Arctic Ocean and over the Eurasian landmasses (Figure 5h). Regression plots for the Tropical SST experiment (not shown) show an increase in precipitation over the North Pacific and a decrease over part of Eurasia during DJ following summer SIE loss. A more detailed discussion of these changes is beyond the scope of this manuscript.
 Regression patterns of snowfall are consistent with the precipitation patterns (compare Figures 5 and 6). The experiments with time-varying sea ice forcing show an increase in snowfall associated with decreasing sea ice over the Arctic Ocean and over adjacent Siberia during ON (Figures 6a–6c). Furthermore, a decrease in snowfall is found over western Eurasia during ON and DJ in “GHG+SST+ICE” experiment and during ON in “SST+ICE” experiment.
 Regional time series of SAT and PRECIP over Siberia (60°N–75°N, 95°E–135°E) are shown in Figure 7. This region coincides with the area where a loss of September SIE corresponds to an increase in SNOWDP (recall Figure 2). Temperature trends over Siberia appear, at first glance, to be related to radiative and SST forcing, but not to sea ice forcing: significant temperature trends are found only in those experiments where ice forcing is complemented by other forcings (Table 2). However, this may be due to the fact that, while the other forcings have trends during the entire period of this analysis, the sea ice forcing becomes strong only after 2000. Figure 7shows that during the more recent period, Siberian temperature and precipitation both increase in concert with decreasing sea ice extent. In experiments without time-varying sea ice forcing, no significant trend is found with the exception of “GHG” during DJ (Table 2).
Table 2. Linear Trend Coefficients for Seasonal Surface Air Temperature (SAT, °C/Decade), Precipitation (PRECIP, mm/Day/Decade), and Snow Depth (SNOWDP, cm/Decade) Over Siberia Between 1979 and 2008a
Significance levels are shown in parentheses; values in bold are significant at p < 0.05 (at least 5% confidence).
 In the experiments with time-varying sea ice forcing, precipitation trends over Siberia are significant and positive during ON (Table 2). In those without sea ice forcing, no significant precipitation trends are found over Siberia during this season. There is no indication that sea ice alone is related to precipitation variations in December through March.
 As a result of these temperature and precipitation signals, we find statistically significant positive trends in Siberian snow depth in “GHG+SST+ICE” during DJ and FM; in “SST+ICE” during FM; and in “ICE” during ON (Table 2). Despite the positive temperature trends over Siberia, monthly average temperatures remain below −10°C (Figure 7a), resulting in atmospheric conditions that remain favorable for precipitation falling as snow (Figure 6), and for the persistence of snow on the ground. This explains why the trend in SNOWDP becomes stronger during FM in “GHG+SST+ICE” and “SST+ICE,” as is observed in Figure 2 as well. Synergistic effects between SST and sea ice trends may result in an extension of the duration of the SNOWDP signal into late winter in these experiments compared to the “ICE” experiment, but further investigation is needed to substantiate this claim. In the “GHG” and “Tropical SST” experiments, we find no significant trend in Siberian SNOWDP in any season (Table 2).
4. Conclusions and Discussion
 We have presented results from a suite of five CAM3/CLM experiments designed to isolate the impact of changes in sea ice and SSTs in the Arctic Ocean and marginal seas on Siberian snow cover. The experiments include two groups of forcings—radiative and surface boundary—in different combinations. Snow depth is found to increase over central and eastern Siberia following summer sea ice loss only in those experiments which include time-varying Arctic sea ice and SSTs (e.g., “GHG+SST+ICE,” “SST+ICE” and “ICE”): the other experiments show no such signal. This leads us to conclude that, according to the CAM3/CLM model, Arctic surface boundary forcing plays a major role in generating this Siberian snow increase. These results are consistent with the observational findings ofGhatak et al.  which show an empirical connection between summer Arctic sea ice loss and autumn/early winter snow cover duration over Siberia.
 Statistically significant warming is found over the Arctic Ocean as well as over Siberia. This suggests a positive feedback in which the loss of summer sea ice results in atmospheric surface level warming over the Arctic Ocean and Eurasia. These results are consistent with the recent warming observed over the Arctic primarily during autumn [Serreze et al., 2009; Screen and Simmonds, 2010]. This suite of experiments suggests that the effect of sea ice variations on temperature, water vapor content, precipitation and snowfall over the Arctic Ocean and Siberia is most extensive during ON. Stroeve et al.  suggest more frequent and intensified cyclones over the Atlantic sector of the Arctic during years of low sea ice; this leads to an increase in autumn cyclone associated precipitation, which is also apparent over Eurasia. Screen and Simmonds  also showed an increase in Arctic precipitation in September–October–November.
 Increased snow cover over Siberia in these three experiments is caused by increased surface air temperature and precipitation during ON. The temperature connection may seem counterintuitive, but in a scenario with significant warming, when average monthly temperatures remain below freezing during the snow season, as they do in this region, conditions are favorable for increased snowfall and more persistent snow on ground. Atmospheric warming over these regions increases the availability of moisture, facilitating increased snow fall and a deeper snowpack over Siberia.
 In the experiment with only Arctic Ocean surface forcing, the snow signal over Siberia is apparent. In the two other experiments that include Arctic Ocean surface forcings plus other surface and radiative forcings, the signal is more geographically extensive during winter. In the experiments with no Arctic Ocean surface forcing, no Siberian snow signal is apparent. Thus, Arctic Ocean surface forcing is necessary and sufficient to produce a Siberian snow signal, but other surface and radiative boundary forcings can modulate the magnitude and geographic extent of the signal.
 It should be noted that our conclusions are based solely on CAM3/CLM simulations and thus may be model-dependent. For example, a recent study that performed similar “ICE” experiments with the UK-Australian Unified Model version 7.3 (UM7.3) finds considerably weaker fall/winter warming over Siberia in UM7.3 compared to CAM3 as well as somewhat different precipitation responses [Screen et al., 2012b].
 This study addresses one aspect of a feedback cycle that has previously been identified in the high-latitude climate system. Other studies have established a remote teleconnection between autumn Siberian snow cover and the following winter's Northern Annular Mode atmospheric circulation pattern [Gong et al., 2003; Fletcher et al., 2009; Cohen et al., 2012]. In fact, a statistical model using October mean snow cover was able to generate a better forecast of the winter air temperature over the extra-tropical Northern Hemisphere than a dynamical forecast system [Cohen and Fletcher, 2007]. Our study contributes to our understanding of the large-scale feedback system by explaining the response of fall snow to the loss of summer sea ice loss. It is hoped that additional examination of these model results will shed more light on the physical mechanisms responsible for the relationships identified in this study.
 This work has been supported by the NASA Cryospheric Sciences Division (NNX08AQ70G), NASA MEaSUREs award NNXO8AP34A, and NOAA grant NA08OAR4310678. Part of the funding for Debjani Ghatak came from an Institute of Marine and Coastal Sciences postdoctoral fellowship. James Miller is thanked for helpful discussions. NCAR is sponsored by the National Science Foundation. C. Deser gratefully acknowledges support from the NSF Office of Polar Programs. We thank the reviewers for their constructive comments, which improved the quality of the manuscript.