Climate Dynamics

, Volume 44, Issue 1, pp 359–369

Observed and SST-forced multidecadal variability in global land surface air temperature


  • L. H. Gao
    • Key Laboratory of Regional Climate-Environment for East Asia, Institute of Atmospheric PhysicsChinese Academy of Sciences
    • University of Chinese Academy of Sciences
    • Key Laboratory of Regional Climate-Environment for East Asia, Institute of Atmospheric PhysicsChinese Academy of Sciences
  • X. W. Quan
    • Cooperative Institute for Research in Environmental SciencesUniversity of Colorado
    • NOAA/Earth System Research Laboratory

DOI: 10.1007/s00382-014-2121-9

Cite this article as:
Gao, L.H., Yan, Z.W. & Quan, X.W. Clim Dyn (2015) 44: 359. doi:10.1007/s00382-014-2121-9


The characteristics of multidecadal variability (MDV) in global land surface air temperature (SAT) are analyzed based on observations. The role of sea surface temperature (SST) variations in generating MDV in land SAT is assessed using atmospheric general circulation model simulations forced by observed SST. MDV in land SAT exhibits regional differences, with amplitude larger than 0.3 °C mainly over North America, East Asia, Northern Eurasia, Northern Africa and Greenland for the study period of 1902–2004. MDV can account for more than 30 % of long-term temperature variation during the last century in most regions, especially more than 50 % in parts of the above-mentioned regions. The SST-forced simulations reproduce the observed feature of zonal mean MDV in land SAT, though with weaker amplitude especially at the northern high-latitudes. Two types of MDV in land SAT, one of 60-year-timescale, mainly observed in the northern mid-high-latitude lands, and another of 20–30-year-timescale, mainly observed in the low-latitude lands, are also well reproduced. The SST-forced MDV accounts for more than 40 % amplitude of observed MDV in most regions. Except for some sporadically distributed regions in central Eurasia, South America and Western Australia, the SST-forced multidecadal variations are well in-phase with observations. The Atlantic Multidecadal Oscillation and Pacific Decadal Oscillation signals are found dominant in MDV of both the observed and SST-forced land SAT, suggesting important roles of these oceanic oscillations in generating MDV in global land SAT.


Multidecadal variabilityEnsemble empirical mode decompositionAtmospheric general circulation modelSea surface temperature

1 Introduction

An important issue regarding the practice of decadal climate prediction is to estimate the possible role of low frequency climate variability. The global-mean surface air temperature (SAT) records of the last century exhibit a warming trend with superimposed multidecadal variability (MDV) (e.g. Latif and Keenlyside 2011). A recent study by Wu et al. (2011) suggested that the global-mean surface temperature series consists of a nonlinear warming trend and a multidecadal oscillation of ~65-year period. Their results revealed that up to one-third of global-mean warming during the late twentieth century was contributed by MDV, indicating the importance of MDV in global-mean surface temperature and its role in modulating global warming. How the global warming trend is modulated by low frequency climate variability at the regional scale remains an open question (e.g. Henriksson et al. 2012; Karoly and Wu 2005). One primary purpose of this study is to quantitatively assess the relative importance of MDV in temporal evolution of regional climate.

A major factor that may cause MDV in regional land surface temperature arises from the oceans. Recent studies investigated two dominant MDV patterns in the oceans known as the Pacific Decadal Oscillation (PDO) or Inter-decadal Pacific Oscillation (IPO) (e.g. Zhang et al. 1997) and the Atlantic Multidecadal Oscillation (AMO) (e.g. Kerr 2000). The oceanic oscillations could influence terrestrial climate through atmospheric circulations. Long-term oscillations were found in continental regions surrounding the North Atlantic (Schlesinger and Ramankutty 1994). Analyses showed the impact of AMO on regional climate (Knight et al. 2006), such as the US (Enfield et al. 2001) and North American and western European summertime climate (Sutton and Hodson 2005). Modeling studies further suggested that internal MDV originated from the North Atlantic-Arctic sector might influence surrounding regions and even global (Semenov et al. 2010). Observed MDV in mean temperature of the Northern Hemisphere could be reproduced by models driven by the Atlantic oscillations (Zhang et al. 2007). Most of the variability in observed global mean land temperatures during the past 128 years is found resulted from Sea Surface Temperature (SST) variations (Hoerling et al. 2008). These results potentially support an idea of realizing decadal climate prediction by using atmospheric general circulation models (AGCM) forced with predictions of future SST variations (e.g. Hoerling et al. 2011) using a two-tiered climate prediction approach, in which skillful SST prediction may be obtained by various dynamical and/or statistical methods. In this study we further explore the potential of applicability of this approach and investigate the role of SST in generating MDV in land SAT by examining MDV in a set of multi-AGCM ensemble simulations forced with historical SST variations, and identifying regions where the models make reasonable simulation of observed MDV.

The ensemble empirical mode decomposition (EEMD) method (Huang and Wu 2008; Wu and Huang 2009) is applied to decompose the annual SAT time series into three parts, an inter-annual to decadal component, a MDV component and a secular trend (ST) to facilitate discussions in the present study. The method and detailed procedure are described in Sect. 2. Spatial distribution, zonal and regional mean of the magnitude of MDV in observations are examined and compared with secular trend in Sect. 3.1. AGCM-simulated MDV is examined in Sect. 3.2. The role of SST in generating MDV in land SAT is discussed in Sect. 3.3. Conclusions are summarized in Sect. 4 with more discussions of existing problems and possible improvement in applications of the AGCM approach.

2 Data and methods

2.1 Data

The observational dataset of gridded monthly land SAT anomalies used in the study is the product (CRU TS 3.1) of the Climatic Research Unit (CRU) at the University of East Anglia (available from; Mitchell and Jones 2005), in 0.5° longitude by 0.5° latitude resolution. The set of model monthly land SAT analyzed in this study consists of 40 ensemble simulations of three AGCMs that are started from different initial conditions. They are all driven by the observed global SST and sea ice evolution with constant greenhouse gas concentrations (see Table 1 for details). The analyses presented in this study are performed for the period of 1902–2004 for which all the observed and model data are available. Monthly anomaly series of simulated grid SAT is obtained firstly by removing local 1902–2004 climatological mean annual cycle at each grid and then turned into annual anomaly series. All the observed and simulated data are interpolated into T42 grids of approximately 2.8° latitude by 2.8° longitude to facilitate comparative analyses between observations and model output. An ensemble mean annual grid land SAT dataset of 40 simulations is then calculated and used in the subsequent analyses in Sect. 3.2.
Table 1

The AGCM simulations used in this study


No. of runs

Resolution (lon × lat)


Configuration of SST, Sea ice and CO2 concentration




2.8 × 2.8


Kapalan SST in Tropical Pacific (20S–20N) and HADISST elsewhere; Time varying sea ice;

CO2 Fixed to pre-industrial level.

Kiehl et al. (1998), Seager et al. (2005)



2.5 × 2


Hurrell SST which merges HADIST and NOAA OI-SST;

Time varying sea ice;

CO2 Fixed to 1961–1990 mean level.

Delworth et al. (2006)



3.75 × 3



1961–1990 climatological annual cycle of sea ice;

CO2 Fixed to pre-industrial level.

Schubert et al. (2004)

The AMO index (available at and the PDO index (available at are used in this study. The former is defined by annual SST anomalies averaged over the North Atlantic (0–60°N, 0–80°W; Trenberth and Shea 2006). The latter is derived from the leading principle component (PC) of monthly SST anomalies in the North Pacific Ocean in which the global warming signal has been removed. In the present study, the 12 monthly PDO indices are averaged to represent the annual signals of PDO.

2.2 The EEMD method and MDV component

EEMD is a self-adaptive analysis method that decomposes a time series into several oscillatory components on various timescales and a non-linear ST. The method has been used to extract signals of climate change and variability of various timescales in global and regional climate (e.g. Qian et al. 2011; Xia et al. 2013).

In this study, an ensemble size of 1,000 and a white noise with amplitude of 0.4 times the standard deviation (STD) of the annual SAT series are used to conduct the EEMD. For the 103-year length SAT series, five intrinsic mode functions (IMFs) denoted as IMFn (n = 1,5) and a ST component are obtained and divided into three generalized modes: (1) interannual to decadal variability including IMF1, IMF2, and IMF3; (2) MDV including IMF4 and IMF5; and (3) the ST as a nonlinear long-term trend. Temporal variations of the three generalized modes are then reconstructed by adding all the IMFs in each category, respectively. Figure 1 shows the three reconstructed modes of a sample time series using the EEMD method. A power spectrum analysis indicates that the dominant timescales of the reconstructed MDV component in observed and simulated SAT series range from 20 to 80 years (figure not shown). The timescale of a component series can also be roughly estimated by the time-window between two peaks or two troughs. The sum of ST and MDV component well represents the long-term SAT variations (Fig. 1, bottom panel), demonstrating the efficiency of this method. MDV components of the land SAT series, AMO index and PDO index for the period of 1902–2004 are also calculated to facilitate comparative analyses. Note that the original series of AMO index includes a secular trend and that of PDO involves strong inter-annual and decadal variability.
Fig. 1

An example of ensemble empirical mode decomposition analysis for a sample SAT series (from top: the inter-annual component; MDV component; secular trend and the original SAT anomaly series (dot line) superposed with the sum of MDV component and secular trend

3 Results

3.1 Observed MDV and its role in modulating long-term land SAT variation

The geographic distribution of the amplitude of MDV represented by its STD is shown in the upper panel of Fig. 2. Regions of strong MDV are found mainly in the Northern Hemisphere, specifically in Northern Eurasia, North America, East Asia, North Africa and Greenland. The amplitudes of MDV in these regions generally exceed 0.3 °C. Weak MDV exists in central Eurasia, south Asia and most regions in the Southern Hemisphere.
Fig. 2

Geographic distribution of standard deviation of MDV (STD_MDV, upper panel), STD_ST (middle panel) and R = STD_MDV/(STD_MDV + STD_ST) (lower panel) of land SAT based on CRU dataset. Zonal mean results are plotted in the right panels

Zonal mean STD of MDV is roughly about 0.2 °C, larger in high-latitudes in the Northern Hemisphere (Fig. 2, right panel). The result is comparable with the analysis of Northern Hemisphere mean SAT series by Zhang et al. (2007). In general, the amplitude of MDV is larger in the Northern hemisphere than in the Southern hemisphere, and the MDV is weaker in the tropics than in mid-latitudes of both hemispheres.

As a measure of the importance of MDV relative to long-term climate change, we calculate the ratio R that compares the STD of MDV to the sum of STD of MDV and STD of the secular trend, R = STD_MDV/(STD_MDV + STD_ST). The value of R reflects the extent to which a secular trend may be affected by the temporal variation of MDV. A larger R indicates less dominance of the secular trend. When R is larger than 0.5, the variance of MDV is larger than the variance of ST, the condition of MDV becomes a dominating factor. The global pattern of R is shown in the lower panel of Fig. 2, compared with that of ST in the middle panel. In general, R is larger than 0.3 in most of the land areas over the globe and is larger than 0.5 in southeastern North America, East Asia, western tropical Africa and South America, Northern Eurasia and Greenland. From the point of view of inter-decadal prediction, correctly predicting possible future evolution of MDV in the upcoming decade is critical for these regions, where the signal of climate change in terms of a secular trend is potentially easily modified by MDV. Note that the pattern of R appears quite different from that of MDV because the magnitudes of the secular trend are not the same in different regions. The R ratio is small in other regions in the Northern Hemisphere, especially in the central part of the Eurasian continent, partly because the secular trend is stronger in these regions (Fig. 2, middle panel). In these regions, MDV is thus less important in the practice of decadal prediction.

To see the main temporal structure of MDV, we examined the zonal mean multidecadal variation in five latitudinal bands. The results are shown in Fig. 3. The multidecadal variations in both the northern middle and high-latitudes show roughly 60-year variability with amplitude about 0.2 °C and 0.5 °C, respectively. In the southern latitudes, the MDV has smaller amplitude, with less significant 60-year variability superimposed with considerable 20–30-year variability.
Fig. 3

Zonal mean multidecadal variations of land SAT based on CRU dataset at specified latitude bands

Four regions including East Asia (90°E–130°E, 20°N–50°N), North America (60°W–130°W, 20°N–50°N), western tropical Africa (0°E–30°E, 10°S–10°N) and Central Eurasia (40°E–90°E, 30°N–60°N) are selected to further examine the characteristics of MDV at regional scales. The first three regions have relatively higher values of R and the last region has low value of R. Figure 4 shows the multidecadal variations and corresponding ST, obtained from the EEMD analysis of the regional mean SAT observations. MDV in East Asia and North America have the amplitude of about 0.2 °C with the ratio R of about 0.4 (Table 2, first two rows). The timescale of MDV in these two regions is about 60 years or longer. In contrast to these two regions, MDV is relatively less important in Central Eurasia, where the STD of MDV is about 0.1 °C while that of the ST is about 0.4 °C (3rd row in Table 2). The timescale of MDV in Central Eurasia is mainly about 20–30 years. The amplitude of MDV in western tropical Africa is about 0.1 °C, but it is still relatively important because the magnitude of the secular trend in this region is also at the level of 0.1 °C, making the R ratio larger than 0.5 (Table 2, bottom row).
Fig. 4

Multidecadal variations (solid line) and secular trend (dashed line) of regional mean land SAT based on CRU dataset

Table 2

The STD of MDV and ST and the ratio (R = STD_MDV/(STD_MDV + STD_ST) of the annual SAT series for four regions, based on the CRU data





East Asia




North America




Central Eurasia




Western tropical Africa




3.2 Global features of SST-forced MDV in land SAT

3.2.1 Amplitude

In order to assess the ability of SST-forced atmospheric model inter-comparison project (AMIP) simulations to reproduce observed MDV in land SAT, we first examine the amplitude of MDV in the 40 AMIP simulations for all the grid points over the global land. For each grid point, the amplitude of MDV is estimated in two ways: (1) calculate the STD of MDV for each of the individual runs and then calculate the mean of the 40 STD values (Fig. 5, upper left panel); (2) calculate the ensemble mean first, and then get the STD of MDV of the ensemble mean (Fig. 5, lower left panel). The average amplitude of MDV obtained from method-i is larger than the amplitude of the ensemble mean from method-ii, while both are smaller than the observations (Fig. 2, top panel). In general, the AMIP runs are capable of reproducing the zonal mean structure of the amplitude of MDV in observations (Fig. 6). Similar to the observed feature, the zonal mean amplitude of simulated MDV also shows a trough near the equator and higher values in the northern middle- and high-latitudes. An exception is seen in the latitudes higher than 50°N, where observations show an increase towards higher latitudes while the simulations show a slight decrease in average. The amplitudes of zonal mean MDV of individual runs show diverse results for the northern high-latitudes, which are also indicated by the geographic pattern of STD of the results of 40 individual runs (Fig. 5, upper right panel). The simulated results also capture the observed feature of smaller amplitude of MDV in the Southern Hemisphere than that in the Northern Hemisphere. However, the models are less successful in reproducing some local features, e.g., the weak MDV in central Eurasia.
Fig. 5

Standard deviation (STD) of MDV of land SAT obtained by (1) average of the STD of MDV in each of the 40 individual runs, named as ave_STD (upper left) and the spread of the 40 individual STD values with regard to ave_STD (upper right); (2) STD of MDV of the multi-model ensemble mean, named as STD_ens (lower left) and the spread of the 40 individual STD values with regard to STD_ens (lower right)
Fig. 6

Zonal mean standard deviation of MDV of land SAT for CRU dataset (red line), 40 simulations’ average (black line) and corresponding individual simulations (thin green line), respectively

The signal of MDV forced by SST is better represented by the multi-model ensemble mean because “noise” of atmospheric internal variability of short timescales and model dependent uncertainties are diminished in the ensemble mean. The pattern of MDV of the model ensemble mean land SAT is examined in the lower left panel of Fig. 5. In the case of the ensemble mean, larger amplitudes of the SST-forced MDV occur in the northern mid-latitudes and Greenland, indicating greater SST impact in these regions. In northern Eurasia and northern North America, MDV is much weaker in the ensemble mean than the average of MDV of individual runs (Fig. 5, upper left panel), suggesting a larger uncertainty in the phase of individual model’s local MDV in response to the SST forcing. This larger uncertainty is also indicated by the larger difference between the amplitude of MDV in each individual run and the amplitude of MDV in their ensemble mean in these regions (Fig. 5, lower right panel). The amplitude of MDV of the models ensemble mean is compared to that of observed MDV in the left panel of Fig. 7. The ratio of the SST-forced MDV in the model ensemble mean to the observed MDV is generally larger than 0.4 in most regions except at high-latitudes where the ratio is lower than 0.3.
Fig. 7

Geographic distribution of the ratios of MDV of the model ensemble mean land SAT to that of CRU dataset (left) and global correlations between multidecadal variaitons of land SAT in the multi-model ensemble mean and those in the CRU data

3.2.2 Temporal structure

Global correlations between multidecadal variations in the model ensemble mean and observations are examined in the right panel of Fig. 7. In general, the temporal evolution of observed MDV is well reproduced by the multi-model ensemble in most global land areas, although disagreement is noticed in some sporadically distributed regions in central Eurasia, South America and Western Australia. The zonal mean multidecadal variations in the model ensemble mean show roughly 20–30-year and 60-year timescale variability, as in observations (Fig. 8). Except in northern high-latitudes, the phases of multidecadal variations in observations and simulations coincide well. In particular, the model ensemble mean reproduces observed 60-year variability well in regions such as East Asia and North America (Fig. 9). The model ensemble mean fails to reproduce observed 20–30-year variability in central Eurasia (Fig. 9), implying possible stronger local climate variability in comparison with SST-forced variability or imperfectness of current AGCM experiments.
Fig. 8

Zonal mean multidecadal variations of land SAT based on CRU dataset (red line) and the model ensemble mean (green line) at specified latitude bands. The black line indicates the MDV component of AMO
Fig. 9

Multidecadal variations and secular trend of regional mean land SAT for CRU dataset (red line) and the model ensemble mean (green line)

3.3 Understanding the role of SST variation in generating MDV in land SAT

The present AGCM modeling is forced by observed SST evolution, without other factors such as changing greenhouse gases in the simulations. Therefore, the results of MDV in the land SAT are mainly due to SST variations. The process is mainly that the oceans heat the surface atmosphere and then influence the land SAT through global atmospheric circulation by energy transportation (Compo and Sardeshmukh 2009; Dommenget 2009). To further explore possible links of SST variations to generate MDV in land SAT, we perform principle component analysis (PCA) to compare dominant modes of MDV in global land SAT in the AGCM results with those in the observations. Several principle components (PCs) of multidecadal variations in both observations and the model ensemble mean results are calculated.

PC1 is dominant in global land SAT for both observations and the model ensemble mean, explaining 49.5 and 67.8 % of the total variance, respectively. The PC1 time series exhibit roughly 60-year variability, albeit with a slight phase discrepancy between the model ensemble and observations (Fig. 10, left panel). Geographical pattern of the observations corresponding to PC1 shows nearly global coherent variation (Fig. 11, upper left panel). Exceptions only occur for some regions in central Eurasia and the Southern Hemisphere. The geographic pattern of PC1 of the model ensemble also shows global coherent variation similar to observations (Fig. 11, upper right panel). The MDV component of PC1 for observations and that for the model ensemble are stronger in the Northern Hemisphere than in the Southern Hemisphere. Differences between model and observations mainly occur in central Eurasia and high Northern latitudes. The PC1 time series and patterns suggest that dominant MDV in land SAT is of a timescale of about 60-years and mainly exists in the Northern Hemisphere. It also implies that the SST variation is inclined to force climate variability of this timescale especially in America, North Africa and East Asia. The PC1 time series of this study is similar to the result of empirical orthogonal function analysis by Kravtsov and Spannagle (2008) based on detrended regional mean surface temperatures of 9 selected regions including both ocean and land areas.
Fig. 10

Left Normalized PC1 timeseries derived from principle component analysis of MDV components of the CRU land SAT (red) and the model ensemble mean (green). The black line is the normalized MDV component of AMO. Right same as the left panel but for PC2; the black line is normalized MDV component of PDO
Fig. 11

Spatial patterns corresponding to the first two PCs of the MDV components of the CRU land SAT (left) and the model ensemble mean (right)

As Fig. 10 (left panel) shows, both the PC1 time series of observation and that of the model ensemble mean SAT are well in-phase with the MDV component of AMO. This suggests that AMO may have played a major role in generating MDV of global land SAT during the last century. In particular, the SST-forced multidecadal variations of land SAT in northern middle- and high-latitudes match well with the MDV component of AMO (Fig. 8a, b).

PC2 of the land SAT observations and that of the model ensemble explain 14.1 and 17.6 % of the total variance, respectively. The PC2 time series in observations and the model ensemble are also close to each other, albeit with a slight phase discrepancy (Fig. 10, right panel). The geographical pattern of PC2 in observation shows positive phase in northern North America, northern Eurasia and Australia, and negative phase in most of northern mid-latitude regions (Fig. 11, lower left panel). PC2 of the model ensemble shows a similar pattern (Fig. 11, lower right panel). Both the PC2 time series of observation and that of the model ensemble are close to the MDV component of PDO (Fig. 10, right panel). This suggests that PDO is another important factor for generating MDV in land SAT. Note that both of the variance explained by PC1 and PC2 in the model ensemble are larger than those in land SAT observations, because the current model simulations mainly include SST-forced variability, while the observations should include variability arising from other factors not within the scope of the present study.

4 Summary and discussion

This study presents an overview of MDV in observed global land SAT and in SST-forced AGCMs. The MDV in land SAT shows obvious regional differences, with strong MDV larger than 0.3 °C in Northern Eurasia, North America, East Asia, North Africa and Greenland and weak MDV in central Eurasia and the Southern Hemisphere (Fig. 2, upper panel). MDV plays an important role in modulating trends in the temporal evolution of regional SAT. The contribution from MDV can account for more than 30 % of the long-term temperature variation during the last century in most regions (Fig. 2, lower panel), and even more than 50 % in parts of southeastern North America, East Asia, western tropical Africa, western South America, Northern Eurasia and Greenland. The geographic pattern of this fraction helps to determine whether and where MDV weighs more in regional decadal climate prediction.

By analyzing the results of AGCMs forced by historical SST variation, we find that AGCMs are capable of reproducing the main features of observed MDV in land SAT. The simulations capture the zonal mean structure of the amplitude of MDV in observations including larger amplitude in some regions in the Northern Hemisphere (Fig. 6). The model ensemble also shows stronger MDV in mid-latitudes and Greenland, indicating a robust role of SST for these regions. Two types of observed variability, one with 60-year timescale in the northern middle- and high-latitudes and another with 20–30 year timescale in southern latitudes, are well reproduced by the AGCMs. The zonal mean multidecadal variations in the model ensemble match in phase with those of observations, though there is considerable discrepancy for northern high-latitudes (Fig. 8). The contribution of SST-forced MDV to the observation is found larger than 0.4 in most regions and the simulated temporal structure is coincident with observation except for sporadically distributed regions (e.g., central Eurasia and parts in the Southern Hemisphere, Fig. 7).

Two major modes of MDV in global land SAT observations, derived from PCA, are closely related to the well-known AMO and PDO signals, respectively, and are well reproduced by each model simulation (figures not present) and the multi-model ensemble (Fig. 10). These results further confirm the important role of SST variations (particularly AMO and PDO) in generating MDV in land SAT especially in the Northern Hemisphere, which has been discussed in other studies (e.g. Knight et al. 2006; Kravtsov and Spannagle 2008; Zhang et al. 2007). Robustness of the present results partly depends on the model performance. The present study involves three different ACGMs forced by different observed SST and sea ice datasets. Note that there are different regional variations and trends in different observational SST datasets (Jha et al. 2013; L’Heureux et al. 2013). The multi-model ensemble mean is emphasized in the present study, as it is expected to diminish uncertainties arising from either the models or the forcing.

While the present results exhibit potential of decadal climate prediction via SST-forced AGCM especially for regions such as parts of North America, East Asia, northern Eurasia and Greenland, it is worthwhile noting potential problems or improvements for application. One is large uncertainty of the simulated amplitude of MDV in the Northern high-latitudes, which may be caused by relatively large atmospheric noise related to sea-ice-atmosphere interaction absent in the present models. Our results also indicate apparent disagreements between observations and the model ensemble in inland regions such as central Eurasia, regarding not only stronger MDV in observations but different dominant time-scales of MDV. Whether these are model dependent remains a question. Another important issue deals with the generally weaker amplitude of MDV of the model ensemble mean compared to the observations. This is either due to imperfectness of current AGCMs in responding to SST forcing or because there are components of MDV associated with other forcings for land SAT. Further studies with considerations of these issues will be beneficial.


The authors thank Philip Pegion, Taiyi Xu, and Tao Zhang at NOAA/ESRL/PSD for helps in preparing the data. Thanks are also due to two anonymous reviewers for their constructive comments. This work was supported by grants CAS-SPRP XDA05090000 and MOST-NBRPC 2012CB956200.

Copyright information

© Springer-Verlag Berlin Heidelberg 2014