1. Introduction
The Atlantic meridional overturning circulation (AMOC) is an important oceanic system that strongly influences climate over large parts of the Atlantic Ocean and Northern Hemisphere through the meridional transport of heat and freshwater. Increasing concentrations of greenhouse gases are projected to reduce the strength of the AMOC in most general circulation models (GCMs), although the fraction of weakening varies from model to model (Randall et al. 2007; Gregory et al. 2005). A weakening of the AMOC would reduce the northward transport of heat in the Atlantic Ocean and lead to a relative cooling of the North Atlantic and surrounding landmasses (Stouffer et al. 2006). Even multidecadal fluctuations of the AMOC can have significant impacts on Northern Hemisphere climate variability (Vellinga and Wu 2004; Frankcombe et al. 2010; Knight et al. 2005).
Understanding the internal variability of the AMOC and the processes that control it is important in order to distinguish whether any observed changes are forced, or part of the natural variability of the climate system. This knowledge is necessary for the attribution of anthropogenic weakening of the AMOC, as well as for the development of methods to detect any changes (Roberts and Palmer 2012). The understanding of which processes and locations control the internal variability of the AMOC, both in the models and in the real world, can also assist with targeting observations and improving decadal prediction systems. An improved understanding of the processes and feedbacks that play a role in the internal variability of the AMOC can also shed light on processes that might reinforce, damp, or modulate larger, anthropogenically forced AMOC changes.
Many previous studies have shown the internal variability of the AMOC in various GCMs from intraseasonal to centennial time scales. A number of these studies have related decadal and multidecadal AMOC variability to low-frequency variability of the North Atlantic Oscillation (NAO) that alters the advection of freshwater and heat in the ocean to the deep-water formation regions (Delworth and Greatbatch 2000; Dong and Sutton 2005; Dai et al. 2005; Danabasoglu 2008). The variability attributed to the NAO in these studies occurs on times scales of 20–35 yr, apart from the study of Delworth and Greatbatch (2000), which differs in that their model used flux adjustment. Lower-frequency multidecadal to centennial variability is also described in various studies of the AMOC in GCMs. Several studies attribute this to the export of salinity anomalies from the Arctic (Delworth et al. 1997; Jungclaus et al. 2005; Hawkins and Sutton 2007; Jahn et al. 2010), though mechanisms for the Arctic variability differ, and Delworth et al. (1993) identified a different mechanism from Delworth et al. (1997) for the same variability. A different mechanism was proposed by Vellinga and Wu (2004, hereafter VW) involving feedbacks from the tropical Atlantic. They found that increases (decreases) in the AMOC change the equatorial sea surface temperature (SST) gradient, which, through changes to the intertropical convergence zone (ITCZ) and its associated precipitation, cause a fresh (saline) anomaly in the tropical and subtropical North Atlantic. This fresh anomaly is advected northward to deep-water formation regions in the North Atlantic, where it acts as a negative feedback on the AMOC strength. This mechanism has also been demonstrated in other GCMs (Menary et al. 2011; Mignot and Frankignoul 2009).
In this study, the multidecadal to centennial variability of the AMOC in the third version of the Hadley Centre Coupled Model [HadCM3; Pope et al. (2000); Gordon et al. (2000)] is revisited. VW examined the multidecadal variability of the AMOC and found significant variability at 70–200-yr periods. They related this variability to salinity changes around 60°N that were generated by the VW mechanism discussed above. Hawkins and Sutton (2007) also examined the AMOC in HadCM3 and proposed a different explanation for the multidecadal variability. They also connected salinity changes in the subpolar North Atlantic [particularly in the Greenland–Iceland–Norwegian (GIN) seas] with the multidecadal variability of the AMOC; however, they related these salinity changes to changes in the freshwater exported from the Arctic and proposed that the Arctic controls the AMOC variability through a storage and release cycle of freshwater. Pardaens et al. (2008) also note the variability in the freshwater exported from the Arctic. The role of the GIN seas as the source of AMOC variability is also supported by Hawkins and Sutton (2008), who see large multidecadal changes in the AMOC as a result of changes in density in the GIN seas, and hence the increased flow through the Denmark Strait’s overflow.
The purpose of this study is to 1) reconcile and further understand these two sources of variability and 2) investigate their sensitivity by using a multiple perturbed physics ensemble. Section 2 presents the models and methods used in this paper. Section 3 discusses the AMOC variability in HadCM3, and the role of the Arctic and equatorial Atlantic. A simple prediction model for the AMOC variability is also presented. The sensitivity of these results to model physics is investigated in section 4 using a multiple perturbed physics ensemble. Section 5 presents a discussion of the results and conclusions.
2. Methods
a. Models
HadCM3 is a global coupled GCM with resolutions of 2.5° × 3.75° and 1.25° × 1.25° in its atmosphere and ocean components, respectively. Ocean parameterizations include the Visbeck et al. (1997) version of the Gent and McWilliams (1990) thickness diffusion scheme, the Redi (1982) along-isopycnal diffusion scheme, and a modified convective adjustment scheme for North Atlantic overflows (Roberts et al. 1996). HadCM3 has some known shortcomings typical of its generation of climate models that might influence the mechanisms described in this paper. They include an Arctic polar island (Connolley et al. 2006), no barotropic flow through the Bering Strait, no throughflow through the Canadian Arctic Archipelago and Davis Strait, and issues related to the resolution of the North Atlantic overflows (Roberts et al. 1996; Roberts and Wood 1997). Despite these shortcomings, HadCM3 produces a reasonable surface climate and simulates the strength of the AMOC and ocean heat transport well compared with observations without the need for flux adjustments (Gordon et al. 2000). An advantage of using a GCM of this resolution and complexity is that its speed makes conducting multi-thousand-year simulations possible, allowing low-frequency climate variability to be studied. It also allows the sensitivity to perturbing physics to be examined using multicentury ensembles.
Results are presented here from a 5000-yr simulation of HadCM3 with constant greenhouse gas and aerosol forcing. Unless otherwise specified, data shown are bandpass-filtered, annual mean data. A Chebyshev filter with cutoffs of 50 and 300 yr is used to capture the multidecadal variability identified by VW.
Section 4 also presents results from a perturbed physics ensemble (PPE). This is an ensemble based on HadCM3, where parameters in the atmosphere, ice, and land models are varied to investigate uncertainty associated with these processes. Multiple parameters are perturbed for each ensemble member, with perturbations to components including atmospheric dynamics, cloud physics, convection, radiation, boundary layer physics, land surface, and sea ice albedo. Flux adjustment, which can have an unphysical impact on deep-water formation, a critical process for the AMOC (Marotzke and Stone 1995), is not used in the ensemble, nor in the standard HadCM3. Most ensemble members only have 140 yr of control simulation, which is not sufficient for examining multidecadal to centennial variability. Eight members have longer (~300–400 yr) control simulations, which are compared with the standard control run in section 4. The ensemble is described more fully in Vellinga and Wu (2008) with the eight members considered being columns (2, 5, 6, 9, 13, 16, 20, and 22) from Tables B1 and B2 of that study. An examination of the climates and AMOCs of the different ensemble members is presented in Jackson et al. (2011).
b. Significance tests
Within this paper, when calculating correlations and regressions against time series, the statistical significance is calculated using a moving-block bootstrap (Wilks 1997). This takes into account autocorrelation in time, which is especially important when considering filtered time series. A confidence limit of 95% is used.
c. PPE significance
To understand whether results from the PPE differ significantly from the standard HadCM3 ensemble, a method of randomly sampling the long standard time series is used.
Since one member only has 290 yr, the standard deviations are calculated from the first 290 yr of each PPE member (though results do not change qualitatively when using all available years). Segments of 290 yr are randomly selected (allowing overlapping segments) from the 5000-yr-long standard time series. These are processed identically to the time series for the PPE (first apply a bandpass filter and then take the standard deviation). A distribution of the quantity of interest is then created from these results and the limits enclosing 95% of the data are found.
3. AMOC in HadCM3
The AMOC streamfunction in HadCM3 in shown in Fig. 1a; the maximum time-mean strength of the AMOC in the Atlantic is 17.9 Sv (1 Sv ≡ 106 m3 s−1) around 35°N at 1000-m depth. On the multidecadal to centennial time scales examined in this paper there is basin-wide variability, with standard deviations of up to 0.5 Sv in the North Atlantic (Fig. 1b). An index M is chosen to measure the variability of the AMOC at 50°N and 1000-m depth, which is where the variability is the largest. The first 1000 yr of the 5000-yr time series are shown in Fig. 1c (gray) with the low-pass-filtered time series overlain (black). Large changes in the AMOC index of 2–4 Sv over 30–100 yr are clearly visible and are picked out by the filtered time series. In some periods such as that shown in Fig. 1c, this variability is intermittent. However in other periods (such as years 2000–2800; see Fig. 8), this AMOC variability is more regular. These changes are also seen at lower latitudes; however, there is also additional, localized variability at lower latitudes that is found to be related to wind forcing (not shown). There is a slight drift over the first 1000 yr of the time series, associated with the long spinup of the ocean. The AMOC index used is bandpass filtered to remove this long-term trend as well as the higher-frequency variability in order to concentrate on multidecadal to centennial time scales.
(a) Time-mean (of full 5000 yr) AMOC streamfunction (Sv). (b) Std dev of the low-pass-filtered AMOC streamfunction (Sv). (c) Time series of AMOC strength (Sv) for the first 1000 yr of the time series considered. Shown are annual means (gray) and the low-pass-filtered time series (black).
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
a. Role of salinity
Hawkins and Sutton (2008) examined several large, multidecadal changes of the AMOC from HadCM3 and identified density changes in the GIN seas as being the driving factor. Figure 2 shows the simultaneous correlation of the AMOC index M with the salinity averaged over the top 535 m, which shows highly significant correlations in the GIN seas and the mid–North Atlantic. Lagged correlations (not shown) show that the relationships are strongest preceding a change in the AMOC in the GIN seas, and 5 yr after changes in the AMOC in the mid-Atlantic, suggesting that the latter may be a response to changes in the AMOC. Note also that there are regions of significant negative correlations in the Arctic and in the Caribbean Sea occurring simultaneously with the AMOC changes. The Arctic changes will be discussed in section 3b.
Correlation (with no lag) of salinity (averaged over the top 535 m) with the AMOC index. The white box indicates the region for calculating SGIN (see Table 1). Only regions that are considered significant (see section 2b) are shaded.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
A GIN seas salinity index SGIN is defined as the average salinity in the top 535 m in the region marked by the white box in Fig. 2. There is a strong and significant simultaneous correlation between SGIN and M (0.73). Temperature changes lag those caused by salinity and are also highly correlated with the AMOC, though their effects on density are much smaller and act in the opposite sense to salinity changes. These results suggest that changes in density (caused by changes in salinity) drive sinking and convection in the GIN seas, which drive the large changes in the AMOC. In HadCM3, deep-water formation and convection in the GIN seas are associated with AMOC changes, with less deep-water formation in the Labrador Sea than in the observations. Increased convection in the GIN seas would bring warmer subsurface waters to the surface, which would explain the lagged negative correlation with temperature.
This picture is also supported by the good correspondence between the Denmark Strait’s outflow (not shown) and both the AMOC index and SGIN. The Denmark Strait overflow (defined as the volume flux denser than 1027.8 kg m−3 between Greenland and Iceland) has a significant simultaneous correlation of 0.8 with M, supporting the relationship seen in Hawkins and Sutton (2008) for large changes in AMOC strength. It should be noted that HadCM3 has artificially deepened overflow channels to improve the volume flux by the overflows.
To explain how these salinity anomalies form, an understanding of the water masses in the GIN seas is important. The North Atlantic Current brings relatively warm, salty water from the Atlantic, through the east section of the GIN seas, and then through the Fram Strait (between Greenland and Svalbard), though some passes south of Svalbard into the Barents Sea. The second important water mass is from the East Greenland Current (EGC), which brings very cold and fresh water out of the Arctic and around the east coast of Greenland in the surface waters. These two water masses mix across the GIN seas leading to profiles with colder, fresher surface water overlaying warmer, saltier subsurface water. In HadCM3 this results in cold freshwater in the upper 200 m and warmer, saltier water from 200 to 1000 m. Salinity anomalies could be advected into this region by either pathway, or could be generated locally. A budget analysis (not shown) indicates that advection plays the dominant role in the salinity changes in the GIN seas on multidecadal time scales, with other processes (e.g., surface fluxes, ice formation and melting, and mixing) having weaker influences.
Time–salinity plots indicate movement of positive and negative salinity anomalies from both the Arctic (via the EGC) and the subtropical Atlantic (see Fig. 3 in Pardaens et al. 2008). VW investigated these subtropical anomalies using tracers to illustrate this advective pathway: the tracer propagates from the surface equatorial water to the subtropical gyre where it recirculates, spreading through the gyre and being subducted to subsurface waters. The tracer then spreads northward along the coast of Europe, entering the GIN seas and Arctic. The tracer takes ~50 yr to reach the GIN seas, and the time scale for the entire VW mechanism is ~110 yr (Menary et al. 2011).
b. Arctic anomalies
There have been various observational and model studies examining the variability of salinity in the Arctic, mainly on seasonal to decadal time scales (Proshutinsky and Johnson 1997; Proshutinsky et al. 2002; Dukhovskoy et al. 2006; Polyakov et al. 2008), and the export of these anomalies to the Atlantic via the EGC (Bigg and Wadley 2007). Other model studies have seen salinity anomalies originating from the Arctic driving variability in the AMOC (Jungclaus et al. 2005; Delworth et al. 1997; Hawkins and Sutton 2007; Jahn et al. 2010). Theories for the Arctic salinity variability have suggested an ocean mode forced by the atmosphere or a coupled ocean–atmosphere mode, and have involved various combinations of a stochastically forced Arctic salinity mode (Frankcombe et al. 2010; Frankcombe and Dijkstra 2010), forcing by sea level pressure anomalies over the Beaufort Sea (Delworth et al. 1997; Proshutinsky et al. 2002; Proshutinsky and Johnson 1997; Jahn et al. 2010), precipitation anomalies over the Arctic (Dukhovskoy et al. 2006), and feedback from atmospheric heat transports (Dukhovskoy et al. 2006).
Previous studies have shown an export of salinity anomalies from the Arctic to the GIN seas in HadCM3 (Pardaens et al. 2008; Hawkins and Sutton 2007). The latter postulated that salinity anomalies in the Arctic are a lagged response to the AMOC, with the Arctic acting to store and release the anomalies, which then feed back onto the AMOC.
An analysis of the salinity budget of the Arctic (not shown) reveals that the salinity anomaly averaged over the entire Arctic over 0–535-m depth SARC (see Table 1) is nearly entirely driven by advective changes, with a small lagged contribution from sea ice processes. Contributions from surface freshwater fluxes are much smaller. The advection in HadCM3 is almost exclusively through the Fram Strait, with little transfer through the Bering Strait and between Svalbard and Norway, where the water is shallow (note that straits joining the Labrador Sea directly to the Arctic are closed in HadCM3). The total advection VS between the GIN and Arctic seas will be referred to in this study as that through the “Fram Strait,” although it comprises not only that through the actual Fram Strait (78.75°N, 20.625°W–23.125°E) but also a small amount through the passage between Svalbard and Norway (71.25°–79.75°N, 23.125°E).
Definition of indices.
Figure 3a shows the rate of change of total salinity in the Arctic 
(a) Time series (106 psu m3 s−1) of first 1000 yr of 
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
To understand the origin of the Arctic variability, the development of salinity anomalies within the Arctic is shown in the top two rows of Fig. 4. Thirty years before a positive anomaly in SEGC, the Beaufort gyre (in the center of the Arctic basin) is more saline. The saline anomaly moves toward the Greenland–Canadian coast at lag −10 and is exported out of the Fram Strait via the EGC, around Greenland and into the Labrador Sea, before recirculating in the subpolar gyre. At the same time a fresh anomaly starts developing on the Siberian side of the Beaufort gyre at lag 0. Frankcombe et al. (2010) find similar spatial patterns of salinity anomalies in the Arctic (their Fig. 11) with a multidecadal time scale in another GCM that participated in the Coupled Model Intercomparison Project 3 (Geophysical Fluid Dynamics Laboratory Climate Model, version 2.1). They relate the salinity variability to “saline Rossby waves” induced by the background salinity gradient and initiated by the stochastic variability in the inflow from the Atlantic.
Lagged correlations of salinity (averaged over the top 535 m) and SLP with the salinity exported from the Arctic, SEGC. Years before changes in SEGC are shown above each panel. Only regions that are considered significant (see section 2b) are shaded.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
The Beaufort gyre is a strong, anticyclonic ocean gyre in the Arctic basin. It is geostrophically balanced, with low salinity and density in the center of the gyre and higher salinity and density near the coasts. This geostrophic balance is modified ageostrophically by Ekman pumping, in analogy to the wind-driven subpolar and subtropical gyres. Since in the Beaufort gyre the currents do not meet any boundaries to dissipate vorticity input by the wind, the current becomes deep and interacts with topography. Hence, inertia becomes important.
In HadCM3, there is a clear relationship between SEGC anomalies and sea level pressure (SLP) over the Arctic (see bottom row in Fig. 4). A similar picture is also seen in Fig. 5a, where the salinity in the top 535 m is correlated with a Beaufort gyre SLP index (averaged over 70°–90°N, 90°–270°E) showing high instantaneous correlations with the salinity in the EGC. When there is a high sea level pressure anomaly over the Arctic basin, an anticyclonic wind stress anomaly is generated around the Beaufort gyre, which leads to downwelling in the center of the basin and upwelling at the coasts (Fig. 5c). Since the surface water is colder and fresher than subsurface water, this leads to a freshening in the center of the gyre and salinification near the coasts (Fig. 5b). Hence, at lag 0 there is a saline anomaly north of Greenland (from a Beaufort gyre saline anomaly generated ~30 yr previously) that is further salinified by the present coastal, wind-driven upwelling. This is a mechanism by which multidecadal variability in SLP could drive the salinity pattern in the Arctic Ocean and hence the multidecadal variability in the salinity exported from the Arctic. It appears unlikely that the SLP is responding to the variability in the Arctic Ocean (rather than driving it) since relationships between Arctic surface air temperature (SAT) and the SLP index are weak and strongest when SAT lags SLP (not shown).
Simultaneous correlations of (a) salinity (averaged over the top 535 m), (b) rate of change of salinity (averaged over the top 535 m), and (c) wind stress curl, with a Beaufort gyre SLP index (average of 70°–90°N, 90°–270°E). Only regions that are considered significant (see section 2b) are shaded.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
Although the multidecadal variability in SLP is correlated with the salinity anomalies, a spectral analysis of the annual mean and winter Arctic Oscillation (first principal component of the empirical orthogonal functions of Northern Hemisphere SLP variations) and SLP over the Beaufort gyre show red noise spectra with no significant spectral peak at low frequencies (not shown). The Arctic salinity, on the other hand, has very significant power on multidecadal time scales, particularly in the Beaufort gyre, where the filtered time series accounts for 79% of the amplitude of the unfiltered time series. Proshutinsky et al. (2002) describe how the Beaufort gyre accumulates freshwater anomalies and acts as a “flywheel.” This could allow the low-frequency part of the spectrum of stochastic SLP variability to induce significant multidecadal variability in the Arctic in HadCM3.
The Arctic variability in this study appears to be similar to the ocean mode discussed in Frankcombe et al. (2010) and Frankcombe and Dijkstra (2010); however, in those studies the variability was seen in subsurface (400-m depth) salinity and they proposed stochastic forcing from the Atlantic inflow. In HadCM3, the salinity anomalies can be seen at the surface and may be induced by anomalous upwelling or downwelling, and the associated Ekman transport, caused by anomalously high or low SLP over the Arctic. It should be noted that the sea ice model incorporated in HadCM3 allows wind stresses to communicate directly with the ocean surface, with sea ice being advected by ocean currents. More sophisticated representations of sea ice dynamics could alter the response of the ocean to changing wind stresses and may affect this mechanism. Nevertheless, multidecadal variability of freshwater in the Arctic has been seen in observations and other modeling studies (Polyakov et al. 2008; Frankcombe et al. 2010).
c. Local feedbacks
Salinity anomalies generated in the Arctic and Caribbean are eventually advected into the GIN seas where the resulting salinity anomaly is modified by local feedbacks. A positive salinity anomaly preconditions the GIN seas for greater wintertime convection, and since the subsurface water is warmer and more saline than the surface water, convection will act to heat and salinify the surface water (which is a positive feedback on the salinity). This heat is lost to the atmosphere (Fig. 6a), leading to a local decrease in sea level pressure (Fig. 6b) and cyclonic wind anomalies over the GIN seas. Positive wind stress curl (Fig. 6c) over the center of the basin acts to drive more Ekman upwelling, bringing more saline subsurface water to the surface and again acting as a positive feedback on the initial salinity anomaly. There is also a strengthening of the cyclonic circulation and upper-ocean velocities (Fig. 6d), which could affect the convergence of salinity and ice from advection into the Greenland Sea region. The increased velocity in the south of the GIN seas would advect more salty water from the North Atlantic into the region, providing the positive feedback noted in Hawkins and Sutton (2007). There is also anomalous southward flow through the center of the Fram Strait. The net effect of these velocity changes is a positive feedback on the salinity in the GIN seas (see Fig. 3b, dotted line).
Regression coefficients from simultaneous regressions with SGIN, scaled with one standard deviation of SGIN: (a) SAT (°C), (b) SLP (hPa), (c) wind stress curl (N m−3) with wind stress vectors (N m−2), and (d) barotropic streamfunction (Sv) with velocity (depth integrated over the top 535 m) vectors (m2 s−1). Only regions that are considered significant (see section 2b) are shaded and plotted.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
d. Lagged regression model
The previous sections have detailed how salinity anomalies drive multidecadal variability in the AMOC. These processes are illustrated by the lagged correlations of salinity anomalies SCRB, SEGC, and SGIN (see Table 1) with the AMOC index M in Fig. 7a. There are positive salinity anomalies in the Caribbean, EGC, and GIN seas of approximately 45, 20, and 5 yr, respectively, before an AMOC maximum. This is consistent with the picture of salinity anomalies from the Caribbean and EGC being advected to the GIN seas where they drive changes in the AMOC. At 10 yr after the AMOC maximum there is a fresh anomaly seen in the Caribbean. This is consistent with the negative feedback seen in the VW mechanism, and also with the time scale of ~55 yr for half a period. A weak, but significant, signal is also seen in SEGC between ~10 and 50 yr after the positive SGIN anomaly. When there is an export of salinity from the Arctic to the GIN seas, it results in freshening the Arctic as well as salinifying the GIN seas, resulting in a future fresh anomaly in SEGC. The broad range of lags over which this negative feedback acts reflects different responses for different sections of the time series. In particular, responses of SEGC to AMOC changes are not significant during periods of intermittent AMOC variability (e.g., first 1000 yr; Fig. 7b), but are during periods of regular multidecadal variability (e.g., years 4000–5000; Fig. 7c). In all periods, the relationships between the AMOC changes and the salinity signals that precede them remain relatively robust, suggesting similar mechanisms for initiating the AMOC changes during both intermittent and regular AMOC variability.
Lagged correlations of AMOC index M with SGIN (solid), SEGC (dashed), and SCRB (dot–dash) for (a) full 5000-yr time series, (b) years 0–1000, and (c) years 4000–5000. Thin dotted lines show correlations that are considered significant for all three time series (see section 2b). The AMOC leads for positive lags.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
A very good degree of correspondence can be seen between the AMOC index and the modeled index (Fig. 8) with the simultaneous correlation between the time series over the final 2000 yr being significant (0.68). The coefficients are found to be A = 16.0 and B = 22.4 Sv psu−1, suggesting a greater sensitivity to anomalies originating in the tropical Atlantic. If the model instead fits SGIN (rather than M) to SEGC and SCRB, then it will indicate the magnitude of the salinity anomaly in the GIN seas compared with the anomalies in the EGC and Caribbean. The magnitudes obtained are A = 1.7 and B = 2.2, which (since both are >1) suggests an amplification of the anomalies, possibly as a result of positive local feedbacks seen in section 3c.
Time series of AMOC index M (Sv) for the 5000 yr of data considered (solid) with modeled index M′ (dashed).
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
e. Sensitivity to time scales
Wavelet analysis of VW identifies significant power in the variability of the AMOC at 70–200-yr periods and intermittent variability at 10–20-yr periods. Power spectra (not shown) confirm this finding, with significant variability [compared with the 0.05 level of an autoregressive (AR1) process] between 80 and 250 yr, with a particularly large signal at around 120 yr that is also seen in the power spectra of SGIN. Around 90 yr, SEGC has significant power (and less significantly at 120 yr) and also significant decadal variability at ~12–25 yr. In addition, SCRB has significant interannual and decadal variability, but when this is removed by using a decadally averaged time series, significant variability is also seen at 110–200 yr.
Analyses of the spectra support the use of the cutoff periods of 50 and 300 yr employed for bandpass filtering. Sensitivity to these cutoff periods was also assessed for the time series in the linear model. Changing the high-frequency cutoff from 50 to 70 yr does not qualitatively change the results. Reducing it to 40 yr and below does increase the influence of SEGC while increasing the error E, possibly as a result of including the significant variability in SEGC at 12–25-yr periods. Results are also insensitive to increasing the low-frequency cutoff from 300 yr, though reducing this cutoff filters out some of the significant AMOC variability and reduces the influence of SCRB.
4. A perturbed physics ensemble
This, and previous studies, have highlighted a number of processes that play important roles in the multidecadal variability of the AMOC in HadCM3. There are many differences across GCMs in both the frequencies and mechanisms of AMOC variability, although there are also some similarities as discussed in the introduction. To investigate the sensitivity of the variability discussed here to differing model physics, a PPE is used. This allows us to gain an understanding of the sensitivity to physics perturbations within one structural model where processes are more likely to be similar, rather than across a multimodel ensemble where structural differences are also likely to play a role. This can allow a greater understanding of how different climate states and the representation of processes can affect variability and, hence, inform multimodel comparisons.
The PPE is based on HadCM3, but with perturbations to the atmospheric, land surface, and sea ice physics. Note that the ocean is identical across the ensemble, so that any differences are caused by differences in surface fluxes and winds. Since multiple parameter perturbations are prescribed for each ensemble member, the small number of members makes it impossible to attribute any differences to specific perturbations. Also, because the PPE simulations are relatively short (290–390 yr) for examining multidecadal to centennial variability, we cannot repeat the analysis that we carried out for the HadCM3 control. Instead, we investigate whether there are any statistically significant differences from the standard HadCM3 and, if so, what can be learned from these differences.
Figure 9 shows the bandpass-filtered AMOC time series for the eight PPE members with the longest time series (solid line), with Fig. 10a showing the standard deviation of these time series. This gives an indication of the magnitude of the multidecadal variability of the AMOC. The letters show the standard deviation for each of the PPE members (y axis is arbitrary). Also plotted is the distribution of the equivalent measure in the standard 5000-yr HadCM3 run (see section 2c), along with the 95% limits (dashed lines). The PPE members x and q have larger multidecadal variability of the AMOC than any 290-yr segment of the standard model. This large variability can also be seen by inspecting the time series (Figs. 8 and 9). Both ensemble members also have large variability in SGIN and SEGC, although only x has significantly larger variability in SEGC (Fig. 10a) and SBFT (not shown). Member q has significantly larger variability than the standard HadCM3 in the Irminger Sea, which may account for additional sinking and the large AMOC variability.
Bandpass-filtered time series of AMOC anomalies (Sv) for eight members of the PPE (solid). Also shown are AMOC strengths constructed from the model discussed in section 3d (dashed) and those constructed by allowing lags to vary, but not coefficients (dotted). See text for more details.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
(a) Letters show the value of the std dev of 290 yr of the filtered AMOC time series (a measure of the magnitude of multidecadal variability) in the PPE (y position is arbitrary). Also plotted is a frequency distribution of the std dev of randomly selected 290-yr segments from the standard HadCM3 control (solid line). Dashed lines show 95% limits. Other panels show the same, but for (b) SGIN, (c) SEGC, and (d) SCRB.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
Analysis of the spectra of the unfiltered time series of these ensemble members (not shown) suggests that over half of the ensemble (e, f, q, x, and z) have significant variability at around 100 yr, with the remainder only having significant variability at higher frequencies. Although there are also periods where the standard HadCM3 has little low-frequency variability (e.g., years 1000–1400 in Fig. 8), it is possible that the physics perturbations have altered the climate and significantly reduced the magnitude of the low-frequency variability in the remaining ensemble members. Two ensemble members that do not show low-frequency variability (b and m) have significantly less variability (as measured by the standard deviation of the filtered time scale) than the standard model in SGIN or SEGC (Figs. 10b and 10c), and in SBFT (not shown). Although they do show multidecadal variability in SCRB, which appears to lead to an AMOC response, this is insufficient to generate a low-frequency spectral peak. These two are also the coldest members with the greatest concentrations of sea ice (Fig. 11a). Ensemble members with greater global mean temperatures than b and m have less sea ice (e.g., member x, which is slightly warmer; shown in Fig. 11b). This sea ice generally forms around the Arctic coast in winter, from where it is advected into the center of the Beaufort gyre by the predominantly anticyclonic winds, and out through the Fram Strait. The seasonal pattern of ice formation and melting results in the pattern of annual mean freshwater flux from ice to ocean seen in Fig. 11d: there is a net salinification near the coasts in the Arctic and a net freshening in the center of the Beaufort gyre and south of the Fram Strait. This strengthens the salinity gradient across the Beaufort gyre, which strengthens the gyre circulation through geostrophic balance (Fig. 11f). In the two coldest members (b and m) the summertime sea ice is much more widespread and covers the whole Arctic. There is a greater export of sea ice through the Fram Strait with more summertime ice melt in the Atlantic and GIN seas, and less in the Beaufort gyre. This lack of freshening in the center of the gyre leads to a much weaker salinity gradient across the basin than in other ensemble members and, hence, a weaker gyre circulation (Figs. 11c and 11e). It should be noted that this relationship between the global temperature and salinity gradient or gyre strength is not seen across the remaining ensemble members. There appears to be a different regime for the coldest ensemble members, with a change in the pattern of freshwater flux from ice to ocean that may be controlled by the summertime sea ice extent.
(a),(b) Summertime sea ice concentration as a fraction. (c),(d) Annual mean change in salinity from sea ice formation and melting (1012 psu s−1) with negative regions indicating net freshening and positive regions indicating net salinification. (e),(f) Annual mean salinity (averaged over the top 535 m; psu) with contours overlain of the barotropic streamfunction (Sv). Contours are only plotted for positive values and are in steps of 2 Sv. Shown in (a),(c),(e) are results for ensemble member b, and in (b),(d),(f) for x.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
The mode of multidecadal variability in Arctic salinity identified by Frankcombe and Dijkstra (2010) is dependent on the salinity gradient across the Beaufort gyre, with a weaker salinity gradient leading to higher-frequency variability. This is possibly the cause of the small multidecadal variability of the Arctic salinity, and hence of the AMOC, seen in these ensemble members.
Another way of comparing the PPE with the standard HadCM3 simulation is by making use of the linear model described in section 3d. Figure 9 shows the AMOC strengths for each PPE member modeled using Eq. (2) with the coefficients and lags calculated from the standard model (dashed lines). A visual inspection reveals a good correspondence between the modeled index and the actual AMOC strength (solid lines) in some ensemble members (e.g., f), but less in others (e.g., i). This can be quantified by defining a weighted error function (see the appendix). Figure 12a compares the errors E between the AMOC time series and those predicted by the linear model for each of the ensemble members (letters) to the distribution of E across the standard model. Four of the ensemble members have errors greater than 95% of the distribution of the standard model, suggesting that this linear model is not successful for all members of the PPE. In some ensemble members, in particular q (Fig. 9), the linear model captures the magnitude of changes in the AMOC, but not the timing. This suggests that there may be a difference across the PPE in the lags between positive (negative) salinity anomalies in the EGC and the Caribbean, as well as increases (decreases) in the AMOC.
(a) A comparison of model errors E (see the appendix) for the PPE (letters; y position is arbitrary) against the frequency distribution of E in the standard HadCM3 (solid line). A total of 95% of the samples have a value of E less that that indicated by the dashed lines. (b) As in (a), but with variable lags (see text). Scatterplots of (c) lags between SCRB and M against control AMOC strength and (d) lags between SEGC and M plotted against global average surface air temperature for each of the PPE models. Gray letters in (d) indicate ensemble members that are not used for the correlation.
Citation: Journal of Climate 26, 7; 10.1175/JCLI-D-11-00601.1
To examine the effects of using different lags, the linear model in Eq. (2) was again used, but this time only fixing the coefficients to be the same as in the long standard model. The lags were chosen to minimize E for each ensemble member. The reconstructed AMOC time series are also plotted in Fig. 9 (dotted line), and are more successful at capturing the timing of large AMOC changes. Since lags were chosen to minimize the errors, it is not surprising that there is an improvement. If the samples from the standard model are treated in the same way (allowing the lags to vary), then these errors also decrease (and the distribution of E shifts toward zero); however, all the PPE members now have errors within the 95% limit for the standard model (Fig. 12b). This suggests that the prediction of AMOC variations from nonlocal salinity anomalies [from the linear model in Eq. (2)] does not give significantly different results when model physics is perturbed as long as differences in lags are taken into account and, hence, that the magnitude of the response of the AMOC to salinity anomalies is consistent.
Figures 12c and 12d compare the lags for SCRB and SEGC (t1, t0) with the mean climate for each PPE member. There is clearly a very good simultaneous correlation between AMOC strength (significant at the 0.05 level using the t test) and the lag for the Caribbean salinity anomaly. There is also a very good simultaneous correlation with global SAT, with colder ensemble members having stronger AMOC strengths (Jackson et al. 2011) and, hence, shorter lags. It might be expected that an increase in AMOC strength would decrease the time for an anomaly to be advected; however, the fourfold difference in lag time across the ensemble cannot be explained simply by changes in the AMOC strength. The meridional transport of salinity in the North Atlantic is also greater in the colder ensemble members (not shown) with transport by the AMOC dominating in the subtropics and that by the gyre dominating in the subpolar regions. Again, this is not a large enough difference to explain the large variation in lag, suggesting that other processes are involved. Vellinga and Wu (2004) show that the Caribbean salinity anomaly has a long residence time in the subtropical gyre, where it is subducted to subsurface waters, before being advected northward along the eastern boundary of the North Atlantic. Hence, the time scale for the anomaly to reach the deep-water formation regions is likely to be affected by vertical/cross-isopycnal mixing as well as advection. The density structure of the upper ocean shows a strong relationship to SAT in the ensemble, with the warmest ensemble members having warmer, and hence lighter, water in the upper layers of the subtropical gyre resulting in greater stratification than that in colder ensemble members. This greater stratification may act as a barrier to the subduction of salinity anomalies and hence increase the time taken for the anomalies to reach the deep-water formation sites. There are also some differences in the regions where deep-water formation occurs: the two coldest members for instance show the greatest wintertime mixed layer depths to the south of Iceland, rather than in the GIN seas. This might decrease the time taken for salinity anomalies from the subtropics to affect the AMOC.
There is initially little relationship between the lags for SEGC (salinity anomalies from the EGC) and SAT (see Fig. 12d). For two of the ensemble members (m and b), however, there is little variability in SEGC, which makes the definition of a lag more ambiguous. If these two members are neglected, then there is significance at the 0.1 level, but not at the 0.05 level. The lags could be affected by SAT through changes in the location of deep convection, or through the speed at which salinity anomalies are advected; however there are insufficient ensemble members to assess whether this relationship is robust.
5. Discussion
Multidecadal to centennial variability of the AMOC is seen in a 5000-yr control run of HadCM3. This variability is driven by the interaction of two mechanisms that cause changes in the salinity in the GIN seas. The first of these is a coupled ocean–atmosphere mechanism described in Vellinga and Wu (2004), whereby increases (decreases) in the AMOC change the equatorial Atlantic SST gradient, which shifts the ITCZ and its associated precipitation northward (southward). This causes a fresh (saline) equatorial salinity anomaly that propagates northward to regions of deep-water formation and has a negative feedback on the AMOC. The second source of salinity anomalies is the significant multidecadal variability in the export of freshwater through the Fram Strait. The geostrophic balance in the Beaufort gyre is modified ageostrophically by the wind stresses associated with low-frequency SLP variability. The proposed mechanism for the variability seen in the Arctic salinity in HadCM3 is that when there is an anticyclonic (cyclonic) SLP anomaly this generates anomalous upwelling (downwelling) near the coast and downwelling (upwelling) near the center of the Beaufort gyre, resulting in the export of saline (fresh) anomalies from the Arctic. The importance of both of these mechanisms for generating salinity anomalies is reflected in the success of a simple, lagged model that predicts AMOC changes in HadCM3 from salinity anomalies generated by the two mechanisms.
The sensitivity of these processes is examined using a small ensemble based on HadCM3 where model parameters relating to physical processes are varied. Results from the lagged model suggest that the AMOC consistently responds to salinity changes from the Arctic and Caribbean sources across the ensemble, although there are differences in the response time of the AMOC. The different lags for the Caribbean source are related to the background climate state, possibly through changes in stratification and meridional velocities that affect the propagation time. Some ensemble members show significantly greater multidecadal variability of the AMOC than in the standard HadCM3, which appears to be caused by the strong variability of the Arctic salinity source. This Arctic variability is significantly reduced in other members. Those members are the coldest and appear to be in a different regime, possibly controlled by the summertime sea ice extent, which appears to result in those members having little Arctic salinity and AMOC variability on multidecadal time scales. This suggests that it is the Arctic salinity anomalies that are important for generating the multidecadal AMOC variability, with Caribbean anomalies playing a role when this variability is present.
Although this study has only considered models related to HadCM3, there have been previous model studies linking multidecadal to centennial AMOC variability in different GCMs to both the salinity anomalies from the Arctic (Jungclaus et al. 2005; Delworth et al. 1997; Jahn et al. 2010) and shifts in the ITCZ (Menary et al. 2011). Although HadCM3 has deficiencies in the Arctic that might affect the processes described, similar variability is also seen in more sophisticated models with better topographical and physical representations (Frankcombe et al. 2010). Multidecadal variability of a ~40–80 yr period is also seen in observations and paleoclimate records, particularly in temperatures in the Atlantic, where it is referred to as the Atlantic Multidecadal Oscillation or variability (Knight et al. 2005; Zanchettin et al. 2010), but also in salinity (Polyakov et al. 2005). There are also indications of similar multidecadal variability in the Arctic (Polyakov et al. 2003, 2004) and Caribbean (Kilbourne et al. 2008). Although the time scales of the variability are slightly longer in this study than observed, aspects of the mechanisms discussed in this paper may play a role in the real climate.
An improved understanding of the mechanisms behind multidecadal variability has implications for detecting, attributing, and predicting changes in the AMOC. The importance of salinity in the GIN seas for multidecadal variability in HadCM3 has been demonstrated here. Although variability in the real ocean may depend on sinking in different locations, this study shows the importance of capturing salinity changes in the deep-water formation regions, as well as highlighting the possible roles of the Arctic and equatorial Atlantic. The dependence of the strength and presence of multidecadal AMOC variability on the mean climate state suggests that achieving a better mean climate in GCMs may improve our predictions of AMOC changes. It also has implications for understanding variability across different GCMs that may have different climate biases.
Acknowledgments
The authors gratefully acknowledge J. Ridley for useful suggestions and discussions. We also thank three anonymous reviewers, whose comments helped improve the paper. This work was supported by the Joint DECC/Defra Met Office Hadley Centre Climate Programme (GA01101). The authors acknowledge help from B. Booth, G. Harris, and D. Sexton in setting up and running the PPE.
APPENDIX
Linear Model
This search is done by using a stepwise regression technique. First, possible factors are defined as lagged salinities [see Eq. (2)] with possible lags for SEGC ranging from 0 to 30 yr before a change in the AMOC, and lags for SCRB ranging from 0 to 62 yr before. Then, a combination of forward selection and backward elimination are used alternately to select significant factors and eliminate redundant factors. Significance is tested using an f test at the 0.01 level. See von Storch and Zwiers (1999, 166–167) for further details.
REFERENCES
Bigg, G. R., and M. R. Wadley, 2007: Simulation of “great salinity anomalies” in coupled climate models. J. Geophys. Res., 112, C03017, doi:10.1029/2006JC003475.
Connolley, W., A. Keen, and A. McLaren, 2006: Results from the implementation of the elastic viscous plastic sea ice rheology in HADCM3. Ocean Sci., 2, 201–211.
Dai, A., A. Hu, G. Meehl, W. Washington, and W. Strand, 2005: Atlantic thermohaline circulation in a coupled general circulation model: Unforced variations versus forced changes. J. Climate, 18, 3270–3293.
Danabasoglu, G., 2008: On multidecadal variability of the Atlantic Meridional Overturning Circulation in the Community Climate System Model version 3. J. Climate, 21, 5524–5544.
Delworth, T., and R. Greatbatch, 2000: Multidecadal thermohaline circulation variability driven by atmospheric surface flux forcing. J. Climate, 13, 1481–1495.
Delworth, T., S. Manabe, and R. Stouffer, 1993: Interdecadal variations of the thermohaline circulation in a coupled ocean–atmosphere model. J. Climate, 6, 1993–2011.
Delworth, T., S. Manabe, and R. Stouffer, 1997: Multidecadal climate variability in the Greenland Sea and surrounding regions: A coupled model simulation. Geophys. Res. Lett., 24, 257–260.
Dong, B., and R. Sutton, 2005: Mechanism of interdecadal thermohaline circulation variability in a coupled ocean–atmosphere GCM. J. Climate, 18, 1117–1135.
Dukhovskoy, D., M. Johnson, and A. Proshutinsky, 2006: Arctic decadal variability from an idealized atmosphere–ice–ocean model: 2. Simulation of decadal oscillations. J. Geophys. Res., 111, C06029, doi:10.1029/2004JC002820.
Frankcombe, L., and H. Dijkstra, 2010: Internal modes of variability in the Arctic Ocean. J. Phys. Oceanogr., 40, 2496–2510.
Frankcombe, L., A. von der Heydt, and H. Dijkstra, 2010: North Atlantic multidecadal climate variability: An investigation of dominant time scales and processes. J. Climate, 23, 3626–3638.
Gent, P., and J. McWilliams, 1990: Isopycnal mixing in ocean circulation models. J. Phys. Oceanogr., 20, 150–155.
Gordon, C., C. Cooper, C. Senior, H. Banks, J. Gregory, T. Johns, J. Mitchell, and R. Wood, 2000: The simulation of SST, sea ice extents and ocean heat transports in a version of the Hadley Centre coupled model without flux adjustments. Climate Dyn., 16, 147–16.
Gregory, J. M., and Coauthors, 2005: A model intercomparison of changes in the Atlantic thermohaline circulation in response to increasing atmospheric CO2 concentration. Geophys. Res. Lett., 32, L12703, doi:10.1029/2005GL023209.
Hawkins, E., and R. Sutton, 2007: Variability of the Atlantic thermohaline circulation described by 3D EOFs. Climate Dyn., 29, 745, doi:10.1007/s00382-007-0263-8.
Hawkins, E., and R. Sutton, 2008: Potential predictability of rapid changes in the Atlantic Meridional Overturning Circulation. Geophys. Res. Lett., 35, L11603, doi:10.1029/2008GL034059.
Jackson, L., M. Vellinga, and G. Harris, 2011: The sensitivity of the Meridional Overturning Circulation to modelling uncertainty in a perturbed physics ensemble without flux adjustment. Climate Dyn., 39, 277–285, doi:10.1007/s00382-011-1110-5.
Jahn, A., B. Tremblay, L. Mysak, and R. Newton, 2010: Effect of the large-scale atmospheric circulation on the variability of the Arctic Ocean freshwater export. Climate Dyn., 34, 201–222.
Jungclaus, J., H. Haak, M. Latif, and U. Mikolajewicz, 2005: Arctic–North Atlantic interaction and multidecadal variability of the meridional overturning circulation. J. Climate, 18, 4013–4031.
Kilbourne, K., T. Quinn, R. Webb, T. Guilderson, J. Nyberg, and A. Winter, 2008: Paleoclimate proxy perspective on Caribbean climate since the year 1751: Evidence of cooler temperatures and multidecadal variability. Paleoceanography, 23, PA3220, doi:10.1029/2008PA001598.
Knight, J., R. Allan, C. Folland, M. Vellinga, and M. Mann, 2005: A signature of persistent natural thermohaline circulation cycles in observed climate. Geophys. Res. Lett., 32, L20708, doi:10.1029/2005GL024233.
Marotzke, J., and P. Stone, 1995: Atmospheric transports, the thermohaline circulation, and flux adjustments in a simple coupled model. J. Phys. Oceanogr., 25, 1350–1364.
Menary, M. B., W. Park, K. Lohmann, M. Vellinga, M. Palmer, M. Latif, and J. Jungclaus, 2011: A multimodel comparison of centennial Atlantic Meridional Overturning Circulation variability. Climate Dyn., 38, 2377–2388, doi:10.1007/s00382-011-1172-4.
Mignot, J., and C. Frankignoul, 2009: Local and remote impacts of a tropical Atlantic salinity anomaly. Climate Dyn., 35, 1133–1147.
Pardaens, A., M. Vellinga, P. Wu, and B. Ingleby, 2008: Large-scale Atlantic salinity changes over the last-half century: A model–observation comparison. J. Climate, 21, 1698–1720.
Polyakov, I., R. Bekryaev, G. Alekseev, U. Bhatt, R. Colony, M. Johnson, A. Maskshtas, and D. Walsh, 2003: Variability and trends of air temperature and pressure in the maritime Arctic, 1875–2000. J. Climate, 16, 2067–2077.
Polyakov, I., and Coauthors, 2004: Variability of the intermediate Atlantic water of the Arctic Ocean over the last 100 years. J. Climate, 17, 4485–4497.
Polyakov, I., U. Bhatt, H. Simmons, D. Walsh, J. Walsh, and X. Zhang, 2005: Multidecadal variability of North Atlantic temperature and salinity during the twentieth century. J. Climate, 18, 4562–4581.
Polyakov, I., and Coauthors, 2008: Arctic Ocean freshwater changes over the past 100 years and their causes. J. Climate, 21, 364–384.
Pope, V., M. Gallani, P. Rowntree, and R. Stratton, 2000: The impact of new physical parameterizations in the Hadley Centre climate model: HadAM3. Climate Dyn., 16, 123–146.
Proshutinsky, A. Y., and M. Johnson, 1997: Two circulation regimes of the wind-driven Arctic Ocean. J. Geophys. Res., 102, 12 493–12 514.
Proshutinsky, A. Y., R. Bourke, and F. McLauglin, 2002: The role of the Beaufort Gyre in the Arctic climate variability: Seasonal to decadal climate scales. Geophys. Res. Lett., 29, 2100, doi:10.1029/2002GL015847.
Randall, D., and Coauthors, 2007: Climate models and their evaluation. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 608–619.
Redi, M., 1982: Oceanic isopycnal mixing by coordinate rotation. J. Phys. Oceanogr., 12, 1154–1158.
Roberts, C., and M. Palmer, 2012: Detectability of changes to the Atlantic meridional overturning circulation in the Hadley Centre climate models. Climate Dyn., 39, 2533–2546, doi:10.1007/s00382-012-1306-3.
Roberts, M., and R. Wood, 1997: Topographic sensitivity studies with a Bryan–Cox-type ocean model. J. Phys. Oceanogr., 27, 823–836.
Roberts, M., R. Marsh, A. New, and R. Wood, 1996: An intercomparison of a Bryan–Cox-type ocean model and an isopycnic ocean model. Part I: The subpolar gyre and high-latitude processes. J. Phys. Oceanogr., 26, 1495–1527.
Stouffer, R. J., and Coauthors, 2006: Investigating the causes of the response of the thermohaline circulation to past and future climate changes. J. Climate, 19, 1365–1387.
Vellinga, M., and P. Wu, 2004: Low-latitude freshwater influence on centennial variability of the Atlantic thermohaline circulation. J. Climate, 17, 4498–4511.
Vellinga, M., and P. Wu, 2008: Relations between northward ocean and atmosphere energy transports in a coupled climate model. J. Climate, 21, 561–575.
Visbeck, M., J. Marshall, T. Haine, and M. Spall, 1997: Specification of eddy transfer coefficients in coarse-resolution ocean circulation models. J. Phys. Oceanogr., 27, 381–402.
von Storch, H., and F. Zwiers, 1999: Statistical Analysis in Climate Research. Cambridge University Press, 484 pp.
Wilks, D., 1997: Resampling hypothesis tests for autocorrelated fields. J. Climate, 10, 65–82.
Zanchettin, D., A. Rubino, and J. Jungclaus, 2010: Intermittent multidecadal-to-centennial fluctuations dominate global temperature evolution over the last millennium. Geophys. Res. Lett., 37, L14702, doi:10.1029/2010GL043717.
