1. Introduction

To improve our ability to forecast climate system variability, an understanding of the oceans is essential. Sea surface temperature (SST) anomalies are considered to be the main boundary forcing mechanism of the atmosphere; consequently, they may significantly impact the atmospheric circulation. As a result, many studies have focused on the effect of SST patterns on climate (Folland et al. 1986; Moura and Shukla 1981; Lau and Nath 1994; Chang et al. 2000; Saravanan and Chang 2000; Haarsma et al. 2003). Relatively little attention has been given to climate variability in the South Atlantic and much remains to be discovered, particularly with respect to ocean–atmosphere dynamics in the subtropics. Furthermore, observed data are sparse with high levels of uncertainty. Consequently, the use of general circulation models becomes essential to improve our understanding of ocean and atmosphere processes.

Ocean–atmosphere interactions related to SST anomalies in the South Atlantic Ocean and their influence on South American climate have been investigated by previous authors, including Venegas et al. (1997), Diaz et al. (1998), Barros et al. (2000), Doyle and Barros (2002), and Haarsma et al. (2003). Using Singular Value Decomposition analysis, Venegas et al. (1997) reported the existence of a coupled pattern between the South Atlantic subtropical high and SST anomalies on interannual to decadal time scales. The authors suggested that the atmosphere is primarily driving ocean variability. However, a number of studies have shown that the atmosphere also responds to changes in the Atlantic Ocean, especially to tropical SST patterns (Moura and Shukla 1981; Nobre and Shukla 1996; Chang 1996; Chang et al. 2000; Saravanan and Chang 2000). Only a few studies, however, have addressed the atmospheric response to subtropical–extratropical SST anomalies. Haarsma et al. (2003), for example, suggested that subtropical SST anomalies in the South Atlantic by themselves can affect atmospheric tropical circulation. In this study, we examine the atmospheric response to the subtropical South Atlantic SSTs. Special attention is given to precipitation, as it has important implications for the South American economy.

Diaz et al. (1998) showed that southwestern South Atlantic SST can affect precipitation on the adjacent continent. Because of the vigorous mixing of subtropical waters advected southward by the Brazil Current and subpolar waters from the Malvinas Current, this oceanic region is believed to have great influence on the atmosphere. Diaz et al. (1998) found a significant positive correlation between southwest South Atlantic SSTs and precipitation over Uruguay and southern Brazil. They also attributed part of the rainfall variations in southeastern South America to the influence of the El Niño–Southern Oscillation (ENSO).

Precipitation variability in Brazil is large, especially during austral summer when the South Atlantic convergence zone (SACZ) is well established. The SACZ is a convective zone characterized by its low irradiance evident in data of outgoing longwave radiation (OLR). This band of low OLR extends southeastward from the Amazon basin into Atlantic and is associated with high precipitation (Kodama 1992; Satyamurty et al. 1998; Carvalho et al. 2002). Previous studies have shown the importance of some factors on the SACZ onset, such as South America topography (Lenters and Cook 1999), enhanced convection over the Amazon basin (Rocha and Gandu 1996), and the persistence of frontal systems in the subtropics (Carvalho et al. 2002). It is also believed that the South Atlantic SST may affect SACZ variability; however, the mechanisms by which this convective zone is formed and maintained are still unclear.

Barros et al. (2000) studied the relationship between rainfall in southeastern South America, the position of the SACZ and the South Atlantic SSTs. The authors concluded that the interannual rainfall variability is related to both remote and regional forcings. The most well-known remote influence is associated with ENSO, also documented by Ropelewski and Halpert (1987), Grimm et al. (2000), and Robertson and Mechoso (2000). Nogués-Paegle and Mo (1997) and Paegle et al. (2000) reported that the Madden–Julian oscillation also remotely influences SACZ variability. In terms of the local SST influence, some studies have shown that the SACZ responds to the southwestern South Atlantic SST anomalies, but there is no consensus about the mechanisms involved in such a processes. Robertson and Mechoso (2000) and Barreiro et al. (2002) found a relationship between the SACZ position and SSTs associated with a cyclonic circulation. Doyle and Barros (2002) suggested that a positive feedback between the positive (negative) SST anomalies in western South Atlantic and weak (intense) SACZ activity might enhance the low-level circulation pattern associated with the SACZ position. Indeed, Robertson et al. (2003) found that the southwest Atlantic SST anomalies can significantly influence interannual variations of the SACZ, as well as the associated changes in circulation and precipitation. A comprehensive study of the SACZ behavior as well as its influence on extreme precipitation events in São Paulo, Brazil, can be found in Carvalho et al. (2002).

Most of the above studies used AGCMs in order to obtain the climate response resulting from certain SST anomalies. Generally, the atmospheric response to these SST anomalies is clearly detectable in the tropics (Shukla 1984, 1993; Stern and Miyakoda 1995; Yang et al. 1998, hereafter YAS98). Tropical variability on interannual time scales is dominated by SST forcing, such as ENSO (Trenberth 1985; Lau and Nath 1994; Saravanan and Chang 2000). Unlike in the tropics, however, subtropical–extratropical variations are dominated by internal atmospheric variability, which is difficult to predict on seasonal time scales (Shukla 1985; Saravanan 1998).

A common method used to increase reliability in forecast simulations is to generate several integrations and to assume the ensemble mean as a substitute for a single forecast. Theoretically speaking, if the number of integrations from a perfect model is infinite the ensemble mean would give the climate response to the forcing signal. However, computational resources are limited and it is impossible to produce an infinite ensemble. For a small set of integrations, the ensemble mean will be composed of both signal and noise (internal variability). If the noise is large compared to the signal, the atmospheric variations will be difficult to predict because they are dominated by internal chaotic variability. Thus it is essential to verify the predictability of the ensemble mean independently if its members present a large or small spread. Stern and Miyakoda (1995) defined the reproducibility, an evaluation of the spread among the ensemble members based on signal-to-noise ratio, as a measure of potential predictability. According to YAS98, a large spread, or low reproducibility, does not imply lack of predictability. The authors proposed a technique to isolate the externally forced response even in regions where it seems negligible compared to internally generated atmospheric noise.

In this work we use the YAS98 technique to examine some aspects related to reproducibility and potential predictability of precipitation simulated by the National Center for Atmospheric Research (NCAR) Community Climate Model, version 3 (CCM3). This evaluation is very important to verify the accuracy of the model performance, before speculating on the mechanisms of how SSTs may impact the climate.

The atmospheric model, the ensemble integrations and the dataset used as boundary condition are described in section 2. The results of reproducibility as well as a brief description of its definition are presented in section 3. Section 4 explains the reconstruction of the ensemble and its consequence with regard to the internal variability. Section 5 discusses the truncation level used for the ensemble reconstruction by including the potential predictability index. Finally, the last section summarizes the major findings of this work.

2. Model and experiment

The atmospheric model used in this study is the CCM3 from NCAR. The NCAR CCM3 is a stable, efficient, and well-documented atmospheric general circulation model designed for climate research on high-speed supercomputers and select upper-end workstations. The standard configuration of the CCM3 (Acker et al. 1996) has T42 spectral truncation in the horizontal (approximately 2.8° latitude × 2.8° longitude) and 18 vertical levels. Land surface processes in CCM3 are represented by a fully interactive land model (Bonan 1998).

Over the last decades, numerous scientific institutions around the world have used GCM simulations for basic research, including climate prediction. Modifications to physical parameterizations for the latest version to the CCM3.6.6 are described in Kiehl et al. (1998a).

The Hadley Centre Global Sea Ice and SST analyses (HadISST1) was used to force the atmospheric model. This dataset contains monthly mean fields of SST and sea ice concentration from 1870 through the present. Its global coverage is a 1° latitude × 1° longitude grid, which was interpolated to the CCM3 grid. The global SST fields were created using a variety of techniques including reduced space optimal interpolation (RSOI). Sea ice concentration fields from different sources were analyzed to make them as consistent as possible through time. The HadISST1 dataset has been shown to be an improvement upon the Global Sea Ice Coverage and Sea Surface Temperature (GISST) dataset. A detailed description can be found in Rayner et al. (2003).

A five-member ensemble was generated with small changes to the initial atmospheric state in order to represent the internal variability of the atmosphere. The initial atmospheric conditions differ randomly across the globe, with a maximum perturbation of 10% from the initial temperature value. This ensemble size is adequate to reproduce the internal atmospheric dynamical processes when a large-scale seasonal simulation is run for a relatively long period (Shukla 1993). Each integration was forced with monthly varying SSTs from September 1949 to October 2001. The SST forcing was specified only in the South Atlantic Ocean between 20° and 60°S, as shown in Fig. 1. Climatological SST values were assigned elsewhere in the global oceans.

As described in the introduction, some authors have previously studied the rainfall response to the tropical Atlantic SSTs using AGCMs (Moura and Shukla 1981; Nobre and Shukla 1996; Chang et al. 2000; Robertson and Mechoso 2000; Barreiro et al. 2002). However, most of these numerical experiments were forced with SSTs from the equator to 60°S, and thus these previous results would have been overwhelmed by the tropical boundary forcing, as its signal is stronger compared to that from the subtropics. It is already known that South American precipitation is influenced by both Pacific and Atlantic Oceans, and also that the larger influence of SSTs on climate is from those at tropical latitudes. However, the impact of the subtropical SSTs still remains unclear. For these reasons, we forced the model with SST anomalies only at subtropical latitudes, excluding the tropics.

3. Reproducibility in ensemble

The methodology used herein includes a technique described in YAS98. These authors isolate the forced modes by applying empirical orthogonal function (EOF) decomposition onto the ensemble and reconstructing it using only the most reproducible modes. The term reproducibility means the level of agreement among the measurements under modified conditions. In this study, reproducibility is related to the level of agreement between the ensemble members produced using different initial conditions. It provides an estimate based on the signal-to-noise ratio. Consider that X_ij is a variable of the ith integration from an ensemble with size M in the jth month among N of a particular season. Also, assume that the ensemble monthly mean is represented by 〈X〉_j, and the ensemble seasonal mean is X. Time series were split into seasons, namely December–February (DJF) as austral summer, March–May (MAM) as autumn, June–August (JJA) as winter, and September–November (SON) as spring. In this case, the time series starts in December 1949, and ends in November 2000, which totals N = 153 months (51 yr) for each season in an ensemble composed by M = 5 members. Hence reproducibility is defined by

View Expanded

View Expanded

is the internal variance or a measure of the ensemble spread and

View Expanded

is the total variance, which satisfies the relationship σ²_S = σ²_I + σ²_E, with

View Expanded

as the external variance. From Eq. (1), the higher the reproducibility, the less dominant is the internal variance (or noise) and better is the agreement among the ensemble integrations.

Figure 3 depicts the spatial distribution of the reproducibility for precipitation, as calculated from Eq. (1), taking into account the original ensemble. Reproducibility is large over tropical latitudes and very low at subtropical–extratropical regions. Broadly speaking, the tropical atmosphere is primarily thermally driven (Gill 1980) and is dominated by external variance (YAS98; Stern and Miyakoda 1995). On the other hand, extratropical atmospheric variations are dominated by internal variance. Consequently, the response of the ITCZ to SST forcing is very strong when compared to the SACZ variations in South America, which in turn, seems to be dominated by internal variance. This result is consistent with Barreiro et al. (2002, 2005).

Another interesting feature in Fig. 3 is the seasonal dependence of reproducibility. It is larger during austral autumn (MAM) and spring (SON) than in summer (DJF) and winter (JJA), especially over South America and Africa. This result may indicate that the summer and winter have relatively larger scatter than the transient seasons. The DJF precipitation shows small reproducibility over most of South America. It is worthwhile noting that, even though precipitation over the SACZ region has a large external variance (not shown), the noise dominates the total variance in Eq. (1), resulting in a low reproducibility. Indeed, in summer convective activity and systems are not predictable because their spatial and temporal scales are not well resolved by the model. In the austral winter, cold frontal passages are the most common transient weather events over South America. Frontal penetrations are more frequent in southern South America, being responsible for a large part of the rainfall in northern Argentina, Uruguay, Paraguay, Bolivia, southern Peru, and southern, southwestern Brazil. Around 20°S, cold air masses gradually lose their identity and merge with the subtropical high (Satyamurty et al. 1998). The high rainfall variability brought from the frontal systems might be one of the causes of the low reproducibility during JJA months.

It is worth noting that the spatial and seasonal distributions of reproducibility vary according to the model and variable studied. For instance, using an ensemble generated with a different AGCM, YAS98 show that the reproducibility is larger in JJA and lower in DJF for 850-hPa temperature anomaly and 300-hPa geopotential height anomaly. Also, the authors show some preferred geographic areas across the globe. For the 850-hPa temperature, for instance, the region including Alaska, northwest Canada, and the southeast United States, as well as the traditional Pacific–North America (PNA) region are potentially predictable during boreal spring and winter (MAM and DJF, respectively). In addition to the geographic regions, differences in reproducibility may arise for several reasons, for example the model resolution, the physical parameterizations, the initial conditions imposed on the simulations and the SST forcing prescribed from the observations.

Figure 4 summarizes both the seasonal and latitudinal dependence of reproducibility. The number of longitude points with reproducibility larger than 0.4 rises from southern high latitudes to the equator. This behavior is more evident during austral winter (line with crosses), with a steep rise in the reproducibility toward low latitudes. Thus, in order to investigate an extratropical phenomenon, it is necessary to apply a technique that increases the reproducibility in those regions.

4. Reconstruction of ensemble

The fact that precipitation in tropical latitudes is more reproducible than in subtropical–extratropical regions does not mean that the subtropical SSTs have no significant influence on the atmosphere. To highlight the forced signal in those areas, we decompose the ensemble using EOF analysis and reconstruct it back using only the first reproducible forced modes. The main difficulty in this method is to determine the number K of EOF modes to be considered in the reconstruction. This problem occurs because, if a large number of modes are included, the reconstructed ensemble will retain the noise. On the contrary, if a small number of EOF modes are taken into account in the truncated reconstruction, the new ensemble mean would eliminate potentially reproducible modes despite the noise reduction. Therefore the truncation level should be high enough to provide a good representation of the forced signal but low enough to eliminate the noise.

A first approach was done by computing the EOF-based Monte Carlo (Press et al. 1986) and North et al. (1982) rule of thumb tests on the explained variance of each decomposed mode from the ensemble mean, as shown in Fig. 5. The Monte Carlo test allows us to distinguish the explained variance obtained from the EOF analysis by testing it against a white noise null hypothesis. Significances of the EOF modes estimated by North et al. (1982) criteria are based on comparing the separation among the neighboring eigenvalues with its sampling error.

According to the Monte Carlo test in Fig. 5, there is a different threshold value for each season. By considering variances in which the North et al. (1982) error bars are above the Monte Carlo threshold, we fixed the truncation levels to K = 16 for DJF, K = 18 for MAM, K = 19 for JJA, and K = 22 for SON. After the ensemble reconstruction, reproducibility increased significantly, as shown in Fig. 6.

A general growth of the reproducible areas can be seen from Fig. 6, indicating an enhanced atmospheric response due to the contribution of the SST forcing. Although the distribution of reproducibility has risen in the truncated ensemble, the seasonal dependence remains the same, with higher values in the transient seasons and lower values in summer and winter. During MAM and SON, the increase in reproducibility is noticeable over both South America and Africa as well as in extratropical latitudes of the Southern Ocean.

The fact that the reproducibility increases with the reconstructed ensemble occurs because of the reduction in the spread among the ensemble members. To understand this decrease, Figs. 7 and 8 are examined. They exhibit the spatial distribution of the internal variance from the original and the reconstructed ensemble, respectively. The internal variance is very pronounced over South America and Africa (Fig. 7), particularly during austral summer. This internal variance is significantly reduced in the reconstructed ensemble (Fig. 8). In SON (Fig. 8d) the internal variance is almost completely eliminated over Africa. A very impressive decrease in the noise can also be seen in MAM and JJA seasons. The reduction of the internal variance is statistically significant at the 95% of confidence level according to a t test.

Although the spread among integrations is reduced after the reconstruction of the ensemble, austral summer still keeps a high internal variance over South America, which explains the low reproducibility, particularly along the SACZ region. Even though the internal variance is reduced in austral winter (Fig. 8c), the spatial distribution of reproducibility (Fig. 6c) is only slightly improved because the external variance is also very low. After the reconstruction of the ensemble, the internal variance in the ITCZ region was significantly decreased, resulting in a marked rise of reproducibility as seen in Fig. 6.

5. Potential predictability

A good approach to determine the truncation level is the examination of the reproducibility of the reconstructed ensemble in terms of the first K EOF forced modes. Figure 9 reveals the percentage of grid points with reproducibility larger than 0.4 versus the number of EOF modes used in the reconstructed ensemble. It is important to note the decrease of reproducible area with the increase of the truncation level. All seasons show a steep reduction in the reproducibility after a first few EOFs, particularly in austral summer and winter.

According to YAS98, however, a low reproducibility may not imply bad predictability. If the ensemble mean is driven by externally forced modes, potential predictability may exist even if the noise is large. Two points are necessary to determine the existence of potential predictability: the first one is that the local variance of the ensemble mean must be dominated by forced modes and the other is that these modes must be highly reproducible. For this reason, YAS98 defined an index for potential predictability, which includes the reproducibility with local variance contribution:

View Expanded

where R^K is the reproducibility of the reconstructed ensemble in terms of the first K EOF modes, and F^K = (σ^K_E/σ_E) is the ratio between the local variance explained by the truncated reconstruction and the original ensemble. By using the PPI, the number K of truncations that yields an acceptable level of potential predictability can be determined.

The spatial distribution of potential predictability (not shown) taking into account the same truncation levels as in Fig. 3 is very similar to that of the reproducibility. Figure 10 shows the latitudinal curve of the percentage of longitude points with potential predictability no lower than 0.4, plotted by season and number of EOF modes used in the reconstruction. The latitudinal distribution of potential predictability shows an increase toward the equator for all seasons. This behavior is more pronounced in DJF and JJA because reproducibility is very low in subtropical–extratropical regions during these months (see Fig. 3). It is worth noting the increase in potential predictability with lower truncation levels (Fig. 10). This can be seen for MAM and SON months (Figs. 10b,d), while summer and winter (Figs. 10a,c) do not show great differences in potential predictability with the truncation levels.

Figure 11 presents the relationship between the percentage of grid points with potential predictability higher than 0.4 versus the number of EOF modes considered for the reconstruction. The most remarkable feature in Fig. 11 is the difference between the potential predictable area in the transition seasons when compared to DJF and JJA. One possible reason for this difference is the effect of the initial conditions in simulating seasons with high oscillations of precipitation (i.e., the intense variations in the SACZ region in summer or the frontal systems in winter, when chaotic processes in the atmosphere dominate the forced signal, as discussed earlier).

According to Fig. 11, the potential predictability reduces with the increase of EOF modes used for the reconstruction. After a certain number of EOFs, predictability seems to be insensitive to the level of truncation. This is particularly clear in the summer and winter curves as reproducibility is lower than during fall and spring and thus they rapidly reach this level of insensitivity.

The choice of truncation level for the reconstruction can be based on Fig. 11. The number of EOF modes that maximize the geographical area of potential predictability is 2 for DJF, 3 for MAM, and 2 for the subsequent months. Considering another high number of truncation levels, it can be fixed, for instance, at 80% of the maximum potential predictability; thus retaining 8 EOFs for DJF, 10 for MAM, 7 for JJA, and 9 for SON. Table 1 presents the reproducibility averaged over the studied region for the original and the reconstructed ensemble with 80% and 100% of the maximum potential predictability. The increase in the reproducibility is evident. The ensemble reconstructed with 80% of the maximum potential predictability presents approximately double the averaged reproducibility than the original ensemble. When considering the maximum potential predictability to reconstruct the ensemble, the region becomes highly reproducible, revealing about 3–4 times more reproducible area than the original ensemble.

Figure 12 shows the increase in geographical area with reproducibility higher than 0.4 after reconstructing the ensemble with 80% of maximum potential predictability index. When considering this level of truncation, it is still possible to note some small reproducible areas in the extratropical latitudes during summer (Fig. 12). During MAM and SON (Figs. 12b,d), precipitation in most of the South Atlantic, South America, and Africa is reproducible after the ensemble reconstruction. Note that the summer reproducibility increased considerably over the SACZ region, especially in its oceanic branch.

To examine the decrease in internal variance as a function of time, we have selected four arbitrary grid points over South America: 20°S, 50°W for DJF, 10°S, 37°W for MAM, 50°S, 75°W for JJA, and 10°S, 75°W for SON. Figure 13 shows the rainfall time series of the ensemble mean with the original (dark gray shade) and the reconstructed (light gray shade) internal variance, using the 80% level of the maximum potential predictability. As expected, the internal variance decreases after the ensemble reconstruction for all seasons because of the reduction in the spread among the integrations. This is in agreement with the reproducibility shown in Figs. 3 and 12 for the original and the reconstructed ensemble, respectively.

The amplitude of the variations in precipitation time series is larger during austral summer (Fig. 13a) because of the presence of the SACZ at 20°S, 50°W. The MAM time series (Fig. 13b) also shows high variations at 10°S, 37°W, as this is the rainy season in northeast Brazil. The JJA months (Fig. 13c) exhibit large variations at 50°–75°S because of the rainy season around the Patagonia region. Extratropical weather systems are responsible for the enhanced rainfall in the southern tip of South America during wintertime. After the ensemble reconstruction, the internal variance is markedly reduced in Fig. 13c. The SON time series (Fig. 13d) reveals a very impressive reduction in the spread in ensemble at 10°S, 75°W. Table 2 quantifies the difference between the spread for the original and reconstructed ensemble with respect to the time series presented in Fig. 13.

The fact that the summer and autumn time series contain larger internal variance than the other two time series suggests that precipitation at those locations is more sensitive to initial conditions in the model. This can also be interpreted as a consequence of the predominance of chaotic components during those seasons. Indeed, the different behavior among the seasons is also sensitive to the grid point selected.

6. Discussion and conclusions

Reproducibility and potential predictability of precipitation over the South America and South Atlantic regions are examined in the NCAR CCM3 model. A five-member ensemble is generated by forcing the atmosphere with prescribed SST from late-1949 to 2001 in the South Atlantic region and a seasonal climatology everywhere else. An important issue could be raised about the size of the ensemble. The choice of ensemble size is dependent on the variable studied, the model used to run it, and, most importantly, the time scale of the simulations (Branković and Palmer 1997; Rowell 1998; Wehner 1998). Chang et al. (2000), Saravanan and Chang (2000), and Barreiro et al. (2002, 2005) have had good results for precipitation and atmospheric circulation with a relatively small ensemble size produced with the NCAR CCM3 model forced with SST in the tropical Atlantic and South Atlantic Ocean, covering similar time scales to those used in this work. Shukla (1984) states that, in large-scale seasonal simulations, the tropical atmospheric patterns are strongly determined by the underlying SST. It is found that, even after very large changes in the initial conditions, the atmospheric rainfall patterns in the tropics converge to values determined by the ocean temperature (Shukla 1998). According to the author, this provides a scientific explanation for why, although it is not possible to make accurate forecasts of the day-to-day weather events beyond 1 or 2 weeks, it should be possible to predict the large-scale seasonal tropical circulation and rainfall for as long as the ocean temperature can be predicted. Here we show that, outside of the tropics, a different approach might be necessary to study the extratropical processes.

Spatial distributions of reproducibility showed higher values in the tropics than in the subtropical–extratropical latitudes. It also showed a seasonal dependence, with greater reproducibility during MAM and SON. This indicates that the forced response of the atmosphere at high and midlatitudes is dominated by internal atmospheric variations. The lack of reproducibility over South America in austral summer is probably due to the chaotic variations originating from the SACZ region. The SACZ is generally associated with high precipitation during DJF season.

With respect to the low reproducibility in winter, it may be argued that it is caused by disorganized precipitation carried via frontal systems, as they are the most common events during this time of year. On the other hand, transient seasons seem to present relatively well-behaved variations, which resulted in a smaller spread among the ensemble members. As a consequence, MAM and SON revealed larger geographical reproducible areas compared to DJF and JJA months.

Nevertheless, this does not mean that external forcing has no effect on atmospheric variations, but that its signal is overwhelmed by atmospheric noise. To highlight this signal in the subtropics, we used the technique of YAS98 to reconstruct the ensemble with the most reproducible modes, thus reducing the internal variance. Special attention was given to the decision of the truncation level for the reconstruction. In this study, we used the EOF-based Monte Carlo test, the North et al. (1982) rule of thumb method, and the YAS98 predictability index to estimate the threshold. We strongly believe that a better evaluation of the truncation level must be done not only with the YAS98 criteria but also with other statistical methods. Although the potential predictability index is a powerful tool for selecting the number of EOF modes for reconstruction, the decision is still arbitrary. Therefore, apart from the approach of YAS98, the Monte Carlo and North et al. (1982) tests were used to reinforce a limit for the truncation level. In this case, the thresholds were fixed at 16 for DJF, 18 for MAM, 19 for JJA, and 22 for SON. Thus, it is possible to choose any number of EOFs between the maximum predictability index given by the YAS98 criteria and the threshold computed by the Monte Carlo test. For instance, we have chosen 80% of maximum potential predictability for the ensemble reconstruction, where the truncation level was fixed at 8 for DJF, 10 for MAM, 7 for JJA, and 9 for SON. Note that the threshold of 0.4 in PPI is empirically selected in the same manner as in YAS98. This parameter can be changed in future studies when necessary.

The spatial distribution of reproducibility for the original ensemble revealed that the SACZ region is dominated by internal variance, while the atmospheric response in the ITCZ region due to SST anomalies is driven by external variance. Since the SACZ is more sensitive to changes in initial conditions of the model, it becomes less predictable than the ITCZ. However, the ensemble reconstruction yields a significant increase in the reproducibility distribution for all seasons, which makes possible the study of regions that were not predictable before.

As the SACZ is one of the most important convective regions in South America, special attention should be given to it. It is worthwhile to mention that the SACZ is a complex phenomenon influenced by several factors, such as the topography (Lenters and Cook 1999) and continentality of South America, and the consequent generation of the thermal low over the Chaco region (Lenters and Cook 1997); the convection over the Amazon basin (Rocha and Gandu 1996); and the persistence of frontal systems in the subtropics (Carvalho et al. 2002); among others. In this study, we examined the reproducibility of the precipitation due to the South Atlantic SST variations, and thus it represents only a component of the many processes that might influence the SACZ. Predictability may be enhanced if other factors responsible for the SACZ variability are included in the model. Another point that should be noted is that the South Atlantic SSTs are also affected by remote forcing, particularly from the tropical Pacific. Whether the results obtained here are partly affected by remote factors is not considered in present study. Barreiro et al. (2002) concluded that the dominant SST-forced precipitation signal in the SACZ comes from the SST anomalies locally generated in the South Atlantic. We showed that the reproducibility and the predictability of precipitation over the SACZ region are larger in its oceanic branch than in its continental part, which in turn is predominantly driven by the internal atmospheric variations.

It is worth noting that rainfall predictability over the equator is quite large, although the SST anomalies were defined south of this region. Precipitation processes depend not only on the underlying SST variations but also on the atmospheric dynamics. It is commonly accepted that a tropical heat source affects remote regions in the globe, for example, warm Pacific SST anomalies during El Niño events. A few studies showed that a subtropical heat source in the Atlantic could impact both tropical and extratropical atmospheric patterns (Huang and Shukla 2005). Robertson et al. (2003) and Haarsma et al. (2003) found that the South Atlantic subtropical dipole, the dominant mode of variability in SST anomalies, does have a significant effect on the equatorial dynamics. They reported a deep baroclinic response over the equatorward pole and a shallow equivalent barotropic response over the poleward pole. According to Haarsma et al. (2003), the anomalous vertical circulation, initiated by diabatic heating from the SST dipole, generates downward motion north of the ITCZ and upward motion southward of it, thus shifting the convergence zone to south. Therefore, enhanced precipitation occurs over the ascend region, southward of the ITCZ mean position, and is maintained by anomalous convergence of low-level moisture. Examining the mechanisms by which the tropical rainfall responds to the subtropical South Atlantic SST is beyond the scope of this study.

Finally, it is important to point out that the results obtained here were based on numerical integrations from the CCM3 model and so depends on the model’s physical parameterizations and its capability to simulate the climate. Therefore, an intercomparison with different AGCMs may be desirable in the near future in order to improve our results.

The ensemble treatment used here is the first step in a detailed investigation into the atmospheric forced response to SST anomalies from the subtropical South Atlantic Ocean on South American precipitation. Moreover, this work also provides additional information for future studies involving AGCM simulations forced by any kind of boundary conditions, such as SST anomalies or soil moisture, for example. Further work is underway to investigate the mechanisms behind the atmospheric response to the SST forcing imposed in this study.