2904 JOURNAL OF CLIMATE VOLUME 10

Atmosphere–Ocean Coupled Variability in the South Atlantic

S. A. VENEGAS, L. A. MYSAK, AND D. N. STRAUB
Centre for Climate and Global Change Research and Department of Atmospheric and Oceanic Sciences,
McGill University, Montreal, Quebec, Canada
(Manuscript received 9 July 1996, in final form 9 April 1997)

ABSTRACT
The climate variability of the South Atlantic region is determined from 40 yr (1953–92) of Comprehensive
Ocean–Atmosphere Data Set monthly sea surface temperature (SST) and sea level pressure (SLP) data using
the empirical orthogonal function (EOF) and the singular value decomposition (SVD) analysis methods. The
EOF method is applied to each field separately, whereas the SVD method is applied to both fields simultaneously.
The significance of the atmosphere–ocean interaction is revealed by a strong resemblance between individual
(EOF) and coupled (SVD) modes of SST and SLP. The three leading modes of coupled variability on interannual
and interdecadal timescales are discussed in some detail.
The first coupled mode, which accounts for 63% of the total square covariance, represents a 14–16-yr period
oscillation in the strength of the subtropical anticyclone, accompanied by fluctuations of a north–south dipole
structure in the SST. The atmosphere–ocean coupling is strongest during the southern summer. The second
coupled mode (20% of the total square covariance) is characterized by east–west shifts of the anticyclone center,
in association with 6–7-yr period fluctuations of SST off the coast of Africa. The coupling depicted by this
mode is weaker than that found in the first and third modes. The third coupled mode (6% of the total square
covariance) is characterized by north–south displacements of the anticyclone, accompanied by SST fluctuations
over a latitudinal band in the central South Atlantic. These oscillations occur on a relatively short interannual
timescale (;4 yr). As with the first mode, the atmosphere–ocean coupling is strongest during the southern
summer. This mode is found to be temporally and spatially correlated with the El Niño–Southern Oscillation
phenomenon. The statistical robustness of the results is tested by using a Monte Carlo approach, which indicates
that the presented results are highly significant.

1. Introduction sure (SLP) fluctuations and their coupled variability

During the past few decades there has been consid-
erable effort devoted to obtaining a better understanding
of natural climate variability on interannual to inter-
decadal timescales. To determine the mechanisms gov-
erning these climatic variations, it is essential to char-
acterize the large-scale interactions between the ocean
and the overlying atmosphere. A number of recent stud-
ies have sought to identify such interactions and their
associated timescales, by analyzing both observational
records and the results of climate model simulations.
The nature of the global air and sea surface temper-
ature variability on interannual to interdecadal time-
scales has been documented in both data studies (Fol-
land et al. 1984; Ghil and Vautard 1991; Houghton and
Tourre 1992; Mann and Park 1993; Mann and Park
1994) and model simulations (Manabe and Stouffer
1994; Mehta and Delworth 1995). The relationships be-
tween sea surface temperature (SST) and sea level pres-
Corresponding author address: Dr. Lawrence A. Mysak, Centre
for Climate and Global Change Research, McGill University, 805
Sherbrooke Street West, Montreal, PQ H3A 2K6, Canada.
E-mail: mysak@zephyr.meteo.mcgill.ca
have also been widely investigated by analyzing a va-
riety of datasets particularly over the North Atlantic and
the North Pacific Oceans (Deser and Blackmon 1993;
Kushnir 1994; Latif and Barnett 1996; Mann and Park
1996) and by performing experiments with atmospheric,
oceanic, or coupled atmosphere–ocean models (Palmer
and Sun 1985; Kushnir and Lau 1992; Delworth et al.
1993; Zebiak 1993; Peng et al. 1995; Halliwell 1996).
Thermodynamic and wind-induced mechanisms such
as air–sea heat exchange, entrainment, vertical mixing,
and wind stress forcing, have been frequently proposed
as ways in which the atmosphere forces upper-ocean
temperature anomalies (Frankignoul 1985; Wallace et
al. 1990; Cayan 1992a; Cayan 1992b; Miller et al. 1994;
Battisti et al. 1995). Other studies, however, have sug-
gested that the ocean may drive atmospheric circulation
anomalies through large-scale changes in the thermo-
haline circulation on interdecadal or longer timescales
timescales (Levitus 1989; Ghil and Vautard 1991;
Stocker and Mysak 1992). A few investigators have also
proposed positive feedbacks between SST and the at-
mospheric circulation on interdecadal timescales (Deser
and Blackmon 1993; Latif and Barnett 1994).
A common feature of these studies is that the major

q 1997 American Meteorological Society

NOVEMBER 1997 VENEGAS ET AL.
modes of large-scale variability in the Northern Hemi-
sphere are characterized by interannual to interdecadal
timescales. It has been generally accepted that inter-
annual variability is primarily driven by the atmosphere,
while interdecadal or longer-scale variability is mainly
associated with changes in the ocean (Deser and Black-
mon 1993; Kushnir 1994). However, recent work by
Halliwell (1996) and results from this paper suggest that
atmospheric forcing of the ocean on interdecadal times-
cales may be more important than previously realized.
As a complement to the Northern Hemisphere focus
of the above studies, this paper (which is partly sum-
marized in Venegas et al. 1996) investigates the natural
climate variability in one region of the Southern Hemi-
sphere, namely, the South Atlantic. The first goal is to
identify the principal modes of behavior of the SST and
the overlying atmospheric circulation, in order to pro-
vide insight into the variability of the South Atlantic
coupled atmosphere–ocean system on interannual to in-
terdecadal timescales. A secondary goal is to determine
whether the South Atlantic modes of variability are con-
nected with any of the well-known tropical and/or
Northern Hemisphere climate oscillations (e.g., El
Niño–Southern Oscillation, North Atlantic Oscillation,
Pacific–North American pattern). The suggestion of
such a connection between hemispheres on interdecadal
timescales was proposed in Mysak and Power (1992).
These connections were further explored by Mann and
Park (1993, 1994).
Empirical orthogonal function (EOF) and singular
value decomposition (SVD) analyses are used to de-
scribe both the independent and coupled variability of
SST and SLP in the South Atlantic. While the EOF
analysis method has been widely applied in geophysics
research (Davis 1976; Wallace et al. 1990; Deser and
Blackmon 1993; Enfield and Mayer 1997), the SVD
analysis method has become more commonly used only
in the last decade (Lanzante 1984; Fang and Wallace
1994; Peng and Fyfe 1996), although it was first pro-
posed in a meteorological context by Prohaska (1976).
The interest in trying to model the type of South
Atlantic variability described in this paper has substan-
tially increased since the creation of two recent large-
scale climate programs: CLIVAR (Climate Variability
and Predictability Program) and ACCP [The Atlantic
Climate Change Program, described in Molinari et al.
(1994)]. It is hoped that this observational study will
offer new insights into the mechanisms and long-range
linkages of interannual to interdecadal climate vari-
ability. These insights will also help the modeling stud-
ies associated with these programs.
The paper is organized as follows. A brief description
of the datasets and methods, together with an introduc-
tion to the main South Atlantic oceanic and atmospheric
features are given in section 2. Results from the inde-
pendent SST and SLP EOF analyses follow in section
3. The results pertaining to the atmosphere–ocean cou-
2905
FIG. 1. Percentage of months in which data was available at each
grid point over the 480-month period analyzed.
pling based on the SVD analysis are discussed in section
4. A summary and the conclusions are given in section 5.
2. Data and methods
a. Data
The data employed in this study include South At-
lantic sea SST, sea level pressure (SLP), and vector wind
extracted from the Comprehensive Ocean–Atmosphere
Data Set (COADS) (Woodruff et al. 1987). These data
are available as monthly means over a grid of 28 lat 3
28 long, and span the period 1854–1992. The 40-yr
period analyzed in this paper extends from 1953–92,
during which time the data coverage over the South
Atlantic basin is fairly good. The area of interest of our
study is the South Atlantic Ocean, from the equator to
508S, and from 708W (or South American coast) to 208E
(or South African coast). The region south of 408S be-
tween 408W and 208E is excluded from the analysis,
however, due to poor data coverage (see Fig. 1).
The SST, SLP, and wind climatologies are calculated
for each calendar month at each grid point by averaging
the data over the 40-yr period. Monthly anomalies are
then defined as deviations from this mean annual cycle.
To reduce the noise inherent in the monthly observa-
tions, the anomalies are temporally smoothed by per-
forming three-month running means (3MRM). At a giv-
en grid point, 3MRMs are computed by averaging
anomalies of three consecutive months and by assigning
that mean value to the central month. If two or more
neighboring monthly anomalies are missing, the cor-
responding 3MRM is not computed and the month is
given a missing value flag.
For a successful application of the EOF and SVD
analysis methods described in the next subsection, the
spatial gaps in the SST and SLP data are filled using
an inverse distance interpolation method (Thiebaux and
Pedder 1987). As a final step before applying the sta-
2906 JOURNAL OF CLIMATE
tistical methods, the data from the original 28 by 28 grid
are averaged onto a coarser 48 lat 3 108 long grid to
reduce the size of the datasets used in our analysis.
b. Methods
Detailed discussions on EOF and SVD analyses can
be found in Bretherton et al. (1992), Wallace et al.
(1992), von Storch and Navarra (1995), and Newman
and Sardeshmukh (1995). A brief description of both
methods follows here.
EOF analysis is a statistical technique based on fun-
damental matrix operations. It has proven to be helpful
in objectively identifying the principal (spatially un-
correlated) modes of variability of a given field. The
analysis applies to a space- and time-dependent field
with zero temporal mean. The covariance matrix of the
field is constructed and diagonalized, resulting in a set
of eigenvalues and corresponding eigenvectors. Each
eigenvector (EOF) can be regarded as a spatial pattern
(a map). To see how a given spatial pattern ‘‘evolves’’
in time, the eigenvector is projected onto the original
field to obtain a time series (the expansion coefficient).
Just as the EOFs are orthogonal in space, the associated
time series are orthogonal in time. The fraction of the
total field variance explained by a given EOF is pro-
portional to its associated eigenvalue. Together, an ei-
genvalue with its corresponding EOF and expansion co-
efficient define a mode of variability. The leading mode
(related to the largest eigenvalue) explains the largest
fraction of the total variance, the second mode explains
the largest fraction of the remaining variance, and so
on.
SVD analysis can be thought of as a generalization
to rectangular matrices of the diagonalization of a square
symmetric matrix (i.e., EOF analysis). It is usually ap-
plied to two data fields together in order to identify pairs
of coupled spatial patterns, which explain as much as
possible the covariance between the two variables. The
SVD of the cross-covariance matrix yields two spatially
uncorrelated sets of singular vectors (analogous to the
eigenvectors, but one for each variable) and a set of
singular values associated with each pair of vectors
(analogous to the eigenvalues). Each pair of singular
vectors describe a fraction of the square covariance (SC)
between the two variables. The first pair describes the
largest fraction of the SC and each succeeding pair de-
scribes a maximum fraction of the SC that is unex-
plained by the previous pairs. The square covariance
fraction (SCF) accounted for by the kth pair of singular
vectors is proportional to the square of the kth singular
value. The kth expansion coefficient for each variable
is computed by projecting the respective kth singular
vector onto each original data field. The correlation val-
ue r between the kth expansion coefficients of the two
variables indicates how strongly related the coupled pat-
terns are.
Both EOF and SVD analysis methods assume the data
VOLUME 10
FIG. 2. Schematic representation of the South Atlantic upper-level
ocean circulation (from Peterson and Stamma 1991).
to be complete. When there are gaps in the data, and
these are taken into account in the construction of the
covariance matrix, the resulting modes are no longer
strictly orthogonal. This problem is overcome here by
filling the gaps as described previously. The differences
in the results using the original and the interpolated
datasets are, however, minimal. The analyses presented
here are based on the interpolated data.
Correlation maps are used extensively in this study.
These consist of gridpoint correlations between a time
series (e.g., the expansion coefficient of an EOF mode,
an index such as SOI, etc.) and gridpoint anomalies of
a given field. Thus, correlation maps indicate how well
the gridpoint anomalies of the field can be predicted
from the knowledge of the time series. In addition, they
provide a measure of the percentage of the variance
explained locally by each mode, represented by the
square of the correlations (Houghton and Tourre 1992).
The spectral analyses and coherences performed in
this study are based on the multitaper method (Thomson
1982), and the confidence levels are obtained relative
to an estimated red noise background (Mann and Lees
1996). Significance levels for all the correlations are
estimated using the method described in Sciremammano
(1979). This method accounts for the month-to-month
autocorrelation inherent in the time series to compute
the actual number of degrees of freedom before the
estimation of significance levels. The 95% significance
level is given in square brackets after each correlation
coefficient. A Monte Carlo approach is used to deter-
mine the significance of the SST and SLP decomposi-
tions in modes of variability. A detailed description of
the method is given in section 4d.
c. Mean South Atlantic conditions
Figure 2 depicts the ocean circulation features of the
South Atlantic. The upper-level circulation is dominated
NOVEMBER 1997 VENEGAS ET AL.
FIG. 3. (a) SST (in 8C) climatology for 1953–92. Isotherm interval
is 28C. (b) SLP (in hPa minus 1000 hPa) and vector wind climatol-
ogies for 1953–92. Isobar interval is 3 hPa.
by the South Equatorial Current, the Subtropical Gyre,
and the Antarctic Circumpolar Current (ACC). The main
components of the Subtropical Gyre are the warm and
saline Brazil Current in the west and the cool and fresh
Benguela Current in the east, joined together by the
westward flowing South Atlantic Current. The cold and
fresh Malvinas (or Falkland) Current originates as a
branch of the ACC and meets the Brazil Current at the
confluence zone (;368S).
The mean annual fields of SST, SLP, and vector wind
are computed as the 40-yr (1953–92) climatologies from
the COADS monthly means, and are shown in Fig. 3.
The SST field exhibits a generally weak southeast–
northwest gradient north of ;358S. The isotherms ex-
tend zonally and are more closely spaced south of this
latitude. The SLP and wind fields are dominated by the
South Atlantic subtropical anticyclone whose center is
located near 308S, 58W. It is surrounded by southeasterly
trades to the north and westerlies to the south. The fol-
lowing analyses of the nonseasonal variability in this
study describe departures from these mean patterns.
3. The variability of SST and SLP: EOF analysis
The purpose of this section is to analyze indepen-
dently the variability of the oceanic surface temperatures
and the atmospheric sea level pressure during the 40-yr
period using the EOF method. By doing this, we can
2907
compare the two sets of EOFs with the set of SVD
modes obtained for the coupled atmosphere–ocean sys-
tem in section 4. Only a brief discussion of the EOF
(independent) modes of variability is given here, since
this study concentrates on the SVD (coupled) modes.
a. Oceanic variability
An EOF analysis is performed on the unnormalized
(i.e., not divided by the standard deviation) monthly
SST anomalies over the South Atlantic Ocean for the
period 1953–92. The three leading EOF modes together
account for 47% of the total monthly SST variance.
Individually, they explain 30%, 11%, and 6% of the
variance. According to North’s rule of thumb (North et
al. 1982), the difference between our third and fourth
eigenvalues is comparable to the magnitude of their
sampling errors, which means that the error in the EOFs
is comparable to the size of the EOFs themselves.
Hence, these two modes are not well separated and the
decomposition between them may not be stable. Nev-
ertheless, we keep the third mode since it will prove to
be of interest in connection with searching for links to
other well-known climatic fluctuations.
The spatial patterns associated with the first three SST
modes are depicted in Fig. 4 as homogeneous correlation
maps. They are, respectively, labeled E1(SST),
E2(SST), and E3(SST). Figure 5 shows the temporal
variability of each EOF, represented by the expansion
coefficients e1(SST), e2(SST), and e3(SST), respec-
tively.
E1(SST) (Fig. 4a) exhibits a monopole pattern ex-
tending over the entire domain. Recalling that the square
of the correlation values represents the local variance
explained (Houghton and Tourre 1992), this mode ac-
counts for up to 64% of the variance in the region of
largest loadings, namely, off the coast of southern Af-
rica. The fraction of local variance explained decreases
toward the south and west. Similar temporal and spatial
structures of the first EOF mode of SST were found by
Houghton and Tourre (1992) and by Enfield and Mayer
(1997) in the common domain (08 to 208–308S) of their
EOF analyses of the tropical Atlantic Ocean (308N to
308S).
E2(SST) (Fig. 4b) displays an out-of-phase relation-
ship between temperature anomalies north and south of
about 258–308S. This mode explains up to 20%–40%
of the variance near its centers of action.
E3(SST) (Fig. 4c) exhibits three latitudinal bands of
centers of action with alternating signs. A 208-wide band
centered around 258S is surrounded by two bands of the
opposite polarity north of 158S and south of 358S. The
temperature fluctuations in the center of action near the
coast of South Africa account for up to 20% of the total
field variance.
The time series e1(SST) and e2(SST) (Figs. 5a,b) are
characterized by a mixture of interannual and interde-
2908 JOURNAL OF CLIMATE
FIG. 4. Spatial patterns E1(SST), E2(SST), and S3(SST) of the
three EOF modes of SST. The spatial patterns are presented as ho-
mogeneous correlation maps; i.e., the contours are scaled such that
the value at each grid point is the correlation coefficient between the
time series of expansion coefficients of Fig. 5 and the SST anomaly
at that grid point. Contour interval is 0.1. Negative contours are
dashed. Zero line is thicker. Correlations higher than (in absolute
value) (a) 60.25, (b) 60.21, and (c) 60.16 are significant at the 95%
level.
cadal fluctuations, while e3(SST) (Fig. 5c) oscillates on
mainly interannual timescales.
b. Atmospheric variability
A similar EOF analysis is performed on the unnor-
malized monthly SLP anomalies over the same domain
and time period as for SST. The first three EOF modes
account for 59% of the total monthly SLP variance,
while they individually explain fractions of 35%, 16%,
and 8%. The same considerations as for SST apply here
VOLUME 10
FIG. 5. Time series of expansion coefficients e1(SST), e2(SST),
and s3(SST) of the three EOF modes of SST. The time series are
smoothed using a 13-month running mean (13MRM), and their am-
plitudes are normalized by the standard deviation.
concerning the separation between the third and fourth
modes. Once again, we keep the third mode since it is
of particular interest to our results. Figure 6 shows the
three spatial patterns E1(SLP), E2(SLP), and E3(SLP),
as homogeneous correlation maps. Figure 7 shows their
associated expansion coefficients e1(SLP), e2(SLP), and
e3(SLP).
E1(SLP) (Fig. 6a) has uniform polarity over the entire
domain. The center of action of the mode is located near
the center of the subtropical anticyclone (see Fig. 3b).
Hence, this mode describes the strengthening and weak-
ening of the anticyclone. E2(SLP) (Fig. 6b) has a dipole
structure, which is related to east–west changes in the
location of the center of the subtropical anticyclone.
Another dipole structure can be seen in E3(SLP) (Fig.
6c). In this case, changes in the position of the anti-
cyclone occur in the north–south direction.
The time series e1(SLP) (Fig. 7a) is dominated by
low-frequency (decadal-scale) oscillations, while
e2(SLP) and e3(SLP) (Fig. 7b,c) exhibit predominantly
interannual fluctuations.
c. Relating oceanic and atmospheric behaviors
To look for possible links between oceanic and at-
mospheric fluctuations, the correlation coefficients be-
tween the time series of the three discussed SST and
SLP modes are calculated and displayed in Table 1.
Correlations with the time series of the Southern Os-
cillation index (SOI) are also included to test for pos-
sible links with the El Niño–Southern Oscillation
NOVEMBER 1997 VENEGAS ET AL. 2909

FIG. 7. Time series of expansion coefficients e1(SLP), e2(SLP),

and s3(SLP) of the three EOF modes of SLP. The time series are
smoothed using a 13MRM, and their amplitudes are normalized by
the standard deviation.

(ENSO) phenomenon. The SOI is defined as the nor-

FIG. 6. Spatial patterns E1(SLP), E2(SLP), and S3(SLP) of the three
EOF modes of SLP. The spatial patterns are presented as homogeneous
correlation maps; i.e., the contours are scaled such that the value at
each grid point is the correlation coefficient between the time series
of expansion coefficients of Fig. 7 and the SLP anomaly at that grid
point. Contour interval is 0.1. Negative contours are dashed. Zero line
is thicker. Correlations higher than (in absolute value) (a) 60.18, (b)
60.16, and (c) 60.13 are significant at the 95% level.
malized sea level pressure difference between Tahiti and
Darwin. The table shows two significant correlations (at
the 95% level) between SST and SLP modes and two
significant correlations between ENSO and the third
modes of SST and SLP.
The time series e1(SLP) and e2(SST) are significantly
negatively correlated (20.34), which is also evident
from a visual comparison of the series in Figs. 7a and
5b. The fact that the first mode of the atmospheric vari-
ability is related to the second mode of the ocean vari-
ability suggests that the two are linked via atmosphere-
to-ocean forcing, since we expect a strong signal to force
a weaker one, and not the other way around. Lagged
correlations between these two expansion coefficients
(Fig. 8a) indicate that the two modes are best correlated
(20.36) when the atmosphere leads the ocean temper-
ature by 1–4 months, consistent with the idea that at-
mosphere-to-ocean forcing links these two modes of
variability.
The series e3(SLP) and e3(SST) are also significantly

TABLE 1. Correlation coefficients between the three first EOF modes of SST and SLP and between these modes and the SOI. Numbers in
brackets are the 95% significance levels calculated for each correlation (Sciremammano 1979). Highly significant correlations are in bold
face.

e1(SLP) e2(SLP) e3(SLP) SOI

e1(SST) 20.06 [0.25] 0.15 [0.16] 20.09 [0.19] 20.13 [0.24]
e2(SST) 20.34 [0.23] 0.17 [0.16] 20.18 [0.18] 0.18 [0.21]
e3(SST) 20.11 [0.15] 20.07 [0.14] 0.24 [0.17] 20.35 [0.20]
SOI 20.05 [0.17] 0.00 [0.16] 20.35 [0.20]
2910 JOURNAL OF CLIMATE
FIG. 8. (a) Lagged correlations between e1(SLP) and e2(SST). SLP leads SST for positive
lags. (b) Lagged correlations between e3(SLP) and e3(SST). SLP leads SST for positive lags.
Dashed lines show the 95% significance levels. Lags are in months. The lagged correlations are
based on the original time series (not the ones smoothed by a 13MRM).
correlated (0.24). Lagged cross correlations between the
two modes (Fig. 8b) exhibit two significant correlations.
One of these shows the atmosphere leading the ocean
by 2 months (0.30), suggesting again the presence of
an atmosphere-to-ocean forcing. The other (at ;221
months) is probably a reflection of the ;3.5-yr peri-
odicity that exists in the two time series. That is, the
two series are out-of-phase at a lag equal to half the
period of oscillation (;21 months).
The third modes of both SST and SLP are found to
be significantly correlated with the SOI (20.35). This
suggests that ENSO has a discernible signal in the South
Atlantic, even though these modes explain only a small
percentage of the variability in their respective fields
(6% and 8% for SST and SLP, respectively). Figure 9
shows lagged correlations between both third mode time
series and the SOI. The highest correlations with ENSO
are found when the variable (SST or SLP) leads the
SOI time series by 1–3 months. Significant correlations
at lag ;21 months presumably have the same origin as
those in Fig. 8b.
4. The variability of the coupled fields: SVD
analysis
To better assess and confirm the relationships between
SST and SLP variations obtained in section 3, a singular
FIG. 9. (a) Lagged correlations between e3(SST) and SOI. SST
leads SOI for positive lags. (b) Lagged correlations between e3(SLP)
and SOI. SLP leads SOI for positive lags. Dashed lines show the
95% significance levels. Lags are in months. The lagged correlations
are based on the original time series (not the ones smoothed by a
13MRM). The SOI monthly time series is smoothed by a 3MRM to
be consistent with the other data.
VOLUME 10
value decomposition (SVD) is performed on the cross-
covariance matrix between SST and SLP. In contrast to
the individual EOF analysis performed in the previous
section, the SVD analysis on the two fields together will
identify only those modes of behavior in which the SST
and SLP variations are strongly coupled.
The three leading SVD modes of the coupled SST
and SLP variations account for 89% of the total square
covariance (TSC). Each mode independently accounts
for 63%, 20%, and 6% of the TSC, respectively. We
label the spatial patterns as Sk, and the expansion co-
efficients as sk, k 5 1, 3. Table 2 displays the square
covariance fractions (SCF) explained by each mode and
the correlation coefficient (r) between the expansion co-
efficients of both variables [sk(SST) and sk(SLP)], as
indicators of the strength of the coupling. The first mode
exhibits a strong and highly significant coupling be-
tween SST and SLP. By contrast, the coupling coeffi-
cient associated with the second mode is smaller and
closer to its significance level than those of the other
two, which suggests a relatively weaker coupling. Even
though the third SVD mode explains the least variance,
its corresponding coupling coefficient is higher than that
of the second mode. This reinforces our decision to
include it in our discussion even though its associated
singular value is not well separated from the fourth one
(following North et al. 1982).
Seasonal means were computed from the monthly
SST and SLP data and the SVD analysis was repeated
to test the sensitivity of the results to the change of
seasons. The southern winter is defined as the 6-month
period from May to October, and the summer as the
TABLE 2. Square covariance fraction (SCF) and coupling correlation
coefficient between the expansion coefficients of both variables (r)
corresponding to the three leading SVD modes. In brackets are the
95% significance levels for the correlations.
Mode SCF r
1 63% 0.46 [0.29]
2 20% 0.30 [0.20]
3 6% 0.41 [0.21]
NOVEMBER 1997 VENEGAS ET AL. 2911

TABLE 3. Square covariance fraction (SCF) and coupling correlation

coefficient (r) for the three SVD modes using 6-month winters and
summers. In brackets are the 95% significance levels.

Winter Summer
Mode SCF r SCF r
1 59% 0.59 [0.40] 68% 0.74 [0.56]
2 18% 0.42 [0.42] 19% 0.52 [0.43]
3 10% 0.58 [0.44] 5% 0.68 [0.47]

FIG. 11. Same as Fig. 10 but for the second SVD mode. Correlations
higher than (in absolute value) 60.25 (SST) and 60.15 (SLP) are
significant at the 95% level.

6-month period from November to April. Table 3 shows

FIG. 10. Spatial patterns and expansion coefficient of the first SVD
coupled mode of SST and SLP. Spatial patterns are homogeneous
correlation maps. Contour interval is 0.1. Negative contours are
dashed. Zero line is thicker. Correlations higher than (in absolute
value) 60.20 (SST) and 60.19 (SLP) are significant at the 95% level.
Time series are smoothed by a 13MRM. Amplitudes are normalized
by the standard deviation.
the same parameters as Table 2 but after the seasonal
breakdown of the data.
The coupling of the three modes is most significant
during the southern summer, which is a somewhat sur-
prising result. It is likely that the seasonality existent in
the data coverage may account in part for the stronger
coupling in summer. However, this argument should ap-
ply to studies on both hemispheres, and the majority of
the Northern Hemisphere studies find the strongest at-
mosphere–ocean couplings during winter. Our interpre-
tation is that the coupling in the Southern Hemisphere
does not necessarily depend on the strong air–sea con-
trasts expected to occur during winter, since these are
2912 JOURNAL OF CLIMATE VOLUME 10

probably not as strong as in the Northern Hemisphere,

mainly due to the different distribution of land and ocean
coverage of the Southern Hemisphere. Instead, we argue
that the coupling is strongest during summer because
of possible links with major climatic oscillations ob-
served in the Northern Hemisphere, which are strongest
during the northern winter (NAO, PNA, etc.). Therefore,
the signals in both hemispheres seem to show their
strongest amplitudes simultaneously. This idea will be
explored further in the next subsections.
A difference should be made, however, when ana-
lyzing the second mode in more detail. Although the
simultaneous coupling is only significant in summer, a
lagged correlation analysis of the seasonal coupling co-
efficients shows that the coupling is similarly significant
in winter when SST leads SLP by one winter (0.51,
[0.43], not shown). Hence, in contrast to the first and
third modes, the second mode shows a similar degree
of coupling in both seasons, although the forcing seems
to act with different time lags in winter and summer.
The coupled spatial patterns and expansion coeffi-
cients corresponding to each variable and each mode
are displayed in Figs. 10, 11, and 12. Lagged cross
correlations between the SST and SLP time series as-
sociated with each mode are shown in Fig. 13. The three
modes of coupled SST and SLP variability are discussed
separately in the following sections.

FIG. 12. Same as Fig. 10 but for the third SVD mode. Correlations
higher than (in absolute value) 60.22 (SST) and 60.16 (SLP) are
significant at the 95% level.
a. First mode
Figure 10 depicts the spatial patterns and the expan-
sion coefficient of the first SVD mode of the coupled
fields. The SLP time series and pattern closely resemble
those obtained for the first EOF mode of SLP (Fig. 6a).
On the other hand, the SST time series and pattern for
the first SVD mode resemble those of the second EOF
mode of SST (Fig. 4b), with the signs reversed. Hence
this mode is related to the correlation found between
E1(SLP) and E2(SST) (Table 1). The correlation be-
tween s1(SLP) and e1(SLP) is 0.93 [0.22], while the
correlation between s1(SST) and e2(SST) is 20.63
[0.31]. Thus, the EOF mode of SLP alone provides more

FIG. 13. Lagged correlations of the coupling coefficient r corresponding to each SVD mode. SLP leads SST for positive lags. Dashed
lines show the 95% significance levels. Lags are in months. The lagged correlations are based on the original time series (not the ones
smoothed by a 13MRM).
NOVEMBER 1997 VENEGAS ET AL. 2913

FIG. 14. Correlation maps between s1(SLP) and SST gridpoint anomalies: (a) for SLP
leading SST by two months and (b) for SLP lagging SST by two months.

information on the coupling between the two fields than
the EOF mode of SST alone.
A lagged correlation analysis between s1(SST) and
s1(SLP) (Fig. 13a) indicates that the coupling is strong-
est when SLP leads SST by 1–2 months (0.51 [0 29]),
corroborating what we inferred from Fig. 8a. Therefore,
the first SVD mode seems to depict an atmosphere-to-
ocean forcing, in which the ocean response to the at-
mospheric changes appears with an intraseasonal time
lag. According to Wallace and Jiang (1987), a signature
of this kind of forcing is the marked asymmetry with
respect to the zero lag exhibited in Fig. 13a. The sig-
nificance of this asymmetry can be illustrated by con-
trasting the 12- and 22-month lag correlation maps
between s1(SLP) and the gridpoint SST anomalies
(shown in Fig. 14). In agreement with what we infer
from Fig. 13a, the SST pattern obtained when SLP leads
SST by two months (Fig. 14a) exhibits much larger
amplitudes than its counterpart pattern (Fig. 14b) ob-
tained when SLP lags SST by two months. This contrast
clearly indicates that a maximum in the amplitude of
the SST signal appears as a response to a prior maximum
in the amplitude of the SLP signal. However, the op-
posite is not true, supporting the hypothesis of an at-
mosphere-to-ocean forcing.
To illustrate the behavior of this mode of variability
in terms of atmospheric configurations, Fig. 15 shows
composites of the pressure distribution of the months
corresponding to the five highest and the five lowest
values of s1(SLP). These composites describe the at-
mospheric conditions in the two extreme situations of
the modal fluctuations. It is clear that this mode depicts
an oscillation between strengthening and weakening of
the subtropical anticyclone, accompanied by a slight
displacement of the center toward the northeast in the
weak anticyclone situations. The difference in the cen-
tral pressure between both extrema reaches 6 hPa.
Spectral analyses of both s1(SST) and s1(SLP) (see
Fig. 4 in Venegas et al. 1996) exhibit highly significant
peaks at low frequencies (14–16 yr), indicating that in-
terdecadal timescales dominate this mode. A compari-
son with the results obtained by Mann and Park (1996)
suggests the possibility of a relation between s1(SLP)
and their interdecadal joint mode in Northern Hemi-
sphere surface temperature and sea level pressure. Fig-
ure 16 shows the South Atlantic mode s1(SLP) leading
the Mann and Park interdecadal reference signal by ap-
proximately 4–5 yr. Even though the correlation be-
tween the two time series shown in Fig. 16 is not highly
significant (0.89 [0.84] using a 5-yr running mean to
smooth the data), partly due to the shortness of our time
series, we speculate that there may exist a connection.
Furthermore, Mann and Park’s analyses reveal that the
Northern Hemisphere SLP anomalies associated with
the interdecadal signal are pronounced only during the
northern winter. This is simultaneous with the strongest
expression of the South Atlantic signal during the south-
ern summer (Table 3). An independent analysis of global
surface temperature variability (Mann and Park 1994)
shows a consistent global-scale temperature signal in
phase with s1(SST), whose associated spatial pattern
exhibits large SST variability near 358S off the coast of
South America, consistent with S1(SST).
The time series s1(SLP) is found to be highly cor-
related (0.55 [0.17]) with the SLP variations at Gough
Island (40.48S, 9.98W) (as depicted in Fig. 17). This

FIG. 15. SLP composites corresponding to extreme positive and negative values of s1(SLP).
Isobars are in hPa minus 1000 hPa. Contour interval is 3 hPa.
2914 JOURNAL OF CLIMATE VOLUME 10

FIG. 17. Time series of monthly SLP anomalies at Gough Island

FIG. 16. Comparison of the South Atlantic first coupled mode with (thin line), located at 40.48S, 9.98W, and s1(SLP) (thick line). Both
the North Pacific projection of the interdecadal joint mode of Mann time series are smoothed by a 13MRM, and the amplitudes are nor-
and Park (1996). malized by the standard deviation. A roughly 15-yr period oscillation
is evident in both series.

island is located just east of the center of action of

S1(SLP) (Fig. 10c), therefore its SLP time series pro- a positive feedback between atmosphere and ocean.
vides a useful index for describing the first mode of Such an SST anomaly induces changes in the atmo-
variability in the South Atlantic. Since these observa- spheric circulation by modifying the north–south SST
tions are independent from those of COADS, the cor- gradient and these atmospheric changes in turn reinforce
relation with s1(SLP) lends support to the existence of the existent SST anomaly by air–sea heat exchange fluc-
such oscillations in the strength of the subtropical an- tuations. The resemblance between the SST and SLP
ticyclone. patterns of Fig. 10 and their North Pacific SST and SLP
To investigate possible air–sea forcing mechanisms patterns and the similar (interdecadal) timescales sug-
related to the first SVD mode, the wind anomalies as- gest that such a feedback may also be present in the
sociated with this mode are presented in Fig. 18 together South Atlantic. However, further work is needed to con-
with S1(SST) (same as Fig. 10a). The correlations are firm this hypothesis.
computed with the wind leading the SST by 1 month,
since this lag was found to depict the strongest inter-
b. Second mode
action between atmosphere and ocean (Fig. 13a). The
first EOF mode of the wind anomalies over the South Figure 11 shows the components of the second SVD
Atlantic (not shown) displays a distribution of wind mode of the coupled fields. The SST time series and
anomalies similar to that observed in Fig. 18, indicating pattern closely resemble those of the first EOF mode of
that such a wind pattern is a preferred mode of vari- SST, with signs reversed (Fig. 4a). There is also some
ability in the atmosphere. similarity between the SLP time series and pattern and
The observed relationship between the wind and SST those of the second EOF mode of SLP, again with signs
anomalies is mainly local. Stronger-than-normal trade reversed (Fig. 6b). The correlation coefficient between
winds (strong anticyclone episode) coincide with cool
ocean temperatures in the northern part of the basin,
while weaker-than-normal westerlies coincide to warm
ocean temperatures in a band between 308S and 408S,
lying west of 08 long. Also, southerly (northerly) winds
are associated with negative (positive) SST anomalies.
The local nature of the wind-SST relationship suggests
that the anomalies in the ocean temperature are driven
by wind-induced mechanisms (Wallace et al. 1990; De-
ser and Blackmon 1993; Kushnir 1994). These results
also suggest that the atmospheric forcing of the ocean
temperatures on interdecadal timescales may be more
important than previously thought. Recent work by Hal-
liwell (1996) dealing with the generation of SST anom-
alies in the North Atlantic supports this hypothesis. FIG. 18. Wind and SST anomalies correlated with s1(SST). The u-
and v-components of the wind anomalies are correlated with s1(SST)
Latif and Barnett (1994) have recently suggested that at each grid point, and ‘‘vector correlations’’ are obtained by plotting
SLP and SST anomalies in the North Pacific distributed the components in vectorial form. The SST pattern is the same as in
similarly to those in Fig. 10 could be in association with Fig. 10 (top). The wind anomalies lead the SST anomalies by 1 month.
NOVEMBER 1997 VENEGAS ET AL.
FIG. 19. Same as Fig. 15 but using s2(SLP).
s2(SST) and e1(SST) is 20.91 [0.37], while that be-
tween s2(SLP) and e2(SLP) is 20.65 [0.14]. Thus, two
facts seem to suggest that this mode of variability is
associated with an ocean-to-atmosphere forcing: (i) the
correlation betwen the two SST time series is stronger
than that between the two SLP time series, and (ii) this
coupled mode is associated with the first mode of in-
dependent (EOF) SST variability. Further support for
this hypothesis comes from the lagged coupling cor-
relation (Fig. 13b), which shows a significant simulta-
neous correlation between s2(SST) and s2(SLP) (0.30
[0.20]). The zero-lag correlation and the symmetry with
respect to the zero lag depicted by this figure is sug-
gestive of an essentially immediate adjustment of the
atmosphere to changes in the ocean (Wallace and Jiang
1987).
The SLP composites in Fig. 19 are based on the five
highest and lowest values of s2(SLP), and illustrate the
behavior of the atmospheric component in this mode.
An east–west displacement of the anticyclone center is
clearly seen in these composites. A pressure trough ap-
pears (disappears) off the Argentine coast and the cir-
culation becomes more meridional (zonal) when the an-
ticyclone is at its easternmost (westernmost) position,
which occurs at the highest (lowest) values of s2(SLP).
FIG. 20. Same as Fig. 18 but for the second mode. SST and Wind
correlations are simultaneous in this case.
2915
The central pressure remains at the same value during
the whole cycle.
Repeating the procedure used in section 4a to further
investigate the atmosphere–ocean relationship, Fig. 20
displays superimposed wind and SST correlations with
s2(SST). Both correlations are simultaneous here, since
the coupling in this mode was found to be strongest at
zero lag. To facilitate the interpretation, s2(SST) has
been multiplied by 21. Stronger-than-normal trade
winds are observed in the northeastern part of the basin
over warmer-than-normal ocean temperatures. Thus, the
SST-wind relationship cannot be explained by local
wind-induced processes, since such processes would im-
ply a negative SST anomaly being forced by anoma-
lously strong winds (Wallace et al. 1990; Deser and
Blackmon 1993; Kushnir 1994). Additionally, a lagged
correlation between wind and SST indices computed at
08, 108S [the center of action of S2(SST)] shows no
significant correlation for wind leading SST up to 10
months. This lack of evidence for wind driving the SST
anomalies also supports the ocean-to-atmosphere forc-
ing hypothesis.
The spectra of both s2(SST) and s2(SLP) (see Fig. 4
in Venegas et al. 1996) exhibit a significant peak at 6–
7 yr. A spectral coherence (or cospectrum) analysis of
s2(SLP) (only southern summer data) and the winter
NAO index (as defined in Rogers 1984) exhibits a peak
at the ;5-yr period, significant at the 95% level (not
shown). This period is consistent with the peak in the
NAO spectrum at ;7 yr found by Rogers (1984). On
the other hand, a temporal lagged correlation between
summer s2(SLP) and the NAO index suggests that the
North Atlantic signal leads the South Atlantic one by
;3 yr (Fig. 21). Even though this correlation is mar-
ginally significant at the 95% level, the evidence given
by such correlations in the frequency and time domains
suggest the possibility of a relationship between these
North and South Atlantic modes of variability. Further
work is needed to understand the nature of this con-
nection.
2916 JOURNAL OF CLIMATE
FIG. 21. Lagged correlations between the winter NAO index (De-
cember–March) and summer s2(SLP) (December–March). SLP leads
NAO for positive lags. Dashed lines show the 95% significance levels.
Lags are in years.
c. Third mode
Figure 12 shows the components of the third SVD
mode. The SLP and SST time series and patterns bare
some resemblance to the respective third EOF modes
(Figs. 4c and 6c), which were previously found to be
correlated with one another (Table 1). The correlation
coefficient between s3(SLP) and e3(SLP) is 0.55 [0.15]
and that between s3(SST) and e3(SST) is 0.40 [0.18].
Both correlations are significant but much smaller than
the ones obtained in the first two modes. Also in contrast
to the previous modes, neither the SLP nor the SST
coupled patterns appear clearly from the independent
EOF analyses.
Figure 13c shows the lagged coupling correlations of
this mode, with the highest values (;0.42 [0.21]) oc-
curing when SLP leads SST by 1–2 months. There is
also a marked asymmetry with respect to zero lag in
this figure, with significant correlations for SLP leading
SST by up to 10 months. This suggests that this mode
may be associated with an atmosphere-to-ocean forcing,
as we discussed in section 4a (Wallace and Jiang 1987).
Further support for this hypothesis comes from the com-
parison of the SST pattern with the wind anomalies
FIG. 22. Same as Fig. 15 but using s3(SLP).
VOLUME 10
related to this mode (Fig. 25). A westerly wind anomaly
associated with the strong SLP north–south gradient
around 258S [see S3(SLP) in Fig. 12] produces a weak-
ening of the climatological easterly winds (see Fig. 3b).
The weaker-than-normal winds result in smaller-than-
normal ocean-to-atmosphere heat fluxes, which produc-
es a zonally oriented positive SST anomaly in a band
around 258S. Furthermore, a lagged correlation between
SST and zonal wind indices computed at 258W, 258S
[the center of action of S3(SST)] shows a significant
correlation for zonal wind leading SST by 2 months
(0.15 [0.11]), which reveals the local nature of the wind
forcing.
The behavior of the atmospheric field in this mode
is illustrated by the SLP composites shown in Fig. 22.
A comparison of the two composites shows that a north-
ward (southward) displacement of the high pressure cen-
ter and stronger (weaker)-than-normal trade winds near
the equator correspond to high (low) values of s3(SLP).
The north (south) shift in the location of the subtropical
anticyclone leads (by 1–2 months) the warming (cool-
ing) of the ocean over the latitudinal band centered
around 258–308S and extending from coast to coast (Fig.
12a).
Based on Table 1, we anticipate a relationship be-
tween this mode and ENSO. This is confirmed in Fig.
23, which shows lagged correlations of s3(SLP) and
s3(SST) with the SOI. The highest correlations (20.37,
[0.20]) correspond to a 1-month lag, both SST and SLP
leading SOI. Since the significant correlations are neg-
ative, and negative values of SOI roughly correspond
to the mature phase of ENSO, this result suggests that
a discernible warming in the central South Atlantic oc-
curs ;1 month before the maximum warming in the
equatorial Pacific.
The spatial distributions of the SST and SLP anom-
alies in the South Atlantic during ENSO episodes are
depicted in Figs. 24a,b by correlation maps between SOI
and the grid point SST and SLP anomalies. The main
centers of action displayed by these two maps are anal-
ogous to those of S3(SST) and S3(SLP) (Fig. 12), with
the signs reversed. This lends support to the suggestion
NOVEMBER 1997 VENEGAS ET AL. 2917

FIG. 23. (a) Lagged correlations between s3(SST) and the SOI. SST leads SOI for positive lags.
(b) Lagged correlations between s3(SLP) and SOI. SLP leads SOI for positive lags. Dashed lines
show the 95% significance levels. Lags are in months. The lagged correlations are based on the
original time series (not the ones smoothed by a 13MRM). The SOI monthly time series is smoothed
by a 3MRM to be consistent with the other data.

that this mode represents the ENSO signal in the South
Atlantic. Similar ENSO-related SST and SLP spatial
patterns are shown in Fig. 16.8 of Peixoto and Oort
(1992).
The large southeasterly wind anomalies in the trades
observed in the wind pattern associated with this mode
(Fig. 25) are consistent with the findings of Enfield and
Mayer (1997), who propose that the ENSO signal in the
tropical Atlantic consists of a northward migration of
the intertropical convergence zone (ITCZ) accompanied
by strong southeasterly trades.
The power spectra of s3(SST) and s3(SLP) (see Fig.
4 in Venegas et al. 1996) exhibit significant peaks at
;4 yr, which coincides with the typical ENSO period
of oscillation. The spectral coherence between SOI and
s3(SLP) confirms their correlation in the frequency do-
main, showing a highly significant peak at 4–5 yr (not
shown)
The spectral coherence between s3(SLP) and the PNA
index (as defined in Wallace and Gutzler 1981) has two
significant peaks (not shown). The first one, at ;4 yr,
corresponds to the ENSO timescales and is consistent
with the correlation between the positive phase of the
PNA pattern and the equatorial Pacific warmings (Horel
and Wallace 1981). Hence, this connection is likely a
by-product of the fact that ENSO is the common de-
nominator. The second one is a broad significant peak
between 2.3 and 2.7 yr, which suggests a strong con-
nection between this South Atlantic mode and the PNA
on the quasibiennial timescale. Lagged temporal cor-
relations of the PNA index with s3(SLP) and s3(SST)
(Fig. 26) corroborate this connection, suggesting that
the positive phase of the PNA (low pressure over eastern
Pacific, high over North America, low over western
North Atlantic) occurs ;6–7 months after the northward
shift of the South Atlantic subtropical anticyclone and
the warming in the central South Atlantic.
A positive trend is also detectable in both s3(SST)
and 3(SLP) (Fig. 12). This trend during the 40-yr period
1953–92 is in good agreement with the 90–100 yr os-
cillation found by Mann and Park (1994, Fig. 4). They
suggest that a strong cooling trend over the northern
North Atlantic from ;1940 to the present is associated
with a warming trend (of a much weaker amplitude) in
the South Atlantic.
d. Statistical significance test
To asses the statistical robustness of the results ob-
tained from the SVD analysis, we perform here a sig-

FIG. 24. Correlation maps between the SOI and (a) gridpoint anomalies and (b) gridpoint SLP anomalies.
2918 JOURNAL OF CLIMATE VOLUME 10

FIG. 25. Same as Fig. 18 but for the third mode.

nificance test using a Monte Carlo approach. The focus

is on the SCs (the total square covariance, as well as
that corresponding to each mode) rather than the SCFs
(square covariance fractions) or the r coupling coeffi-
cients (Wallace et al. 1992). The SC is a direct measure
of the coupling between the two fields, while the SCF
and r are only meaningful when they are associated with
a significant SC. We thus test the significance of the
total SC between the SST and SLP anomalies (the sum
of the squares of all the singular values) and the SC
associated with the three first SVD modes (the square
of the corresponding singular values).
The procedure is similar to that described in Wallace
et al. (1992) and Peng and Fyfe (1996). We create a
‘‘randomized’’ SLP dataset by scrambling the SLP maps
of the 40 yr in the time domain, in order to break the
chronological order of SST relative to SLP. The scram-
bling is done on the years and not on the months; that
is, the 12 SST maps of a given year will be paired with
12 SLP maps of another (randomly chosen) year, but
keeping the order of the months inside the year. In this
way, we keep the month-to-month autocorrelation in-
herent in the time series and do not deteriorate the in-
traseasonal variability, which would lower the signifi-
cance levels and make our results appear more signif-
icant than they really are. We choose to scramble only
the SLP dataset since it has a smaller month-to-month
FIG. 27. Total SC and SC accounted for by the first three SVD
modes from the observed SST and SLP datasets (asterisks) and from
the 100 scrambled datasets (dots). All the SCs are normalized by
dividing by the number of grid points of each variable (NM 5 117
3 117).
autocorrelation than the SST data, in order to minimize
the increase of degrees of freedom inherent in the scram-
bling procedure.
We then perform an SVD analysis on the scrambled
datasets, and repeat the same procedure 100 times. An
SC value from the original datasets will be statistically
significant at the 95% level if it is not exceeded by more
than five values of the corresponding SC from the
scrambled datasets. The results of the 100 scrambled
SVD analyses are shown in Fig. 27, together with the
results of the original SVD based on the observations.
The total SC and the SC accounted for by the three
modes in the original run are plotted with an asterisk,
while the corresponding values of the 100 randomized
runs are marked with small dots. The observed SC, SC1,
and SC2 are found to be highly significant (at more than
99% level). Only two values of the scrambled runs ex-
ceeded the observed SC3, indicating that it is significant
at the 98% confidence level.

FIG. 26. (a) Lagged correlations between s3(SST) and the PNA index. SST leads PNA for
positive lags. (b) Lagged correlations between s3(SLP) and the PNA index. SLP leads PNA for
positive lags. Dashed lines show the 95% significance levels. Lags are in months.
NOVEMBER 1997 VENEGAS ET AL.
In conclusion, the results discussed in this work are
based on a statistically significant decomposition of
modes of variability of the coupled SST and SLP fields.
5. Discussion and conclusions
The three leading EOF modes of variability of the
individual SST and SLP fields in the South Atlantic
region have been identified and analyzed. The leading
modes of oceanic variability account for 47% of the
total SST variance. A slightly larger percentage (59%)
of the total SLP variance is accounted for by the leading
modes of atmospheric variability. It is interesting to note
that both decompositions leave a relatively large fraction
of the total variance unexplained.
The principal features that characterize the individual
modes of SST and SLP behavior are found again in the
SVD of the coupled fields. The main patterns of vari-
ability of both variables independently provide consid-
erable information on the coupling, but only one of the
two variables dominates each of the two first coupled
modes. While the first coupled mode keeps the char-
acteristics of the main SLP (independent) pattern, the
second coupled mode repeats the behavior of the main
SST (independent) mode, suggesting the relative im-
portance of each component of the system in each dis-
cussed coupled mode.
The first coupled mode of variability between the
ocean temperature and the atmospheric pressure can be
described as a strengthening and weakening of the sub-
tropical anticyclone, which seems to force fluctuations
in a north–south dipole structure in the ocean temper-
ature by wind-related processes. The atmospheric forc-
ing of the SST changes is detectable in the ocean with
a lag of 1–2 months. The observed fluctuations are dom-
inated by interdecadal timescales (14–16-yr period).
The atmosphere-ocean coupling is found to be strongest
during the southern summer. A possible relationship be-
tween this South Atlantic mode and interdecadal oscil-
lations of Northern Hemisphere SST and SLP is also
found. The possibility of an atmosphere–ocean feedback
in the coupling depicted by this mode is suggested.
The second mode of coupled atmosphere–ocean vari-
ability in the South Atlantic is characterized by east–
west displacements of the subtropical anticyclone center,
accompanied by large SST fluctuations over a broad
region off the coast of Africa, which occur on a 6–7-yr
timescale. This mode seems to depict an immediate re-
sponse of the atmospheric circulation to changes in the
ocean, although the coupling between the SST and SLP
fluctuations is rather weak. A significant frequency-do-
main correlation with the NAO index indicates that sim-
ilar timescales characterize the fluctuations associated
with these South and North Atlantic modes of vari-
ability.
The third South Atlantic mode is characterized by
north–south displacements of the subtropical anticy-
clone, accompanied (with a lag of 1–2 months) by SST
2919
fluctuations over a latitudinal band in the central South
Atlantic Ocean. Relatively high-frequency interannual
timescales (;4 yr) dominate the SST and SLP oscil-
lations. Evidence is presented that suggests an atmo-
spheric forcing of the ocean in this mode. However, the
interaction between the ocean and atmosphere cannot
be described by local wind-related mechanisms. The
spatial patterns and temporal behavior of the SST and
SLP anomalies are found to be strongly associated with
ENSO. The ENSO-related signal in the South Atlantic
leads the Pacific warmings by 1–2 months. A relation-
ship with the Northern Hemisphere PNA atmospheric
pattern is also suggested. A warming trend identified in
this mode may be associated with a global century-scale
oscillation.
The processes involved in the air–sea interactions de-
picted by these coupled modes of variability and the
nature of the links between Southern and Northern
Hemisphere fluctuations constitute the main subjects of
investigation for further work.
Acknowledgments. We thank Halldór Björnsson for
his input and assistance with the EOF and SVD analyses
and Michael Mann for valuable discussions and help
with the spectral analyses. Special thanks go to Yo-
chanan Kushnir for kindly reviewing the first draft of
the manuscript. We also thank Shiling Peng, John Fyfe,
David Enfield, and the two anonymous reviewers for
their help and suggestions. This work was supported by
research grants awarded to LAM and DNS from the
Canadian Natural Sciences and Engineering Research
Council and Fonds FCAR (Quebec.)
REFERENCES
Battisti, D. S., U. S. Bhatt, and M. A. Alexander, 1995: A modeling
study of the interannual variability in the wintertime North At-
lantic Ocean. J. Climate, 8, 3067–3083.
Bretherton, C. S., C. Smith, and J. M. Wallace, 1992: An intercom-
parison of methods for finding coupled patterns in climate data.
J. Climate, 5, 541–560.
Cayan, D. R., 1992a: Latent and sensible heat flux anomalies over
the northern oceans: The connection to monthly atmospheric
circulation. J. Climate, 5, 354–369.
, 1992b: Latent and sensible heat flux anomalies over the north-
ern oceans: Driving the sea surface temperature. J. Phys. Ocean-
ogr., 22, 859–881.
Davis, R. E., 1976: Predictability of sea surface temperature and sea
level pressure anomalies over the North Pacific Ocean. J. Phys.
Oceanogr., 6, 249–266.
Delworth, T., S. Manabe, and R. J. Stouffer, 1993: Interdecadal vari-
ations of the thermohaline circulation in a coupled ocean–at-
mosphere model. J. Climate, 6, 1993–2001.
Deser, C., and M. L. Blackmon, 1993: Surface climate variations over
the North Atlantic Ocean during winter: 1900–1989. J. Climate,
6, 1743–1753.
Enfield, D. B., and D. A. Mayer, 1997: Tropical Atlantic SST vari-
ability and its relation to El Niño–Southern Oscillation. J. Geo-
phys. Res., 102(C1), 929–945.
Fang, Z., and J. M. Wallace, 1994: Arctic sea ice variability on a
timescale of weeks and its relation to atmospheric forcing. J.
Climate, 7, 1897–1913.
2920 JOURNAL OF CLIMATE
Folland, C. K., D. E. Parker, and F. E. Kates, 1984: Worldwide marine
temperature fluctuations 1856–1981. Nature, 310, 670–673.
Frankignoul, C., 1985: Sea surface temperature anomalies, planetary
waves, and air-sea feedback in the middle latitudes. Rev. Geo-
phys., 23(4), 357–390.
Ghil, M., and R. Vautard, 1991: Interdecadal oscillations and the
warming trend in global temperature time series. Nature, 350,
324–327.
Halliwell, G. R., 1996: Atmospheric circulation anomalies driving
decadal/interdecadal SST anomalies in the North Atlantic. ACCP
Notes, III(1), 8–11.
Horel, J. D., and J. M. Wallace, 1981: Planetary scale atmospheric
phenomena associated with the interannual variability of sea
surface temperature in the equatorial Pacific. Mon. Wea. Rev.,
109, 813–829.
Houghton, R., and Y. Tourre, 1992: Characteristics of low-frequency
sea surface temperature fluctuations in the tropical Atlantic. J.
Climate, 5, 765–771.
Kushnir, Y., 1994: Interdecadal variations in North Atlantic sea sur-
face temperature and associated atmospheric conditions. J. Cli-
mate, 7, 141–157.
, and N. C. Lau, 1992: The general circulation model response
to a North Pacific SST anomaly: Dependence on time scale and
pattern polarity. J. Climate, 5, 271–283.
Lanzante, J. R., 1984: A rotated eigenanalysis of the correlation be-
tween 700 mb heights and sea surface temperature in the Pacific
and Atlantic. Mon. Wea. Rev., 112, 2270–2280.
Latif, M., and T. P. Barnett, 1994: Causes of decadal climate vari-
ability over the North Pacific and North America. Science, 266,
634–637.
, and , 1996: Decadal climate variability over the North
Pacific and North America: Dynamics and predictability. J. Cli-
mate, 9, 2407–2423.
Levitus, S., 1989: Interpentadal variability of temperature and salinity
at intermediate depths of the North Atlantic Ocean: 1970–1974
versus 1955–1959. J. Geophys. Res., 94, 6091–6131.
Manabe, S., and R. J. Stouffer, 1994: Multiple-century response of
a coupled ocean–atmosphere model to an increase of atmospheric
carbon dioxide. J. Climate, 7, 5–23.
Mann, M. E., and J. Park, 1993: Spatial correlations of interdecadal
variation in global surface temperatures. Geophys. Res. Lett., 20,
1055–1058.
, and , 1994: Global-scale modes of surface temperature
variability on interannual to century timescales. J. Geophys.
Res., 99, 25 819–25 833.
, and J. Lees, 1996: Robust estimation of background noise and
signal detection in climatic time series. Climate Change, 33,
409–445.
, and J. Park, 1996: Joint spatiotemporal modes of surface tem-
perature and sea level pressure variability in the Northern Hemi-
sphere during the last century. J. Climate, 9, 2137–2162.
Mehta, V. M., and T. Delworth, 1995: Decadal variability of the
tropical Atlantic Ocean surface temperature in shipboard mea-
surements and in a global ocean–atmosphere model. J. Climate,
8, 172–190.
Miller, A. J., D. R. Cayan, T. P. Barnett, N. E. Graham, and J. M.
Oberhuber, 1994: Interdecadal variability of the Pacific Ocean:
Model response to observed heat flux and wind stress anomalies.
Climate Dyn., 9, 287–302.
Molinari, R. L., D. Battisti, K. Bryan, and J. Walsh, 1994: The At-
lantic Climate Change Program. Bull. Amer. Meteor. Soc., 75,
1191–1199.
Mysak, L. A., and S. B. Power, 1992: Sea-ice anomalies in the western
VOLUME 10
Arctic and Greenland–Iceland sea and their relation to an inter-
decadal climate cycle. Climatol. Bull., 26(3), 147–176.
Newman, M., and P. D. Sardeshmukh, 1995: A caveat concerning
singular value decomposition. J. Climate, 8, 352–360.
North, G. R., T. L. Bell, R. F. Cahalan, and F. J. Moeng, 1982: Sam-
pling error in the estimation of empirical orthogonal functions.
Mon. Wea. Rev., 110, 699–706.
Palmer, T. N., and Z. Sun, 1985: A modelling and observational study
of the relationship between sea surface temperature in the north-
west Atlantic and the atmospheric general circulation. Quart. J.
Roy. Meteor. Soc., 111, 947–975.
Peixoto, J. P., and A. H. Oort, 1992: Physics of Climate. American
Institute of Physics, 520 pp.
Peng, S., and J. Fyfe, 1996: The coupled patterns between sea level
pressure and sea surface temperature in the midlatitude North
Atlantic. J. Climate, 9, 1824–1839.
, L. A. Mysak, H. Ritchie, J. Derome, and B. Dougas, 1995: On
the difference between early and middle winter atmospheric re-
sponses to sea surface temperature anomalies in the northwest
Atlantic. J. Climate, 8, 137–157.
Peterson, R. G., and L. Stramma, 1991: Upper-level circulation in
the South Atlantic Ocean. Progress in Oceanography, Vol. 26,
Pergamon Press, 1–73.
Prohaska, J. T., 1976: A technique for analyzing the linear relation-
ships between two meteorological fields. Mon. Wea. Rev., 104,
1345–1353.
Rogers, J. C., 1984: The association between the North Atlantic Os-
cillation and the Southern Oscillation in the Northern Hemi-
sphere. Mon. Wea. Rev., 112, 1999–2015.
Sciremammano, F., 1979: A suggestion for the presentation of cor-
relations and their significance levels. J. Phys. Oceanogr., 9,
1273–1276.
Stocker, T. F., and L. A. Mysak, 1992: Climatic fluctuations on the
century time scale: A review of high-resolution proxy data and
possible mechanisms. Climate Change, 20, 227–250.
Thiebaux, H. J., and M. A. Pedder, 1987: Spatial Objective Analysis
with Applications in Atmospheric Science. Academic Press, 299
pp.
Thomson, D. J., 1982: Spectrum estimation and harmonic analysis.
IEEE Proc., 70, 1055–1096.
Venegas, S. A., L. A. Mysak, and D. N. Straub, 1996: Evidence for
interannual and interdecadal climate variability in the South At-
lantic. Geophys. Res. Lett., 23(19), 2673–2676.
von Storch, H., and A. Navarra, Eds., 1995: Analyses of Climate
Variability—Applications of Statistical Techinques. Springer,
334 pp.
Wallace, J. M., and D. S. Gutzler, 1981: Teleconnections in the geo-
potential height field during the Northern Hemisphere winter.
Mon. Wea. Rev., 109, 784–812.
, and Q. R. Jiang, 1987: On the observed structure of the inter-
annual variability of the atmosphere–ocean climate system. At-
mospheric and Oceanic Variability, H. Cattle, Ed., Royal Me-
teorological Society, 17–43.
, C. Smith, and Q. R. Jiang, 1990: Spatial patterns of atmosphere–
ocean interaction in the northern winter. J. Climate, 3, 990–998.
, , and C. S. Bretherton, 1992: Singular value decomposition
of wintertime sea surface temperature and 500-mb height anom-
alies. J. Climate, 5, 561–576.
Woodruff, S. D., R. J. Slutz, R. L. Jenne, and P. M. Steurer, 1987:
A Comprehensive Ocean–Atmosphere Data Set. Bull. Amer. Me-
teor. Soc., 68, 521–527.
Zebiak, S. E., 1993: Air–sea interaction in the equatorial Atlantic
region. J. Climate, 6, 1567–1586.