## 1. Introduction

During the last decade, the scientific community has shown an increasing interest in Pacific decadal climate variability. This variability is characterized by decadal variations of the intensity of the Aleutian low, positive (negative) SST anomalies in the central and western extratropical Pacific, and negative (positive) anomalies in the tropical and eastern extratropical Pacific (see Graham 1994; Trenberth and Hurrell 1994; Zhang et al. 1997; Mantua and Hare 2002). In the ocean, Pacific decadal variability is characterized by significant variations in gyre-scale ocean circulation. For example, the decadal variations of temperature at a depth of 200– 400 m result from decadal shoaling (deepening) of thermocline in the western extratropical Pacific, corresponding to the intensification (weakening) of subpolar and subtropical gyres (see Miller et al. 1998; Deser et al. 1999), and meridional shifts of the Kuroshio Extension (Seager et al. 2001).

The mechanism of decadal variability in the Pacific Ocean is the subject of debate. A number of theories explaining Pacific decadal variability were recently proposed (see Miller and Schneider 2000): passive ocean response to atmospheric forcing; ocean–atmosphere coupling in the Tropics and extratropics; and different types of tropical–extratropical interactions, for example, ocean and atmospheric teleconnections or advection of anomalies by the mean circulation.

It is known that tropical ocean–atmosphere coupling causes strong variability on interannual time scales, that is, the El Niño–Southern Oscillation (ENSO) phenomenon. Unlike the tropically confined ENSO, Pacific decadal variability has a wider latitudinal extent (Zhang et al. 1997). The mechanism of extratropical ocean–atmosphere coupling is poorly understood today, and it is not clear if it plays an important role in producing midlatitude decadal variability. A number of recent studies suggest that extratropical coupling is weak, in which case the decadal variability would be a result of the oceanic response to stochastic variations in atmospheric forcing (see Davis 1976; Hasselman 1976). This assessment implies that the coupling feedbacks are of secondary importance to midlatitude decadal variability.

Frankignoul et al. (1997) used a simple linear model with planetary geostrophic dynamics in a semi-infinite basin (without a western boundary) to estimate the decadal response of the extratropical ocean to stochastic, zonally independent wind forcing. They found that the baroclinic response has a red spectrum with large amplitude at periods longer than Rossby wave transit time across the ocean. Jin (1997) and Qiu (2003) used a planetary wave approximation to study the ocean response to stochastic in time forcing with basin-scale zonal variation, and found that the response has a spectral peak at the frequency of the basin-scale Rossby wave.

Cessi and Primeau (2001) showed that the linear quasigeostrophic formulation in a bounded ocean allows for the existence of the low-frequency basin modes that are weakly damped in the presence of weak dissipation. These modes can be resonantly excited by the time-dependent forcing. Their frequencies are integer multiples of the gravest mode frequency, which has a period equal to the longest Rossby wave transit time across the basin. The shallow water formulation by Cessi and Louazel (2001) confirmed the existence of the preferred frequencies determined by the basin eigenmodes, and the ability of the stochastic forcing to produce spectral power near the frequency of the gravest mode. The response of the bounded ocean basin to zonally independent stochastic in time forcing was studied in detail using a geostrophic analytical model by Liu (2003), and using a numerical shallow water model by Yang et al. (2004). In both studies, the spectral peak of the response is located near the frequency of the gravest basin mode.

Thus, there are two different theories for the ocean response to stochastic forcing: excitation of long Rossby waves in a semi-infinite basin as in Frankignoul et al. (1997), and excitation of ocean basin modes in a bounded basin as in Cessi and Louazel (2001). In the absence of a western boundary, the forcing must have specific structure in order to produce a spectral peak at decadal frequency (e.g., Jin 1997; Qiu 2003). When a western boundary is present, a spectral peak emerges naturally due to excitation of ocean basin modes. Both theories described above provide potential mechanisms for decadal variability; however, the related studies were done with very simplified models and comparison of the results of these studies with observations is difficult. Additional studies of basin-mode properties and the oceanic response to stochastic forcing with more realistic models are necessary to understand which of the two mechanisms described above approximate decadal variability better.

Several types of low-frequency ocean basin modes have been identified in analytical and simple numerical models. Using classical equatorial wave dynamics, Cane and Moore (1981) found the equatorial modes. Yang and Liu (2003) found the planetary and Kelvin modes using a shallow water model. Low-frequency basin modes are the slowest decaying modes in the ocean. Since they have decadal periods given basin scales of thousands of kilometers, they may be responsible for extratropical decadal variability. The planetary basin modes are composed of westward-propagating Rossby waves at midlatitudes and equatorial and coastal Kelvin waves wrapping around the edge of the basin (Fig. 1). The damping of planetary modes is due to dissipation along the western boundary where incoming long Rossby waves are reflected into coastal Kelvin waves and short Rossby waves. The frequency of the gravest mode is determined by the long Rossby wave transit time across the basin. The maximum amplitudes of the low-frequency planetary modes are confined to the northwestern part of the basin (see Figs. 7 and 8).

Previous studies related to the resonant excitation of basin modes by stochastic atmospheric forcing were typically done using analytical approximations or simplified numerical models. Comparing the results of these ocean-basin-mode studies with observed or GCM-simulated variability in the Pacific Ocean is difficult due to the extreme simplifications made in either the numerical models or the analytic approximations. Typically, the basin modes were derived in the absence of any mean circulation and in rectangular ocean geometry, despite the fact that both factors can significantly change Rossby wave propagation and the resulting basin-mode properties. Analytical models simplify the dynamics of the ocean, leaving only the subset of basin modes. In numerical models, basin modes were found by putting the discrete model operators in matrix form and finding the eigenvectors of the matrices directly. Simplifications were necessary because the calculation of the eigenvalues in a complex domain is computationally expensive. To understand the role of basin modes in decadal variability and their dependence on the ocean geometry and on the mean flow, we developed a method for finding multiple eigenvectors of a linear numerical model, corresponding to the most unstable eigenvalues, without involving complex matrix manipulations. The method is based on the *breeding* technique. The breeding method (also referred to as the initial value method in the literature) was used by Brown (1969) and by Simmons and Hoskins (1976) to find the most unstable mode in linearized atmospheric models. Later, this method was applied to a nonlinear atmospheric model (Toth and Kalnay 1993), and to coupled general circulation models (Yang et al. 2006; Vikhliaev 2006). Here we combine the breeding method with the linear least squares method and adjoint technique to extract multiple basin modes from the solution of a linear ocean model by iterative elimination of the least stable modes. To study the dependence of basin modes on the model formulation we applied our method to a reduced-gravity linear shallow water model with realistic ocean geometry and idealized mean flow.

The next section will give a description of the breeding method for finding planetary basin modes in an ocean model (see also Vikhliaev 2006 for details). The results of the application of the method to a linear ocean model with realistic geometry and mean circulation will be given in section 3.

## 2. Finding basin modes using the breeding method

As was mentioned above, the ocean basin modes are damped. The decay rates of different modes are different. This allows us to use the breeding method for separation of the slowest decaying modes. Note that the application of the breeding method for finding basin modes in an ocean model is different from the application of this method to an atmospheric model as in Brown (1969) since the atmospheric modes are growing. In addition to the breeding developed by Brown (1969), we use a linear least squares method for finding the eigenvalues and eigenvectors corresponding to the slowest modes, and the properties of the adjoint modes for excluding the slowest modes from the model output. The latter two steps allow us to find multiple basin modes.

### a. Breeding

**Ψ**(

*x*,

*y*,

*t*) is the vector of model variables and

*breeding*of most slowly decaying perturbations was developed in the scope of the present work to avoid the computational difficulty noted above.

The basin modes have different decay rates, except for complex conjugate pairs. Modes whose eigenvalues have a larger real part decay slower, while modes whose eigenvalues have a smaller real part decay faster. Typically, the breeding method involves two integrations: a control run and a perturbed run. The difference between these two integrations is a random initial perturbation. During the integration of the perturbed run the initially random perturbation converges to the fastest growing or slowest decaying mode. Since in our experiments with a linear ocean model described later the mean circulation is zero or stationary and we explicitly solve for anomalies with respect to the mean flow, no control integration is necessary. The model is initialized with a random perturbation added to the mean state and integrated until a few of the slowest decaying modes dominate over others. To avoid a computational underflow the solution is rescaled when the energy of perturbation drops below a certain level. Since the model is linear, a choice of the amplitude of perturbation and rescaling coefficient is arbitrary. In our experiments we multiply the solution by 3.5 when the energy of perturbation is decreased by a factor of 3.5^{2} from the initial level. An example of the amplitude of the basin-mode thermocline anomaly in the reduced-gravity shallow water model is shown in Fig. 2. The interval between two consecutive rescalings is called the *breeding cycle*. Note that the breeding cycle is not uniform in Fig. 2. This is a consequence of the presence of several modes in the dataset. It is possible to integrate the model long enough to leave only the single slowest decaying mode in the output. However, if two modes have similar slow decay rates, a relatively long integration time is needed to separate them. In this case, it is possible that the computational advantage of the breeding method over direct calculation of eigenvectors will be negated. To solve this difficulty, we extended the method of Brown (1969) and combined the breeding with a technique that allows us to extract multiple basin modes from a model output.

### b. Finding slowest basin modes in model output

*σ*are the complex frequencies of the basin modes,

_{m}**Ψ̂**

_{m}are the complex amplitudes of the basin modes,

*f*(

*t*) is equal to the number of rescalings made up to time

*t, k*is the rescaling coefficient, and

*N*is a small number of modes (e.g.,

*N*< 10) presumed to contribute to the sum. We describe below how to determine

*N*from the model output. Here and in the following sections we define the “gravest modes” as the slowest decaying modes, which are not necessarily the lowest-frequency modes.

**Ψ̂**

_{n}and frequencies

*σ*on the right-hand side of (4). Equation (4) is linear relative to the basin-mode amplitudes and nonlinear relative to the basin-mode frequencies, but the problem can be converted into a linear problem using theorem (1) (see the appendix). Let

_{n}**Ψ**(

*t*) be an element of vector

**Ψ**(

*x*,

*y*,

*t*) at some point (

*x*,

*y*). The function Ψ(

*t*) satisfies the

*N*th order autoregressive equation: where

*N*must be the same as the number of modes in the solution. The multiplication by

*k*

^{f}^{(}

^{t}^{)}in Eq. (4) and by

*k*

^{f}^{(}

^{t}^{)−}

^{f}^{(}

^{t}^{+}

^{n}^{Δ}

^{t}^{)}in Eq. (5) is necessary to compensate for the effect of rescaling. If the model output Ψ(

*t*) is known, the coefficients

*a*in Eq. (5) can be found with the linear least squares method. The roots of the characteristic polynomial are related to the eigenvalues of the model operator as The roots of the characteristic polynomial and corresponding eigenvalues can be either real or complex conjugate pairs. The real eigenvalues correspond to the purely decaying (stationary) modes; the complex eigenvalues correspond to a decaying oscillation or propagating modes. The roots of the polynomial (6) can be found numerically. The real and imaginary parts of Eq. (7) can be written as Solving (8) and (9) for

_{n}*σ*and

_{r}*σ*gives In Eqs. (8)–(11) we dropped the indexes

_{i}*m*, but there is a set for each

*m*.

Special attention should be paid to choosing the order of the autoregressive model (5), *N*, and the time step Δ*t*. If the order of the model is chosen less then the number of modes in the dataset, then Eq. (5) will not be satisfied. Figure 3 shows an example from an integration of the reduced-gravity shallow water model. The time series shown in Fig. 3 is a superposition of two decaying oscillations (two basin modes) with a periodic rescaling. The rescaling does not allow the amplitude of the solution to drop below the machine precision. We describe below how to find the appropriate order of the autoregressive model to obtain the decay rate and period of the modes in Fig. 3.

Figure 4 shows the scatterplot of the time series shown in Fig. 3 at time *t* versus the same time series at time *t* + Δ*t*. Clearly, *h*(*t*) and *h*(*t* + Δ*t*) cannot be fitted to a straight line, and this time series cannot be approximated with an autoregressive model of the first order. A similar problem arises when the order of the regression model is greater than 1, but too low to represent all the modes in the dataset.

*m*th component of vector

**x**

*is and*

_{n}**a**is a vector of regression coefficients. Figure 5 shows the dependence of the regression error on the order of the autoregressive model used to fit the time series shown in Fig. 3. The regression error is calculated as the ratio of the sum square error SSE = (

**x**

_{0}− 𝗠 ·

**a**)

^{T}· (

**x**

_{0}− 𝗠 ·

**a**) to the sum square total SST =

**x**

^{T}

_{0}·

**x**

_{0}. The regression error is always decreasing with an increase in the model order. Notice the sharp drop in the error by a factor of 10

^{4}in going from order 3 to order 4, indicating that the order of the regression model should be at least 4 to fit the example dataset.

If the order of the autoregressive model is chosen to be larger than the number of modes in the time series, the regression matrix 𝗠 becomes ill conditioned, since its columns become linearly dependent. In the latter case, some of regression coefficients *a _{n}* are not related to the eigenvalues

*σ*. Figure 6 shows the dependence of the condition number of the matrix 𝗠 on the order of regression model used to fit the time series shown in Fig. 3. When the order of the model is increased from 1 to 4, the condition number of 𝗠 stays below 10. Further increase in the model order results in a sharp increase in the condition number of 𝗠 to 1000, indicating that the regression matrix becomes ill conditioned, and its columns are linearly dependent. Finally, the dependence of the regression error on the order of the model requires the order to be at least 4, whereas the dependence of the condition number of the regression matrix on the order of the model requires the order to be less than or equal to 4. Given the above considerations, we conclude that the example dataset represents four different basin modes (two complex conjugate pairs), and the regression model of the order 4 should be used to find four different eigenvalues.

_{n}Choosing a time step in the regression model (5) that is too short or too long may result in a poor representation of the variability of the time series. A suitable strategy for choosing a time step is to use a plot of the time series to roughly identify a period *T*, and then take a time step Δ*t* ≈ *T*/2*N*, where *N* is the order of autoregressive model used to fit the data. For example, to fit the dataset shown in Fig. 3 with the model of order 4, Δ*t* = 13.7 yr (5000 days) was taken. The resulting eigenvalues are two complex conjugate pairs with the parts corresponding to *e*-folding decay scales of 38.7 and 38.1 yr, respectively, and imaginary parts corresponding to periods of 88.4 and 44.2 yr, respectively. When the eigenvalues *σ _{n}* are known, Eq. (4) can be solved for the amplitudes

**Ψ̂**

_{n}of eigenmodes at each point using the least squares method. [The eigenvectors corresponding to the eigenvalues calculated in the example above are shown in Fig. 14 (middle and bottom).]

### c. Finding faster modes

*σ*and

_{k}**Ψ̂**

_{k}are the eigenvalues and eigenvectors of the operator

*γ*and

_{l}**Φ̂**

_{l}are the eigenvalues and eigenvectors of the adjoint operator

The slowest eigenmodes of the adjoint model are found by integrating the adjoint model and following the breeding steps described in previous section. The integration of the adjoint model takes the same time as the integration of the ocean model, since adjoint operators have the same set of eigenvalues, and corresponding basin modes have the same decay rates and frequencies. After finding the slowest basin modes and corresponding adjoint modes, their orthogonality allows us to remove these modes from the solution during the further model integration.

*a*(

_{n}*t*) can be found by projecting the Eq. (17) onto the

*n*th adjoint eigenmode: where

*N*is the number of the known modes. Then,

*N*known modes can be removed from the solution during the model integration: making the faster decaying modes dominant and keeping their amplitudes finite. The procedure described above can be repeated iteratively to extract as many basin modes as necessary.

## 3. Application of breeding technique to a linear ocean model

To validate our approach, we have applied the breeding method to the shallow water reduced-gravity beta-plane model described in Cessi and Louazel (2001) and compared the result with what they found solving the eigenvalue problem directly (see Figs. 7 and 8).^{1} The similarities between Figs. 8 and 7 are noteworthy and demonstrate the utility of the breeding technique.

The inviscid shallow water equations linearized about zero mean flow conserve energy. Therefore, the decay of the decadal modes arises purely as a result of dissipation. This raises a question of whether the structure of the decadal modes is determined by friction. In our test model [same as in Cessi and Louazel (2001)], dissipation has the form of linear drag with drag coefficient *r* = 8.0 × 10^{−7} s^{−1}. To determine the dependence of decadal basin modes on the form of dissipation, we repeated the mode calculation using the same model, but with horizontal diffusion *A _{H}*∇

^{2}

**u**instead of linear drag. The horizontal diffusion coefficient

*A*was chosen to make the width of the Munk boundary layer

_{H}*δ*> =

_{M}^{3}

*A*/

_{H}*β*

*δ*=

_{S}*r*/

*β*≈ 50 km in the test model, that is,

*A*=

_{H}*r*

^{3}/

*β*

^{2}≈ 2000 m

^{2}s

^{−1}. The resulting decadal modes are shown in Fig. 9. The modes resulting from linear drag (Fig. 8) and the modes resulting from horizontal diffusion (Fig. 9) have identical patterns except in the regions near the western and northern boundaries, where diffusion generates a recirculation. Cessi and Louazel (2001) derive decaying modes using the analytical geostrophic approximation with no explicit dissipation, but there is an implicit energy sink, since they allow no mass flow through the western boundary, but do allow net energy flow through the boundary. The resulting modes have patterns and decay rates similar to those in numerical shallow water model with the explicit dissipation. Thus, the leading decadal modes are nearly insensitive to the form of the dissipation and so they are robust.

Note that in our test model we use a resolution of 100 km, while the width of frictional boundary layer is 50 km; therefore, the boundary layer is not resolved. We calculated the leading modes using the model with resolution of 25 km, resolving the boundary layer, and obtained the modes with patterns, decay rates, and frequencies exactly the same as shown in Fig. 8. Given the extent of the basin of 7500 km such a calculation requires 300 × 300 grid cells and would be impossible to solve with the matrix method on currently available computers.

Since the breeding method for calculating basin modes does not require any matrix manipulations, it may be applied to more realistic ocean models than the model used by Cessi and Louazel (2001).

*ϕ*and*λ*are latitude and longitude;*u*,*υ*are zonal and meridional velocities;*h*is the thermocline depth anomaly with respect to the mean thermocline depth*H*(*ϕ*,*λ*);*g*′ = 0.002 m s^{−2}is the reduced gravity;*f*= 2Ω sin*ϕ*is the Coriolis parameter;*R*is the earth’s radius;*ρ*= 1000 kg m^{−3}is density;*r*= 8.0 × 10^{−7}s^{−1}is the linear drag coefficient.

The mean thermocline depth *H* is taken to be 500 m in the experiments without the mean flow, and is derived from the nonlinear shallow water model with static wind stress forcing in the experiments with the mean flow.

Figure 11 shows the mode that is adjoint to the mode in Fig. 10. The adjoint mode looks like the east–west reflection of the mode itself, since Rossby and Kelvin waves in the adjoint model propagate in the directions opposite to those shown in Fig. 1 (i.e., from the west to the east and clockwise).

The effect of the nonzero mean flow on the basin modes is demonstrated in Figs. 12 and 13. The mean flow that is shown in Fig. 12 results from a latitudinally sinusoidal zonal wind stress, and simulates the subpolar and subtropical North Pacific gyres. If the mean flow is strong enough to stop the westward propagation of the long Rossby waves, the first basin mode is purely decaying. This result is consistent with the study of Spydel and Cessi (2003). The structure of the purely decaying mode is similar to the structure of the subpolar gyre. The second basin mode is oscillatory with a period of 7 yr. Its amplitude has maximum value in the region between the subpolar and subtropical gyres.

The planetary basin modes in the equatorially symmetric basin are equatorially symmetric themselves (e.g., see Fig. 2 in Cessi and Otheguy 2003). The global ocean does not have equatorially symmetric structure and consists of several connected basins. Figure 14 shows the thermocline anomaly field for the gravest planetary modes resulting from the global shallow water model. The maximum amplitudes of these modes are confined to the southern part of the ocean around the Antarctic coast. The dissipation of the Rossby waves is minimal around the Antarctic since they do not hit any boundaries. The first mode is purely decaying, and has almost no zonal variation. The gravest oscillatory modes have 2*π* and 4*π* structures around the Antarctic.

We continued the extraction of the planetary basin modes in the global model, to examine if any of the more rapidly decaying global modes have the structure that is similar to the North Pacific modes from the regional model. Figure 15 shows the eigenvalues of the slowest basin modes from the global model (circles) and from the model with the North Pacific geometry (crosses). Notice that the modes marked with numbers 1 and 2 have similar frequency. The real part of the mode marked with number 1 is shown in Fig. 16, and the spatial structure of the mode marked with number 2 is shown in Fig. 10. The spatial structures of the latter two modes in the North Pacific region are similar, as are their frequencies. Notice that, although the spatial and temporal characteristics of the particular basin mode from the global model resemble the corresponding characteristics of the gravest mode from the regional model, the regional model completely misses a set of the dominant global modes. Since a number of the gravest global basin modes are absent from the regional model, the response of the regional North Pacific model to stochastic wind stress forcing may be different than the response of the global model to the same forcing.

Unlike the regional North Pacific model, the global model has a stationary mode, shown in Fig. 14, and a very low frequency mode, shown in Fig. 17. The period of the latter mode is 250 yr, and its spatial structure represents the North Atlantic Ocean oscillating in antiphase with the Indo–Pacific Ocean.

## 4. Summary

The breeding method described in section 2 allows finding multiple basin modes in linear ocean models. The method has the advantage over finding the eigenvalues and eigenvectors of the model operator directly, since it does not involve any computationally expensive operations, except for the model integration. The breeding technique allows us to separate the rapidly decaying modes and the slowly decaying modes. If the most slowly decaying mode is well separated from the other modes, the model is integrated until the most slowly decaying mode dominates the solution, then its eigenvalue and eigenvector are estimated with the linear least squares method (see section 2). If several modes have similar slow decay rates, then their eigenvalues can be estimated by taking a larger order of the autoregressive model formulated by Eq. (5). The residual of the least squares fit of slowest modes cannot be used to estimate the faster modes accurately, since it has a rather small amplitude. Instead, the eigenmodes of the adjoint model are used to remove the slowly decaying modes from the solution during the model integration, making the faster modes dominant. The method can be applied to the model iteratively to find as many slowest modes as necessary.

The disadvantage of the proposed method is in the need to subjectively choose the parameters [the time step Δ*t* and the order *N* of the autoregressive model (5)]. For the choice of the order of the autoregressive model, two conditions must be satisfied: the regression error must be small, and the regression matrix must be well conditioned. If it is not possible to satisfy these two conditions simultaneously, the model should be integrated for a longer period for a better separation of the fast and slow basin modes. The time step should be chosen to sample the variability in the time series properly, and the plot of time series can be used as a guide.

The method developed here was tested with a linear reduced-gravity shallow water model, described in section 3. It was found that changing the basin geometry and introducing the mean flow can qualitatively change the dominant planetary basin modes. Some of the global basin modes may resemble the gravest North Pacific basin modes in temporal and spatial characteristics; however, the dominant modes in the global model are different from the dominant modes in the North Pacific model. If the variability in the ocean is defined by the dynamics of the gravest basin modes, the response of the global ocean and the response of the North Pacific Ocean to the same forcing may be different.

The basin modes for the global ocean were calculated using a shallow water model; however, the results shown in Fig. 14 cannot be considered realistic, since the global model did not have a realistic mean circulation. The global ocean circulation incorporates the ventilated thermocline dynamics, and cannot be correctly simulated with the 1.5-layer shallow water model. To calculate the realistic global planetary basin modes, a more complex model that takes into account vertical stratification should be used.

It was mentioned in the introductory section that the excitation of the ocean basin modes by stochastic atmospheric forcing is one possible mechanism for decadal climate variability. Though the low-frequency basin modes are dominant in the shallow water model, it is not clear if a complex system such as a general circulation model have the same dominant modes. For example, if the propagation of coastal Kelvin waves is blocked by strong dissipation in the western boundary layer, the planetary basin modes do not exist, and the excitation of long Rossby waves as in Frankignoul et al. (1997) is a likely mechanism for decadal variability. In addition to a passive ocean response to atmospheric forcing, low-frequency coupled modes are possible in a coupled GCM. To answer the question to what extent the decadal variability in a coupled system can be explained by ocean basin modes, low-frequency modes should be studied using a realistic coupled model. An application of the breeding method to a coupled GCM for finding slowly varying dynamical modes is a subject for future work.

## Acknowledgments

This research was supported by grants from the National Science Foundation ATM-0332910, the National Ocean and Atmosphere Administration NA04OAR4310034 and NA05OAR4311135, and the National Aeronautics and Space Administration NNG04GG46G.

## REFERENCES

Brown, J., 1969: A numerical investigation of hydrodynamic instability and energy conversions in the quasi-geostrophic atmosphere. Part I.

,*J. Atmos. Sci.***26****,**352–365.Cane, M. A., , and Moore D. W. , 1981: A note on low-frequency equatorial basin modes.

,*J. Phys. Oceanogr.***11****,**1578–1584.Cessi, P., , and Louazel S. , 2001: Decadal oceanic response to stochastic wind forcing.

,*J. Phys. Oceanogr.***31****,**3020–3029.Cessi, P., , and Primeau F. , 2001: Dissipative selection of low-frequency modes in a reduced-gravity basin.

,*J. Phys. Oceanogr.***31****,**127–137.Cessi, P., , and Otheguy P. , 2003: Oceanic teleconnections: Remote response to decadal wind forcing.

,*J. Phys. Oceanogr.***33****,**1604–1617.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.Deser, C., , Alexander A. , , and Timlin M. S. , 1999: Evidence for a wind-driven intensification of the Kuroshio Current from the 1970s to the 1980s.

,*J. Climate***12****,**1697–1706.Frankignoul, C., , Müller P. , , and Zorita E. , 1997: A simple model of the decadal response of the ocean to stochastic wind forcing.

,*J. Phys. Oceanogr.***27****,**1533–1546.Graham, N. E., 1994: Decadal-scale climate variability in the tropical and North Pacific during the 1970s and 1980s: Observations and model results.

,*Climate Dyn.***10****,**135–162.Hasselman, K., 1976: Stochastic climate models: Part I. Theory.

,*Tellus***28****,**473–485.Jin, F. F., 1997: A theory of interdecadal climate variability of the North Pacific ocean–atmosphere system.

,*J. Climate***10****,**1821–1835.Liu, Z., 2003: Tropical ocean decadal variability and resonance of planetary wave basin modes. Part I: Theory.

,*J. Climate***16****,**1539–1550.Mantua, N. J., , and Hare S. R. , 2002: The Pacific decadal oscillation.

,*J. Oceanogr.***58****,**35–44.Miller, A. J., , and Schneider N. , 2000: Interdecadal climate regime dynamics in the North Pacific Ocean: Theories, observations and ecosystem impacts.

*Progress in Oceanography,*Vol. 47, Pergamon Press, 355–379.Miller, A. J., , Cayan D. R. , , and White W. B. , 1998: A westward-intensified decadal change in the North Pacific thermocline and gyre-scale circulation.

,*J. Climate***11****,**3112–3127.Primeau, F., 2002: Long Rossby wave basin-crossing time and the resonance of low-frequency basin modes.

,*J. Phys. Oceanogr.***32****,**2652–2665.Qiu, B., 2003: Kuroshio extension variability and forcing of the Pacific decadal oscillations: Responses and potential feedback.

,*J. Phys. Oceanogr.***33****,**2465–2482.Seager, R., , Kushnir Y. , , Naik N. H. , , Cane M. A. , , and Miller J. , 2001: Wind-driven shifts in the latitude of the Kuroshio–Oyashio extension and generation of SST anomalies on decadal timescales.

,*J. Climate***14****,**4249–4265.Simmons, A., , and Hoskins B. , 1976: Baroclinic instability on the sphere: Normal modes of the primitive and quasi-geostrophic equations.

,*J. Atmos. Sci.***33****,**1454–1477.Spydel, M., , and Cessi P. , 2003: Baroclinic modes in two-layer basin.

,*J. Phys. Oceanogr.***33****,**610–622.Toth, Z., , and Kalnay E. , 1993: Ensemble forecasting at NMC: The generation of perturbations.

,*Bull. Amer. Meteor. Soc.***74****,**2317–2330.Trenberth, K. E., , and Hurrell J. W. , 1994: Decadal atmosphere-ocean variations in the Pacific.

,*Climate Dyn.***9****,**303–319.Vikhliaev, Y., 2006: Extra-tropical Pacific decadal variability. Ph.D. thesis, George Mason University, 120 pp.

Yang, H., , and Liu Z. , 2003: Basin modes in tropical–extratropical basin.

,*J. Phys. Oceanogr.***33****,**2751–2763.Yang, H., , Liu Z. , , and Zhang Q. , 2004: Tropical ocean decadal variability and resonance of planetary wave basin modes. Part II: Numerical study.

,*J. Climate***17****,**1711–1720.Yang, S-C., , Cai M. , , Kalnay E. , , Rienecker M. , , Yuang G. , , and Toth Z. , 2006: ENSO bred vectors in coupled ocean–atmosphere general circulation models.

,*J. Climate***19****,**1422–1436.Zhang, Y., , Wallace J. M. , , and Battisti D. , 1997: ENSO-like interdecadal variability: 1900–93.

,*J. Climate***10****,**1004–1020.

## APPENDIX

*N*modes where frequencies

*σ*are related to roots of the characteristic polynomial as for

_{m}*m*= 1, . . . ,

*N*.

^{1}

In their paper Cessi and Louazel (2001) did not specify the resolution of their model. We will show later that the properties of decadal basin modes do not depend on resolution. In our test experiment we used a resolution of 100 km.