1. Introduction

The scientific community is now taking on the challenge of using initialized models to produce time-evolving climate predictions for the next 10–30 yr (Smith et al. 2007; Keenlyside et al. 2008; Pohlmann et al. 2009). Such predictions will be a key component of the next Intergovernmental Panel on Climate Change (IPCC) assessment report (Taylor et al. 2009). Compared with traditional climate change experiments, the fundamental difference in these forecasts is that the initial ocean state is determined from observations, and the hypothesis is that the resulting forecasts will substantially benefit from this added information. But the duration of the influence of the ocean initial conditions remains unknown. Since the climate system is chaotic, inevitable errors in the initial conditions grow with time causing the initial signals to fade (Lorenz 1963). Eventually, the impact of the initial conditions become undetectable, placing a fundamental limit on its influence. If one considers a situation where the forcing of the climate system is changing, a second limit on initial condition influence should be introduced. For, if, as in the case with forcing by the ongoing changes in greenhouse gas (GHG) and aerosol concentrations, the system response increases with time, then at some point the influence of the initial conditions becomes of secondary importance compared to the forced response. In this paper, we quantify the forecast range at which these two limits are reached. Our results should help to determine the feasibility and value of decadal predictions (Meehl et al. 2009; Hurrell et al. 2009; Solomon et al. 2011).

Since the observational record is so short, there are no methods of measuring the predictability limits of the natural system. But one can estimate these limits for the numerical models that are used to simulate and predict nature. In general, predictability properties can be different for two dynamical systems that in many ways appear to be similar. Hence, studies, like ours, that quantify the initial-value predictability of a particular model are not necessarily quantifying the predictability of nature. But just as the predictability limits of nature impose limits on the range at which the initial conditions can influence forecasts, so do the predictability limits of the models used to predict nature. Hence, the predictability of models needs to be quantified. And to the extent that a given model is a good surrogate for nature, its predictability limits give some indication of the predictability limits of nature.

Because initial-value predictability concerns how rapidly a cluster of similar initial states evolves to a distribution that is statistically indistinguishable from the system’s climatological distribution, a common approach for quantifying initial-value predictability of a model is to perform ensemble experiments with perturbed initial conditions.¹ Most previous studies have focused on the North Atlantic, particularly the Atlantic meridional overturning circulation (AMOC; Griffies and Bryan 1997a,b; Collins 2002; Collins and Sinha 2003; Pohlmann et al. 2004; M. Collins et al. 2006). Many of these investigations concur that the AMOC is potentially predictable a decade in advance, but the characteristics of the AMOC, including its predictability limits, vary from model to model (M. Collins et al. 2006; Latif et al. 2006; Hurrell et al. 2009). Like the North Atlantic, the North Pacific exhibits strong decadal variability [with the dominant mode called the Pacific decadal oscillation (PDO); Mantua et al. (1997)]. While many studies suggest that the North Pacific also has decadal predictability as a result of ocean Rossby wave propagation (Latif and Barnett 1994; Kwon and Deser 2007; Schneider et al. 2002; Sugiura et al. 2009), ocean advection processes (Saravanan and McWilliams 1998), or tropical–extratropical interaction (Gu and Philander 1997), only a few have attempted to quantified its predictability limit. Our recent ensemble experiments indicate that the PDO, which is EOF1 of both SST and the subsurface temperature intrinsic variability in the model we studied, is only predictable for less than 6 yr (Teng and Branstator 2010), but EOF1 has a tendency to evolve into EOF2 through ocean advection. In combination, these patterns form a mode that is predictable for more than a decade. The Southern Ocean is another region with possible decadal predictability (Boer 2000), but quantitative estimates are even scarcer. Though previous studies like these have provided valuable information about decadal predictability, they are inadequate in several respects. Our investigation is designed to ameliorate some of these shortcomings.

Thus far, most predictability studies have been carried out under equilibrium conditions in which the boundary conditions that control climate are held fixed. Hence, only a few studies (Collins and Allen 2002; Boer 2009, 2010) have provided some indications of the second limit of initial-value predictability that interests us, namely the forecast range at which the influence of changes in forcing becomes larger than the influence of the initial conditions. When considering the global mean temperature, Hawkins and Sutton (2009) suggest that the forced response gives more reliable information than do the initial conditions during the first predicted decade. On the other hand, Troccoli and Palmer (2007) and Latif et al. (2006) suggest that when regional or modal variables, respectively, are analyzed, predictability from the initial conditions may be more important than the forced response, for a decade or longer. These studies have only begun to evaluate the relative importance of ocean initial conditions and increasing greenhouse gases in decadal predictions. Here, we have analyzed several Community Climate Model version 3 (CCSM3) ensemble experiments specifically designed to make it possible to address this issue. We use two different forcing, namely the Special Report on Emissions Scenarios (SRES) A1B and Commitment scenarios (Meehl et al. 2006a). In addition to enabling us to quantify the two limits of initial-value predictability, these experiments allow us to assess whether the changing forcing impacts the duration of the initial condition influence.

Another inadequacy in previous investigations concerns the measure used to quantify predictability. Many studies have investigated predictability by concentrating on the rate at which forecast distributions spread (e.g., Grötzner et al. 1999; Collins and Allen 2002; Pohlmann et al. 2004), while other studies have focused on the pace at which the ensemble mean signal weakens (e.g., Newman 2007; Alexander et al. 2008). Taken together, the conclusion from these two types of investigations is that both factors make significant contributions to initial-value predictability. We take into account both factors (section 3b) by using relative entropy (Kleeman 2002) to measure predictability. Furthermore, relative entropy has the advantage that it can measure predictability for multivariate states. As Teng and Branstator (2010) have pointed out, propagating phenomena contribute to decadal predictability and their predictability is difficult to assess with univariate measures.

A third inadequacy of many predictability investigations stems from decadal predictability limits being sensitive to the variables used to define the system state. SST is usually used as the indicator of the state of the ocean because it clearly has the potential to affect the atmosphere. But one might expect subsurface fields in the mixed layer, which are somewhat shielded from weather noise, to be more predictable and yet might still have the potential to affect the atmosphere on long time scales. Reported numerical experiments (e.g., Griffies and Bryan 1997a) support the contention that subsurface quantities are more predictable than SST. In section 3a our study compares the predictability of SST with the predictability of the layer mean temperature in the upper 300 m and concludes that the latter is the superior field for isolating predictable fluctuations on decadal time scales. For this reason much of our study focuses on subsurface, depth-averaged temperature.

Through its experimental design, use of relative entropy, and selection of state variables, our study contributes to the ongoing research into decadal prediction and specifically to a quantification of the two limits that influence the range of skillful prediction (section 3). But for various reasons there are inherent uncertainties attached to the estimates of predictability that we find. In section 4 and the appendix, we explore these uncertainties. Section 5 summarizes our results and discusses their implications including our finding that, in the model we have analyzed, for the World Ocean the initial-value influence becomes undetectable after about a decade and becomes less important than the forced response at an even shorter range. This is true even though our approach may be biased toward giving optimistic limits of predictability.

2. Methods and measures

b. Initial-value predictability and forced predictability

When quantifying predictability for situations where the system climate is reacting to changing external conditions, it is helpful to think of two distinct time-evolving distributions. The first, P_e(t), is the distribution of predicted states resulting from marching a specific initial distribution of states, P_e(0), forward in time. In Fig. 1a this distribution is depicted schematically by the green region. The second pertinent distribution, P_c(t), represents the time-evolving reaction to the forcing and is independent of any particular initial state. One can estimate it from an ensemble of realizations, each beginning long before t = 0 and each experiencing the same time-dependent external forcing. The red region in Fig. 1a is such an ensemble. As Leith (1978) has explained, for situations where the forcing is not fixed, the appropriate definition of climate corresponds to the statistics of the distribution P_c(t) at any given time rather than the conventional view of climate being defined in terms of the statistics of a single realization during a long time interval. Eventually, assuming the system is transitive, as the influence of the particular initial condition is lost, P_e(t) converges to P_c(t). Measures of “total predictability” deal with comparing P_e(t) with P_c(0), but it is also informative to consider two other comparisons that contribute to predictability. One is a comparison of P_e(t) to P_c(t); it represents “initial-value predictability,” which is the focus of our study. The second is a comparison of P_c(t) to P_c(0); it corresponds to “forced predictability.” Note that depending on the exact measure used, the two components may not add up to the total predictability, but these two components do provide a means of comparing the relative strength of the effects of the initial value and forcing.

For our investigation the CCSM3 ensembles described in the previous subsection give us an approximation to P_e(t). Throughout our study we use annual mean quantities to define such distributions. For example, the red dots in Fig. 1b show an approximation to P_e(t) for the annual-mean depth-averaged upper-300-m ocean temperature (which we denote T0–300) in a small box in the North Atlantic in the A1B experiment. On the other hand, we do not have direct information about P_c(t). Ideally, P_c(t) would be estimated from a large ensemble of twentieth-century climate simulations and their A1B extensions into the twenty-first century. In our case only a single realization of these computer intensive experiments is available. To deal with this problem, we have used the following approach. As explained shortly, in our study we have assumed that distributions are well approximated by Gaussians. Furthermore, we have made two additional assumptions. First, the covariance structure of the system climate does not change as the forcing changes and is equal to the covariance structure in our control experiment. In the schematic this corresponds to only the mean of the red distribution changing with time. Second, the evolution of the climate mean of a given state variable can be well approximated by an analytical function of time whose parameters can be determined from forecast behavior after the effects of the initial conditions weaken. Using this function and the assumption of unchanging covariances leads to an approximation of P_c(t) at all forecast ranges.

The functions that we have assumed are good approximations to the time-evolving climate mean [i.e., the means of P_c(t)] are the linear function,

View Expanded

for the A1B experiment, and the exponential function,

View Expanded

for the Commitment experiment. In these expressions, the model year t varies from 2000 to 2061. We have estimated constant k by least squares fitting a line to A1B ensemble values during 2010–62. Evaluating that line for t = 1999 gives T₁₉₉₉. Next, we have inserted the resulting T₁₉₉₉ into (2) and calculated τ and A by least squares fitting of Commitment values during 2010–61.

Note that measures of initial-value predictability concern departures of the red and blue dots in Fig. 1b from the red and blue lines, respectively, which depict the climate means (T_A1B and T_Commit). We refer to these departures as the initial-value components of a forecast to distinguish them from the “raw” forecast states. When initial-value predictability is lost, the mean of the initial-value components is zero and their distribution about their mean matches the distribution of P_c(t) about its mean. In Fig. 1b early in the forecasts the distribution of the initial-value component is clearly distinguishable from the climatological distribution both because of the narrowness of its spread and the separation of the ensemble mean from the climate mean. The forced predictability can be measured by the departures of T_A1B(t) and T_Commit(t) from T₁₉₉₉ because the spread contribution to the distribution difference is negligible under our first assumption.

While Fig. 1b suggests the coexistence of initial-value predictability and forced predictability for regional variables, the situation is different for global mean temperature (Fig. 1c). The ensemble means are not very different from the climatological means, and the spreads of the dots do not increase dramatically in the first decades. Both suggest that in CCSM3 the initial conditions are less important for predicting global mean temperature, though the presence of year-to-year variability in estimates of observed global mean temperature (Brohan et al. 2006) suggests that internal processes do have some impact on this quantity in nature.

Some of the choices we have made when separating the initial-value and forced components of our ensemble forecasts cannot be strictly justified. For example, to estimate the time-evolving mean forced response, we have assumed that the influence of the initial conditions is relatively small after 2010. As the appendix explains, our results are not completely insensitive to this decision, but it appears to be a reasonable compromise choice. A second choice that is not strictly valid is our decision to assume that covariances are not affected by forcing changes. Meehl et al. (2006b) point out that ENSO variability weakens in CCSM3 in reaction to increased GHG concentrations, and in results not shown here we have found that even larger changes in variability in our A1B experiment occur in the North Atlantic near the end of the integrations. But as we will present in the next section, our major interests, the two limits of initial-value predictability, both occur in the first one to two decades, and even in the North Atlantic variability reactions to GHG changes are so weak at this range as to have no discernable effect on our results.

c. Relative entropy

As described by Kleeman (2002) and Majda et al. (2005), relative entropy is a means of comparing a distribution to a baseline distribution and thus can be employed in predictability studies where one wants to determine whether and by how much a forecast distribution differs from a climatological distribution. In full generality, the relative entropy of the distribution P_x relative to the baseline distribution P_b is

View Expanded

where s is the state and S represents the system state space. In addition to the advantages mentioned in the introduction, relative entropy has a well-defined interpretation. It represents the information, in binary bits, in P_x [say, forecast P_e(t)] over and above the information one would have if one knew no more than a state was a member of P_b [say, climatology P_c(0)]. A common application of relative entropy concerns the average number of bits that are required to represent the current state of a system. If the system has distribution P_x, this number will be larger if one uses a system of representing (in information parlance “coding”) states that was developed under the assumption those states are drawn from the distribution P_b than if one uses a system based on knowledge that the actual distribution is P_x. Relative entropy is the number of extra bits needed when knowledge of P_x (or in our case knowledge of a perfect forecast ensemble) is not taken into account. As an indication of what one bit of information represents, consider a system with a finite number of states, and imagine that they are represented in terms of a binary basis with equal probability of each bit being 0 or 1. If a forecast specifies the actual value of one of these bits, then that forecast has a relative entropy of 1. Note that it corresponds to reducing by a factor of 2 the possible configurations of the system. Similarly, a forecast that specifies m bits has a relative entropy of m and corresponds to reducing the number of possible states by a factor of 2^m. In the more general application we make of relative entropy, it need not take on integer values, but it continues to be a measure of how much more precisely we know the state of the system as a result of having a (perfect) forecast ensemble.

In our study we represent the system state by a vector of finite length, n. Since extremely large samples are required to estimate general multivariate distributions of even modest dimension, we have approximated our distributions by Gaussians. In this case, (3) becomes

View Expanded

where μ_b and μ_x stand for the mean state vectors in the distributions P_b and P_x, respectively, while σ_b² and σ_x² correspond to covariance matrices representing the relationships between elements of state vectors in these same distributions. One feature of this approximation is that it can be decomposed into contributions from the mean, namely the third term in the brackets, and from the covariances, which are the rest of the terms. Commonly, these are referred to as the signal and the dispersion components, respectively (Kleeman 2002). Note that while Kleeman (2002) and Teng and Branstator (2010) chose to express values of relative entropy in base e, here we use base 2. We do this because we are more accustomed to mentally raising 2 to a power than e to a power. This choice leads to the factor of log₂(e) in (4) that does not appear in expressions in those papers.

d. Basins and bases

As we have explained, when considering predictability, it is best to be able to represent propagating phenomena. This implies that a field representation of variables should be used. On the other hand, we wish to be able to distinguish predictability characteristics in different geographical locations. Together these factors suggest it is desirable to measure predictability in different basins of the ocean.

To decide on the boundaries of the basins, we have been guided by local time scales of intrinsic variability in the CCSM3 control experiment. As the four examples at the bottom of Fig. 2 demonstrate, when we have examined the variance spectra of T0–300 at various locations we have found a broad range of patterns of behavior. For example, variance spectra in the North Pacific and tropical Atlantic locations shown in Fig. 2 are essentially red but with different decay rates, the North Atlantic point also has pronounced low-frequency variability but with two distinct peaks, and the tropical Pacific point has a pronounced peak with a period of 2 yr (W. D. Collins et al. 2006). As a means of visualizing the geographical dependence of the spectra, we have summarized each one in terms of a single characteristic time scale. If V( f, x) is the variance per unit frequency for frequency f at location x (the quantity plotted in the bottom panels of Fig. 2), then we assign variability at x the time scale

View Expanded

where k points to each of the frequencies we can resolve in the control experiment. These time scales are simply variance-weighted mean frequencies (expressed as cycles per year) transformed to a period. They are plotted in the top panel of Fig. 2. They make clear the much shorter time-scale behavior in the tropics compared to the extratropics, a contrast that may be enhanced by the underrepresentation of tropical–extratropical connections on decadal time scales that has been found in CCSM3 (Alexander et al. 2006). Reasoning that prominent propagating features are unlikely to cross between regions with very different time scales, we have felt justified in choosing basin perimeters based on continental boundaries and the time-scale separation of tropical and extratropical regions. The eight basins determined in this way that we use to partition the World Ocean are outlined in Fig. 2. In sections 3 and 4 we describe predictability in the North and tropical Pacific and Atlantic basins while in section 5 we summarize our results in terms of global statistics that take into account all eight regions.

When calculating relative entropy, we have chosen to represent fields in terms of EOF bases. These bases are calculated for each basin from the control run variability. One constraint on the appropriate EOF truncation is that for the covariance matrices in (4) to be nonsingular, state vectors can be no longer than the sample size minus 1. Hence, our state vector of principal components cannot be larger than 39. And to guard against including variability that is too weak to be estimated with adequate accuracy, we have chosen to make n even smaller. Indeed, in the following section our calculations of relative entropy employ fields that are represented in terms of the leading 15 EOFs in each region. (For T0–300 these correspond to 73%–89% of the variance in the control run depending on the region.) In section 4 we report on the sensitivity of our results to this choice of truncation.

3. Results

4. Sensitivities and robustness

In arriving at the estimates of predictability and its limits, we have made a number of choices that could affect our results. To determine how robust our findings are to these choices, we have carried out a number of tests.

One source of uncertainty in our results is the composition of the state vector. We have seen very large sensitivity to our results depending on whether we use T0–300 or SST as the state variable. Even if we settle on the layer mean subsurface temperature, there is a question as to what is the appropriate layer to use. We have reasoned that a layer that largely represents the mixed layer is a reasonable compromise choice. But with annual mean mixed layer depths varying between tropical values of less than 100 m and extratropical values of 500 m or more, it is possible that in some regions our results would be strongly affected if another layer thickness were used. When, however, we have employed layers ranging from 0–100 to 0–500 m, we have found little sensitivity in the extratropical basins with saturation times and relative entropy values leading up to saturation being affected only modestly. Figure 9a shows this pattern of behavior for the North Pacific. It is only when we consider a layer extending to 1000 m, well below the mixed layer, that we find substantial increases. By contrast, for tropical regions, with their much shallower mixed layers, our choice of the 0–300-m layer has had a substantial effect. As depicted in Fig. 9b, in the tropical Pacific the saturation time is reduced from the T0–300 value of 7 yr to a value of 3 yr when T0–100 is employed. (Further reduction to a 50-m layer has little effect; not shown.) This suggests that our tropical results may overstate the predictability of that portion of the ocean that readily communicates with the surface, and that the contrast between extratropical and tropical predictability mentioned in section 3b is probably even greater than indicated by our T0–300 results.

Another aspect of the state vector that can affect predictability results is its truncation. One might speculate that a more severe truncation than the one we have employed could enhance predictability given that the leading EOFs of geophysical fields often have longer intrinsic time scales than trailing EOFs. To check this possibility, we have calculated relative entropy for the A1B experiment truncated to five EOFs. The North Atlantic region (Fig. 10a) shows a pattern of behavior typical of most basins. Of course, given the form of (4), it is not surprising that all relative entropy values are smaller than for the case of 15 EOFs, which is also included in the plot. We take this into account in a crude fashion by stretching the scale for the five-EOF truncation by a factor of 3, to make up for the factor of 3 difference in the degrees of freedom that can contribute to information. When this is done, it is apparent that information content, saturation time, and crossover range are similar for the two truncations. The behavior for the tropical Pacific (Fig. 10b) is more difficult to decipher quantitatively given the small values of relative entropy resulting from weak projections onto the leading EOFs in this region. The one dramatic effect of the severe truncation is the weakness of the forced response so that forced predictability practically disappears with this truncation. And though the exact saturation range is murky, it is apparent that restricting the state vector to the leading patterns has not enhanced initial-value predictability.

A second characteristic of our analysis concerns the confidence we can have in our relative entropy estimates when they are based on a single finite ensemble. The shading in Fig. 11 depicts the range of values of relative entropy that result from 10 000 random draws of 10- (top panel) and 20- (bottom panel) member subensembles from our 40-member A1B experiment for the North Pacific. (Here, we have employed a truncation of 5 EOFs rather than our standard 15 so that we can calculate the relative entropy for small ensembles.) In Fig. 11 the range is given by the 5th and 95th percentile boundaries while the solid line shows the mean at each forecast range. We see that for 10-member ensembles there is a broad range of possible relative entropy values because of sampling fluctuations. For long-range forecasts the span of possible bits of information is around eight while for the first few forecast years when the ensemble is tightly bunched the span is much smaller. This uncertainty produces a range of saturation years that is about 9 yr wide. By contrast, for 20-member ensembles, the relative entropy is confined to a range of about four-and-a-half bits for extended forecasts and the saturation year is confined to about five possible years. Of course for our full 40-member ensemble, uncertainty will be reduced still further, one of the benefits of using a large ensemble. Another worthwhile consequence of using a large ensemble for predictions is that the threshold for significance is reduced as a result of the finite sample size error in R being reduced. This will tend to expand the range that forecast distributions can be distinguished from the control distribution. It also means that the emergence of the forced signal from the intrinsic noise can be detected earlier.

The third factor is the sensitivity of initial-value predictability to the structure and amplitude of the ocean initial conditions. As mentioned in section 2a, here we address this factor in an expedient fashion; we simply repeat our analysis for three additional 40-member ensemble experiments with initial states different from the initial state in the A1B and Commitment experiments and determine how much our plots of relative entropy as a function of forecast range change. These experiments, which include GHG and aerosol concentrations from the SRES A1B scenario, are referred to as experiments A1B(II), A1B(III), and A1B(IV). A1B(II) and A1B(III) were originally generated for Teng and Branstator’s (2010) study of North Pacific modal predictability and were chosen to differ from each other and from our standard experiments in the way they project onto the leading intrinsic mode in that region. Specifically, A1B(II) is an ensemble of perturbed integrations starting from the 1 January state in 2008 of one member of our standard A1B experiment that has a very strong positive projection onto the Pacific decadal oscillation. A1B(III) starts from a different member of our standard A1B experiment in 1 January 2008, but here the member has a strong projection onto the second leading pattern of North Pacific variability. Similarly, A1B(IV) starts from a third member of our standard A1B experiment but for 1 January 2010, when this member has a strong projection onto the second EOF of the Atlantic meridional overturning circulation.

When we examine plots of relative entropy for these additional cases (Fig. 12), we see that in a few instances there are cases that stand out. The most prominent example is the strong North Pacific predictability in the A1B(II) case. In a more thorough analysis (Teng and Branstator 2010), we have found that this high predictability results from the very large initial projection onto the leading basin EOF of the intrinsic T0–300 variability. Indeed, the initial projection is in the 99th percentile and produces a signal component in the forecast that lasts for at least a decade. A second example of an outlier is case A1B in the tropical Atlantic with its unusually late saturation. This is noteworthy because the other three cases are more consistent with the idea referred to in section 3b that tropical regions have lower initial-value predictability than extratropical regions. With these few exceptions, we find a surprising lack of sensitivity of the saturation range to the initial conditions. This insensitivity contrasts with previous studies that examined the Atlantic overturning circulation (M. Collins et al. 2006) and found that the initial conditions with a strong overturning circulation had enhanced predictability. The discrepancy may be caused by model differences, insufficient sample size in the previous work, the different state vectors used to represent the system, or simply the small number of initial states considered in both studies.

5. Summary and discussion

Motivated by the current interest in decadal climate predictions and the hypothesized improvement in these predictions that might result from ocean initialization, we have quantified the forecast range that initial states can potentially influence in forecasts made with a coupled climate model similar to models that will be used in decadal prediction studies. We have done this by analyzing the evolution of ensembles of initially similar states. To summarize our main ideas and results, we plot in Fig. 13 relative entropy values that quantify predictability properties of the two climate change experiments that were the focus of our study. The values in Fig. 13 are the sum of relative entropy³ in eight basins (Fig. 2) that span the World Ocean. The relative entropy for raw forecasts (black curves) has the characteristic U shape that we also saw for individual basins (Fig. 5). When we separate this into its initial-value (blue) and forced (green) influences, we find the decrease in relative entropy at the beginning of the forecasts corresponds to the loss of information from the initial state one expects in a chaotic system, while the increase after 6–7 yr results from the influence of external forcing. These two processes prompted us to use two time scales to quantify the limit of influence from the ocean initial states. The first is the time at which information from the initial value becomes undetectable. We referred to this as the saturation range, which for the global ocean is about 12 yr in our experiments. The second time scale is the range at which the initial condition information becomes smaller than the information that results from external forcing. We called this the crossover range; it occurs at year 7. By using relative entropy, we can also quantify the information provided by the initial value at any point in the forecast. For example, the initial state provides at least 10 bits of information for almost a decade.

While most previous predictability studies have focused only on the saturation range, because of the design of the experiments we analyzed, we were able to recognize the importance of the crossover range when considering forecasts within a global warming context. Although, like several other studies, we found the ocean initial conditions may provide predictability for a decade or so, the crossover range suggests that for some SRES forcing scenarios after 7 yr the external forcing can bring more information than the ocean initial state. Consequently, in the model we have analyzed, 10–30-yr predictions fall into the category of “boundary condition problems” rather than “initial-value problems” if predictive skill comes solely from the upper-ocean heat content. Therefore, predictions in this range must rely on accurate estimates of future external forcing, rather than on estimates of the present ocean state.

Another significant finding from our study results from considering both the mean and spread when we measured predictability. As discussed in section 3b, for most basins more information is contained in the mean than in the spread of predicted distributions after the first 1–2 yr. Indeed, based on results like those in Fig. 6, for the global ocean between year 2 and the year of saturation, 70% of the information is in the ensemble mean and only 30% in the spread. This result suggests that the many investigations that have focused on spread in their assessment of decadal predictability have neglected a major contributor.

Another potential advance in our study is the variable used to represent the state of the climate system. We found that the subsurface temperature is more predictable than SST. One way to quantify this difference is to note that if we redraw Fig. 13’s summary of predictability for the global ocean but use SST rather than T0–300, we find saturation occurs about 3 yr earlier and crossover happens 1 yr earlier. In section 4 we reported even greater initial-value predictability when we used layers that were deep enough to extend below the mixed layer. In preliminary research we have found evidence that on decadal time scales predictability in the mixed layer is associated with the predictability of surface conditions in CCSM3. Whether predictability produced below the mixed layer has this same property is open to question.

Any predictability study that uses ensemble experiments as its basic methodology suffers because it must draw conclusions using just a small number of cases. Our investigation has this shortcoming, but to the extent we have been able to test the robustness of our results, we have found them to be insensitive to four key factors. First, as Fig. 13 shows, we found both limits of initial-value predictability to be insensitive to the two scenarios used to drive the model. This insensitivity was aided by the similarity in the forced response during the first 10 yr of prediction (Fig. 8) and by the fact that initial-value predictability did not last much longer than a decade. Second, predictability properties were rather insensitive to the four particular initial ocean states used for our ensemble experiments. The only situation for which we found a very pronounced departure in the saturation range was in a case that had an extremely strong initial anomaly in the North Pacific. Third, our results suggest predictability is not dramatically higher for the most prominent intrinsic patterns of a basin than for patterns that explain somewhat less variance. This insensitivity is consistent with Teng and Branstator’s (2010) finding that the leading intrinsic propagating mode in CCSM3’s North Pacific does not have unusually high predictability. But given the high amplitude of the leading modes, this result does not rule out their important role in predictions. Fourth, in results not described earlier in the body of our paper, applying temporal filtering to T0–300 (e.g., 5- or 10-yr running means) does not alter our major conclusions. Indeed, when running n-yr averages are employed, the saturation ranges tend to be extended by no more than about n/2 yr. This increase in the saturation range is simply what is to be expected from predictable years being including in running averages for n/2 yr after saturation occurs in the raw fields. Time averaging may produce more dramatic increases in investigations where smaller ensemble averages are used than we have employed because then it serves the purpose of reducing the unpredictable components that one would prefer to remove through ensemble averaging. Similarly, our use of depth-averaged quantities means that time averaging has less of an effect on our results.

By contrast, we did discover two factors that do influence our estimates of decadal initial-value predictability. First, there are substantial variations in the predictability properties of different basins. The contrast between higher predictability in the extratropics and lower predictability in the tropics was noteworthy in most basins and experiments. This could be associated with large basin-to-basin variations in intrinsic time scales (Fig. 2). We also noticed substantial variations in predictability on subbasin scales. Second, predictability limits, when they are defined in terms of statistical comparisons of finite ensembles to a background distribution, are affected by ensemble size. This is particularly true for the saturation and emergence times and happens because the smaller the ensemble, the larger will be the relative entropy from randomly chosen states (Fig. 11).

As we emphasized in the introduction and in section 2a, when considering our results, one should remember that predictability is an inherent property of a dynamical system and thus our results are valid only for CCSM3 and not for other models or for nature. However, in conjunction with predictability estimates for other models, our estimates help to establish what predictability limits and behavior it is possible for nature to have. Also, no matter what nature’s predictability is, the inherent limits of any model will affect the limits of its skill. Therefore, these limits should be kept in mind when designing and interpreting the results of decadal prediction experiments like the upcoming Coupled Model Intercomparison Project Phase 5 (Taylor et al. 2009). In fact, the prudent approach would appear to be for groups engaged in decadal prediction to carry out similar determinations of their model’s predictability before undertaking extensive prediction experiments. Also, when applying our results, it is best to remember that we have used an unusually large ensemble; we have concentrated on depth-averaged, subsurface conditions; and we have employed a perfect model assumption. All of these factors mean that the two limits of basin-scale CCSM3 predictability that we estimated are likely to be longer than its range of skillful prediction. On the other hand, our plots of subbasin-scale features (Fig. 7) suggest that on these scales there may be isolated features whose initial-value predictability is longer than the basin and global limits we have emphasized. Moreover, CCSM3’s North Atlantic oceanic variability in our long control integration may be influenced more by atmospheric noise than in nature (Danabasoglu 2008) and it peaks at higher frequencies in the decadal range than does the corresponding variability in the short record from nature. Both factors mean the predictability we have found for CCSM3 may be less than nature’s predictability in this region, predictability that future models may achieve as they become more realistic.