To understand how the effective difference in temporal (or spatial) scales between the components is determined by S and τ, we compute the energy Ea/Eo and period Ta/To ratio with varying values of S and τ. The sub-index “a” denotes the atmosphere and “o” the ocean components.

We use energy E to estimate each component's spatial scale. The energy E of the two components (Eo for X, Y, Z and Ea for x, y, z) of the coupled system are computed as follows: 2E=∑n=1Nxn2+∑n=1Nyn2+∑n=1Nzn2 where N is integration length, and xn, yn, and zn are the component's variables at time n.

We estimate the component's period T, i.e. the dominant time scale, as the period at which the power spectrum density (PSD) reaches its maximum. To estimate the PSD, we use the component's magnitude m: 3m=∥x∥=|x|2+|y|2+|z|21/2.

The sensitivity of energy Ea/Eo and period Ta/To ratios to S and τ is explored in Fig. . The relative energy content of each component (Fig. a) shows that the energy of the ocean (Eo), and hence the spatial scale separation, is mostly inversely proportional to S and that the temporal scale has only a little influence on it. Therefore, the ocean's spatial scale increases as S>1. As it could have been anticipated, the temporal separation is uniquely sensitive to τ. Figure  demonstrates that one can change these two parameters separately to modulate the spatial or temporal scale differences accordingly.

We analyze the dependence of information transfer between components on the spatio-temporal parameters by computing the cross-component instantaneous correlation of the coupled system C over 500 TU for all (S, τ) configurations and calculate the spectral norm of it (Fig. ). We consider the sub-matrix Ccc=cij, where i=1, 2, 3 and j=4, 5, 6. The spectral norm of a matrix A is computed as follows: 4∥A∥2=ρA*A1/2, where ρ(A*A)=max⁡1≤i≤n|λi| is the spectral radius of A, the maximum modulus of the n   eigenvalues λi of A; and A* is the conjugate of A. Since the correlation matrix C∈Rn×n then C*=CT.

The pattern in Fig.  reveals that the information flow across the system's components is heavily influenced by S and τ. The cross-component correlation decreases as the spatial scale difference increases. However, and somehow unexpected, the cross-component correlation increases as the time scale difference increases. The cross-covariance shows a maximum when the ocean component (X, Y, Z) has a smaller spatial scale and slower time scale than the atmospheric component (x, y, z), implying a high information flow. Conversely, a large and fast ocean hampers the flow of information. These findings align with those found by in relation to time-scale difference only.

Here, we assess how the spatio-temporal parameters influence the chaoticity of the coupled system. We quantify the sensitivity of the system to initial conditions by exploring its Lyapunov spectrum. We use the Benettin algorithm , as described by , to compute the Lyapunov exponents (LEs) and the first backward Lyapunov vector (LV1). The computation of LEs and LV1 is performed over an integration of 1×104 TU, thus ensuring convergence of the method. To understand the effect of the coupling on the dynamics of the coupled system, we first compare the coupled L63 system (Eq. ) to two uncoupled L63 systems Eq. . 5x˙=τσ(y-x)y˙=τ(rx-y-Sxz)z˙=τ(Sxy-bz).

For the uncoupled atmosphere, we set (S=10., τ=1.0), for the ocean (S=1.0, τ=0.1), and compare them with the coupled system (Eq. ) with the parameterization in Table . Figure  shows the Lyapunov spectrum of the uncoupled sub-components (atmosphere in red; ocean in green) and the coupled system (blue). Both uncoupled systems possess one positive (λi+), one negative (λi-), and one neutral (λi0) LE. However, while the LEs of the atmosphere are substantially different, in the ocean, they are much closer to each other. The ocean appears only very marginally unstable, with the largest LE just above zero. Thus, the uncoupled-unforced ocean is nearly stable. This is also confirmed by the Kolmogorov-Sinai entropy and the divergence (KS-E and ∇⋅f, respectively, in Table ), which are around one order of magnitude smaller than for the atmosphere. Notably, although the ocean is much stabler than the atmosphere, their attractors' dimensions (measured by the Kaplan–York dimension KY-dim, in Table ) are the same.

The Lyapunov spectrum of the coupled system is a mix of the uncoupled atmosphere and ocean spectra (Fig. ). The coupling does not affect the number of positive Lyapunov exponents, which remains one. Instead, it introduces two near-neutral exponents while retaining the two negative exponents from each uncoupled system. The presence of these additional exponents with values close to zero is commonly referred to as quasi-degeneracy. The emergence of these quasi-neutral modes has been found to be related to the coupling itself in models of higher complexity, such as MAOOAM . Moreover, their presence has tremendous implications when designing efficient DA methods to control error growth .

We now investigate in Fig.  how the chaotic behavior of the coupled system changes with (τ, S) parameters. From the system's Jacobian (Eq. ), we can already anticipate that τ will play a larger role in modulating the system's degree of chaos than S. However, since S appears in the cross-component terms (from the ocean to the atmosphere, Eq. ), it can potentially influence the dynamical properties of the system, regardless of the coupling parameter c (Sect. ).

Changing either τ or S does not alter the shape of the Lyapunov spectrum – i.e., the number of positive, near-neutral, neutral, and negative LEs (not shown). However, changing τ impacts the degree of chaos of the system and the attractor's dimension (KS-E, KY-dim and ∇⋅f in Fig. ), independently of the value of S. As the time scale separation decreases (τ→1), the system becomes more unstable, and the dimension of the attractor decreases. However, when τ gets too large, the KY-dim saturates at 4. Since the KY-dim approaches the phase space dimension (n=6), the entropy decreases (near zero values), and the divergence increases; we have that the dynamics become similar to a Hamiltonian system (it approximates a conservative system) as the ocean becomes slower and that the error will evolve along the complete phase space.

We finally analyze how the projection of LV1 on the state vector varies with the spatio-temporal parameters, which helps to understand or discriminate the source of instabilities in the state vector. To this end, we show the ratio between the projection on the atmosphere (LV1_a) and that of the ocean (LV1_o) component in Fig. . The LV1 projection on the atmospheric variables is generally larger, implying that the dominant source of error propagation relies on the atmosphere. This behaviour is independent of the spatial scale of the ocean (S) as long as the parameter τ is less than one (ocean slower than the atmosphere). When both components have the same time scale (τ=1), the LV1 projection is larger in the component with the largest spatial scale, i.e., the projection on the ocean component decreases when S>1. When τ>0.25, the error propagation is relatively even on both components. However, when τ<0.25, the error propagation towards the ocean is almost two orders of magnitude smaller than towards the atmosphere.

In this section, we use the (linearised) dynamical arguments from to investigate the error propagation from one component to the other, and how it depends on the spatio-temporal scale separation parameters S and τ. In this context the evolution of a perturbation ξ(t) in the system is governed by the tangent linear system: 6ξ˙(t)=Jξ(t) where J is the Jacobian of such system. For our coupled system, in Eq. (), the Jacobian reads: 7J=-σσ0-cS00r-z-1-x0cS0yx-b000-c00-τστσ00c0τ(r-SZ)-τ-τSX000τSYτSX-τb.

Let write our coupled system in Eq. () in the following general form: 8x˙=f(x,X),X˙=τg(x,X), where x and X indicate the atmosphere and ocean state respectively, and τ is the time scale. Therefore, over the interval dt, the forecast error in the atmosphere ζ(t) and in the ocean η(t) evolve according to: 9ζ˙(t)η˙(t)=FaFoτGaτGoζ(t)η(t). After re-arranging the terms in Eq. (), we can identify the form of the cross-component error propagation terms, Fo and Ga as: 10Fo=∂f∂X=-cS000cS0000 and 11Ga=∂g∂x=-c/τ000c/τ0000. In these expressions, Fo represents the ocean → atmosphere, while Ga the atmosphere → ocean propagation of error. We can see that Fo depends explicitly on the spatial parameter S. Thus, the error propagation from the slow (ocean) to the fast (atmosphere) component depends exclusively on S and increases as S does. On the other hand, Ga depends only on the temporal scale τ, and it is inversely proportional to it. Therefore, the error propagation from the fast to slow components increases as the parameter τ decreases. Using this information, we can plot the competing direction of error propagation, which we can take as the ratio of the norm of Ga and Fo. For this, we use the Frobenius norm, defined for a matrix A as ∥A∥=[tr(A*A)]1/2. Thus, the competing direction of error porpagation is ∥Ga∥∥Fo∥-1=(Sτ)-1; we illustrate this dependence in Fig. .

Figure  elucidates the role of S and τ on the error propagation across components. The figure has three separate regions, in which ∥Ga∥>∥Fo∥, ∥Ga∥≈∥Fo∥, and ∥Ga∥<∥Fo∥. Thus, when the temporal scale separation vanishes and the ocean's spatial scale increases (τ-1>S), the error propagation is dominated by the atmosphere → ocean term Ga (blue region in Fig. ). In the opposite case, when the temporal scale separation increases and the ocean's spatial scale decreases (τ-1<S), the error mostly propagates from the ocean → atmosphere term Fo (red region in Fig. ).

We conduct a set of coupled and uncoupled DA experiments using the coupled L63 in Eq. (), and the uncoupled L63-like system in Eq. (), respectively, for different values of the parameters S and τ to reflect spatio-temporal separations between the two components of the system. We used the stochastic Ensemble Kalman Filter EnKF and compared weakly and strongly coupled data assimilation (WCDA and SCDA, respectively). The additional experiment using UCDA is contrasted against WCDA.

We use an idealized perfect twin experiment framework , i.e., the model used for generating synthetic observations is the same as that used for DA. The synthetic observations are generated by adding zero-mean Gaussian noise to a reference simulation (hereafter referred to as True). Here, we only observed (y, Y) variables. This choice follows what was done by, e.g., and that showed these two variables to be more informative in the Lorenz model . Similar performances were found when observing the full system (not shown).

The observation error standard deviation σ is equal to 2.5 % of the system's natural variability (i.e., the time-wise standard deviation of each model variable). This implies that the observation error covariance matrix R depends on the spatio-temporal scales of each component (see Sect. ). The observational error is uncorrelated; therefore, R is diagonal with the observational error variance along the diagonal.

The DA is performed using an observational interval equal to one-fifth of the error-doubling time of an uncoupled L63 system (see Sect.  for further details for this choice). To prevent filter divergence, we used an adaptive inflation scheme (Sect. ) and 20 ensemble members. We run the model for 850 TU, allowing for 150 TU for spin-up before the start of the DA. We use such a long spin-up time to allow all the system's (S, τ) combinations to evolve beyond the transient period (especially for the cases with a very slow ocean). The error statistics (Sect. ) are computed over the last 600 TU in a similar way to the experiments of and . We also repeat the DA experiments 30 times with different initial conditions x0T and different observation perturbations to assess the system's performance robustly.

The Ensemble Kalman Filter EnKF, is a Monte Carlo-like sequential DA methodology consisting of a forecast step alternated with an update phase (analysis). During the first phase, the ensemble of states (the ensemble) is integrated forward in time (forecast) from the previous ensemble of analysis states. During the second phase, observations are used to update (analyze) the ensemble for the next iteration. The method uses ensemble covariance to provide flow-dependent correction and performs a linear analysis update.

We denote the ensemble forecast Xf∈Rn×N. The superscript “f” stands for forecast, N is the ensemble size, and n is the dimension of the state. The model error is assumed to follow a Gaussian distribution with zero mean. The ensemble mean is denoted x‾f and the ensemble anomalies are Af=Xf-x‾f1T, where 1∈RN×1 has all its values equal to 1. Under the aforementioned hypothesis, the ensemble covariance Pf is an approximation of the true forecast error, ϵ, covariance matrix: 12ϵϵT‾≈Pf=(N-1)-1AfAfT.

Furthermore, given y∈Rm×1 to be the observation vector with m observations from which we can construct additional N perturbed observations yi=y+ϵoi, and the ensemble of observations Yo∈Rm×N as: 13Yo=y1,y2,…,yN with an associated ensemble of observation perturbations Γ∈Rm×N: 14Γ=ϵo1,ϵo2,…,ϵoN from which we can construct the error covariance R: 15R=(N-1)-1ΓΓT.

The analysis equation then becomes: 16Xa=Xf+KYo-HXf; where H is the (linear) observation operator which relates the forecast model state variables to the measurements. Finally, K is the Kalman gain: 17K=PfHTHPfHT+R-1.

In ensemble methods, such as the EnKF, the analysis step is based on a flow-dependent forecast error covariance Pf, which is a finite-size ensemble approximation of the true forecast error covariance . This sampling error causes the matrix Pf to be usually an underestimation of the actual covariance. Besides sampling error, other sources of incorrect specifications include model error and nonlinearities. All these factors may lead to filter divergence, whereby the filter incorrectly trusts a “too small” Pf. Filter divergence is usually handled using covariance inflation .

In this study, we use the adaptive covariance inflation method proposed by and based on the following relation: 18〈ddT〉=HBHT+R, where d=y-Hxf are the innovation vectors, R is the observation error covariance, HBHT is the true forecast covariance matrix in observation space, and 〈⋅〉 denotes the expected value. This relation holds if B and R are correctly known . For an EnKF, we can approximate Eq. () using the ensemble-based background covariance matrix Pf and the inflation factor α, according to: 19〈ddT〉=αHPfHT+R. The inflation factor α is estimated by taking the trace Tr(⋅) of Eq. (): 20α=〈ddT〉-Tr(R)TrHPfHT.

Since our experiment uses a small sample of observations (i.e., two when the network includes both the ocean and atmosphere and only one when one component is observed), the estimation of α can be plagued by noise, which is detrimental to the correct functioning of the filter . To mitigate the effect of the noise, we used a simple time smoothing function for α′ formulated as follows: 21αt′=(1-γ)αt+γαt-1. Thus, αt′ is a linear combination of the previous inflation factor αt-1 and the one at current time αt; with γ being a weighting parameter 0<γ≤1, which we tune empirically for the time scale separation τ of our system.

We set the length of the ocean (resp. atmosphere) assimilation cycle ΔTo (resp. ΔTa) as one-fifth of the error doubling time: 22ΔT=15ln⁡(2)λ1 with λ1 being the maximum Lyapunov exponent of the uncoupled L63 system in Eq. (). This choice reflects observational networks designed or optimized to counteract effectively the error growth . Furthermore, and importantly for our goals here, this choice allows us to compare different DA experiments, with varying S and τ, but keeping the observational constraint at a comparable strength.

As shown in Sect. , the characteristic time scale of the system is largely modulated by τ and, to a lesser extent, by S but not by the coupling terms. We therefore decide to determine ΔTo(a) as a function of τ only; results are shown in Fig. . This relation is exponential, which is expected from Eq. () in which τ is a factor multiplying all variables of the L63 system. Thus, we use a constant ΔTa=15dt for the atmospheric component, as τ=1 in all experiments (blue dot in Fig. ). On the other hand, the assimilation cycle of the ocean ΔTo will change with the ocean temporal scale (red line in Fig. ).

We assessed the accuracy of each DA approach by computing the root-mean-squared error RMSE of the ensemble mean analysis state, averaged over the 30 different experiments, as follows: 23RMSE=1Ne∑j=1Ne1Nt-1∑i=1Ntx‾ija-xit21/2, where x‾ia is ensemble mean of analysis state, xit is the truth at analysis step i, and Nt is the integration length. The j indices denote the experiment number performed with a different random seed (here Ne=30, see Sect. ). In particular, we adopt the ΔRMSEn metric, which uses the WCDA experiment as a benchmark and normalizes the error by that of the FREE run (experiment with no assimilation), calculated as follows: 24ΔRMSEnSC=RMSESCDA-RMSEWCDARMSEFREE.

Values of ΔRMSEnSC<0 indicate a reduction of error compared to WCDA, values close to zero indicate no difference. Values ΔRMSEnSC>0 indicate a degradation. We present the results as the mean of ΔRMSEnSC over each component. Hence, we present the average error reduction ΔRMSE‾nSC in the full atmosphere and ocean.

For the comparison between UCDA and WCDA, we use the metric ΔRMSEnUC, defined similarly to Eq. (), but now evaluating the UCDA experiment, thus: 25ΔRMSEnUC=RMSEUCDA-RMSEWCDARMSEFREE.

In this comparison, the RMSE for UCDA is computed using the truth and FREE run calculated using the coupled L63 system in Eq. (). The metric ΔRMSEnUC assesses the capability of UCDA to reconstruct the variability of one of the components of a coupled system, using its uncoupled version.

Figure  shows the error of UCDA compared to that of WCDA. In general, UCDA gives larger errors in both components, indicating that using the coupled model for forecasting is useful for propagating information across model compartments and further decreasing the error. The error of UCDA is different on each component, with the ocean presenting the larger difference between UCDA and WCDA.

We can see that in UCDA-A (UCDA in the uncoupled atmosphere, Fig. a), the error has approximately the same magnitude across all the spatio-temporal scale separations. On the other hand, the error in UCDA-O (UCDA in the uncoupled ocean, Fig. b) shows a clear pattern of increasing error toward the small-slow modes of variability. Since the same pattern is observed when comparing UCDA-O with a partially observed WCDA – i.e. when observing atmosphere or ocean only – (not shown), we can conclude that the coupling is key to further decrease the error growth, via the system’s dynamics. This pattern in the ocean becomes evident due to the dynamic characteristics of the system. The area where UCDA-O performs the poorest is a region where the cross-component correlation is largest (Fig. ), and the dominant error propagation is ocean → atmosphere (Fig. ); therefore, the interaction between both components is vital for efficient error constraint, especially in the small-slow modes of ocean variability. This shows that a coupled analysis provides better assimilation.

A comparison of WCDA with SCDA with observations on ocean and atmosphere components is shown in Fig. . It is important to note that overall, the differences between the SCDA and WCDA are very small (about 0.1 % of climatological error); i.e., both systems perform nearly equally well. It was already reported in that the SCDA benefit over WCDA reduces when both components are well observed. While confirming that finding, our results further demonstrate that WCDA performs slightly better than SCDA in most spatio-temporal configurations.

In the atmosphere (Fig. a), SCDA yields slight degradation as the temporal scale separation decreases and most when the scale separation between the two components is large (small S). Both components are highly chaotic when τ is 1 (Fig. ). When S is small, most of the energy is found in the chaotic ocean component (Fig. a). There is little gain during the assimilation as the cross-covariance of the system is relatively small (Fig. ), and the performance is highly sensitive to spurious covariance in the system (sampling error). This result also aligns with the linear error analysis carried out in that suggested that when the temporal scale separation is not large, WCDA is preferable.

In the ocean (Fig. b), the dependence of the error on the temporal separation is the opposite of that seen in the atmosphere. It increases as the time scale separation becomes larger, i.e., as the “stable” ocean becomes more sensitive to the chaotic atmosphere. When the energy of the system is dominated by the ocean (small S), performance is slightly degraded, but when the energy is distributed or prominent in the atmosphere component (S>1), and the cross-covariance is maximum, the SCDA improves over WCDA.

SCDA has little advantage over WCDA when both components have good observation coverage. Furthermore, the system becomes vulnerable to the approximation inherent in the DA (sampling error, linear analysis update), which can lead to slight degradation. We speculate that if we were increasing ensemble size and softening the linear analysis update (e.g., with iterative approaches), this degradation would be gone, and the SCDA would outperform WCDA, but discrepancies would remain marginal.

Note also that when the ocean and atmosphere assimilate data every 20dt, SCDA improves over WCDA for all (S, τ) configurations, especially for the ocean, but the difference is again very small (not shown). This result agrees with the experiments of , which shows that SCDA provides better analyses than WCDA in a fully observed system, with the same assimilation cycle on both components.

When we observe only the atmosphere (Fig. ), SCDA shows improvement over WCDA in nearly all spatio-temporal scale experiments. The pattern of improvements in the atmosphere and ocean components is similar; however, the ocean shows comparatively larger improvement.

The benefit is largest when τ gets small, as the ocean gets less chaotic and well constrained by the atmospheric state. The region for τ<0.25 shows a very strong sensitivity to S. Improvement of SCDA is largest when the ocean has a large scale (S<0.75), and holds most of the system's energy. The well-constrained atmosphere and atmospheric data (even with moderate cross-covariance) can constrain the predictable ocean. In the region where the ocean has a smaller scale (S≥0.75), SCDA has no impact, unlike in the FULL experiment. This result can be explained as the atmosphere → ocean error propagation is smaller (Fig. ); thus, the atmospheric data has no impact over the ocean. Indeed, the lack of ocean data implies that the ocean cannot influence the atmosphere at analysis time. Furthermore, the ocean's energy is too small to impact the coupling during the model integration step.

Thus, we can conclude that using SCDA is highly beneficial over the non-observed component (the ocean). The ocean state improves with fast atmosphere observations, except when the timescale separation is very large and the energy in the non-observed component is small. In this case, the improvement is negligible on both components.

When observing only the ocean, the skill pattern of the SCDA, Fig. , is overall similar to what is seen in the case of only the atmosphere being observed. Nevertheless, we see now that the impact (positive or negative) of SCDA is larger in the atmosphere and almost negligible in the ocean (∼1 % of the climatological error).

Strongly coupled DA degrades over WCDA when the ocean is fast and has a large amplitude (S≥0.25, τ≥0.5). Both components are chaotic, but the ocean holds most of the system's energy. As the cross-covariance is minimal, and the ocean → atmosphere error propagation is small, the usefulness of the cross-update from the ocean towards the chaotic atmosphere is limited and sensitive to the linear update and sampling error. On the contrary, when the ocean's energy decreases (S<0.25, τ≥0.5), the ocean state benefits from the improved atmosphere's initial condition during the model integration. There, the cross-covariance is high with an increased ocean → atmosphere error propagation so that the atmosphere can benefit from ocean observations. These results resemble those obtained in , which uses SCDA in a real framework to update the atmosphere with ocean observations, improving the ocean-atmosphere tropical interface.

It is somewhat surprising that little improvement in the atmosphere is found when it interacts with a slow and small ocean (S≥0.25, τ<0.5). In this region, the error propagation is largely in the atmosphere, and the cross-covariance maximum and the sensitivity to error growth are largely dominated by the ocean → atmosphere. We conjecture that the ocean assimilation cycle is so large that the cross-update cannot properly constrain the atmospheric state, in agreement with .

In a way, the conclusions of the ocean observation network are quite analogous to those found with the atmospheric observation network (being symmetrical w.r.t. the spatial scale). We could anticipate that when the ocean timescale gets as fast as the atmosphere, the benefit will further increase until a certain threshold.

This study investigates how spatio-temporal scale separation in coupled atmosphere-ocean dynamics – i.e., the instability of the system – and the availability of observational data on either or both of the components influence the skill of different approaches to coupled data assimilation. In particular, we analyze the so-called uncoupled, weakly and strongly coupled data assimilation . In uncoupled data assimilation, observations are assimilated using an uncoupled system. In the WCDA, the observations in one of the model components are used to infer the state of that component only during the assimilation step. In the SCDA, all the observations are used to infer the full coupled model, no matter where they are taken.

We focus on ensemble-based data assimilation using the well-established EnKF and use a prototypical low-dimensional coupled system obtained by coupling together two Lorenz-63 models with different parameters . The model configuration allowed us to modify the parameters affecting the spatio-temporal scales explicitly. The model's low computational cost allowed us to analyse its dynamic properties and made it possible to obtain statistically robust results.

The main conclusions are the following:

The coupling between the system’s components is vital for error constraint, and its consideration that is possible in the coupled data assimilation framework provides an effective method for decreasing initial error compared to UCDA. In particular our findings indicate that the ocean is important for atmospheric improvement, as noted by , and the atmosphere-ocean interactions become increasingly important for constraining the ocean's slow variability.

Our study uses a low-order complexity model to explore isolated cases of spatio-temporal scale separations with different observation networks. Despite its simplicity, our study confirms previous studies' findings and provides a complete picture of configurations where: (1) the benefit of CDA over UCDA is demonstrated, and (2) SCDA is expected to yield improvement over WCDA. However, we acknowledge the simplifications of our experiment design and discuss their expected impact on our conclusions: effects of the coupling strength, the superposition of spatio-temporal scales, the choice of data assimilation method, the choice of the observation network and model biases.

Our experiment assumes a weak coupling and studies coupled processes in isolation, while the coupling strength varies from process to process and often influences each other. The influence of the coupling strength for comparing SCDA and WCDA was studied in , where it was shown that a stronger coupling results in a more stable system and higher cross-covariances among the components. As such, we can anticipate that as the coupling gets weaker, the observed benefit of SCDA over WCDA will fade out and strengthen with a stronger coupling. For example, reports this benefit when increasing the cross-component interaction in a system with equal spatial scale separation. We do not think combining several processes is an issue, as standard DA methods (particularly ensemble methods) are designed for that. However, it is expected that a larger ensemble size is required.

All experiments in our study were carried out with the EnKF, which has the advantage of providing flow-dependent error covariance, a property that is important for SCDA . However, the linear analysis update in the EnKF is suboptimal for strong non-linear dynamics , which can also occur with a too-long assimilation cycle. In our study, we encountered these situations when disparate spatial scales and strongly chaotic components led to a degradation of SCDA. We expect that methods that soften the linear analysis update approximation, such as the iterative Ensemble Kalman Filer iEnKF, or outer-loop 4D-Var e.g., CERA-like method would improve the performance of SCDA. The iEnKF has been tested in a low-complexity coupled model with different spatial scales by , showing a good performance. It would be interesting to test whether the iEnKF could remove the degradation of the SCDA over the WCDA with our system. In situations when the iterative approach is unable to mitigate the approximation from the linear update, one could consider using the lagged cross-correlations between the system's components as proposed by LACC method. It should be acknowledged that the LACC method only allows for a way stream of information (from the fast to the slow component) and is challenging to use with the superposition of processes with different spatio-temporal scales.

Another limitation of the EnKF is sampling error, as large ensemble sizes are needed to accurately sample the variance of the system . We did not investigate this issue in our study as 20 members can already be considered a large ensemble size given the size of our dynamical system . With a realistic system, the sampling error is comparatively much larger. Based on our dynamic analysis, we can infer that the limited ensemble size becomes a more restrictive factor for the successful implementation of SCDA as the timescale separation increases (Fig. ). Different approaches have been proposed to address sampling error. Some methods consider the system's dynamic characteristics, such as , which uses the attractor dimension to estimate the rank of the cross-covariance needed for SCDA. Ad-hoc solutions such as vertical localization, as discussed by , or the correlation-cutoff method, as in , and hybrid covariance are also options to address the same issue.

One of the key findings of our study is the confirmation of the CDA’s higher potential over UCDA in reducing the error in both components, thereby legitimising the transition toward WCDA. Our results have implications for NWP, indicating that including the ocean improves the initial state of the atmosphere . In the case of S2D predictions, where the ocean state is the key source of predictability, this transition – UCDA to WCDA – has already been tested . In this study, we further present implications for the initialization of slower modes of variability, showing the importance of atmospheric coupling for such time scales.

Our study highlighted that the observational network is a significant factor in deciding when SCDA outperforms WCDA. Here, we only assess configurations where observations are limited to the atmosphere, the ocean, or both components. The situation is not as distinct in a real framework, but it still shows a strong imbalance between the observation network of the different components. Historically, ocean observations have been scarce compared to the atmospheric network . We can thus anticipate two situations: first, that the largest benefit of SCDA is expected in the ocean component, meaning that the large-slow ocean modes of variability can benefit from the high-frequency atmospheric variability, potentially improving seasonal-to-decadal predictions, as shown in . Secondly, in the context of NWP, we can infer that WCDA will remain the best strategy for initializing the atmospheric state in medium-range weather forecasts over the upcoming years. The atmosphere marginally benefits from the improved ocean – obtained with SCDA of atmospheric observations. This little improvement is more evident with the configurations where a large and fast atmosphere interacts with a small and slow ocean, characteristic of the extratropics, where the impact of SCDA is negligible compared to that of WCDA. However, we acknowledge that there has been recent rapid progress in the ocean observation network. For example, the SWOT altimetry may allow for constraining ocean fronts at a much finer scale than is currently possible, providing another exciting perspective on SCDA, having the capability to enhance NWP based on the fine-fast ocean observations.

Finally, given our perfect-model assumption, we have not addressed one critical aspect: how model bias can hinder SCDA's potential. Earth System Models have large biases , and coupled processes are often only partially represented. Therefore, as models increase their resolution and the availability of ocean observations changes and better resolve those coupled processes , ocean observations can provide more useful information about the atmosphere's surface processes, making SCDA a powerful tool for initializing the coupled system and making skilful predictions.