## 1. Introduction

The ceaseless development of GCMs aiming to approach the complexity of climates has driven numerical models to become ever more complex themselves. The main difficulty stems from nonlinearity and the strong coupling between the elementary mechanisms represented in the models. Under these conditions, calibration of uncertain parameters, along with model verification and validation, has become a challenging task.

However, results of perturbation analysis bring some hope. In climate warming experiments, evidence has been found of linear behavior of the perturbed fields, suggesting that linear analysis can be applied to time- and scale-filtered variables. On the global scale, feedback analysis introduced in Hansen et al. (1984) and sensitivity factors from Wetherald and Manabe (1988) are based on linearity. Criticisms have nevertheless been voiced concerning the validity of these approaches, and some unfounded conclusions on climate sensitivity can indeed be found in the literature (e.g., Dessler 2010; Lindzen and Choi 2009, and comment articles). The methodological difficulty in separating climate forcing and rapid adjustment from feedback is also addressed in Stocker et al. (2014).^{1}

It is mathematically difficult to justify a linear behavior from high-dimensional, nonlinear systems, because this calls for integro-differential functional techniques. In the first of these two papers (Lahellec and Dufresne 2013, hereafter Part I), we approached the problem from the opposite direction: we analyzed the tangent linear system (TLS) that represents the perturbed climate and proposed the assumption that, on large-enough time and spatial scales, a constant Jacobian matrix system can be used. We proposed some dynamical criteria to estimate the relevance of the TLS. The integration of this system should explain the transition from the unperturbed climate at equilibrium to the perturbed equilibrium state. When this is the case, we may conclude that the linear approach is valid.

The mathematical theory of linear systems provides numerous efficient methods. In particular, linearity allows climate sensitivity and feedbacks to be interpreted in a simple manner. At the scales where the TLS reveals itself, climate response can be decomposed into elementary perturbations, thus opening up the possibility of climate behavior becoming intelligible in terms of interactions between elementary mechanisms.

In Part I, we used the formal TLS to clarify a benchmark set of definitions and methods that are currently used to analyze climate feedbacks. We introduced the distinction between standard exclusive feedback analyses as obtained from an offline radiative computation technique and inclusive feedback as found in the standard suppression methods (e.g., Hall 2004; Hall and Manabe 1999; Hansen et al. 1997). The rigorous links between these methods were established with the severe limitation that the surface had to be at the origin of perturbations. We also stressed that speaking of feedback in loose terms can lead to difficulty in interpreting the results when the corresponding feedback loop is not clearly identified.

The dynamics of the TLS in the Laplace domain can be analyzed in a straightforward manner in terms of its characteristic times, and such an analysis allowed us to demonstrate that the regression method introduced by Gregory et al. (2004) gave correct sensitivity factors, the *λ* parameters, as soon as additional criteria were fulfilled. This can be checked in practice. In particular, we learned that when the correct filtering is applied, the atmosphere appears coupled to the drifting surface temperature; thus opening the way for a formal extension of our strict results of Part I. This second paper describes how we can extend feedback analysis to achieve the desired generality in GCMs.

After recalling the principal results of Part I in the next section, we discuss application in the GCM context in section 3. The formal extension of feedback decomposition necessitates a modification of the definition of feedback gains, as already noted by Lu and Cai (2009) in the CFRAM approach.

The main application concerns the phase 5 of the Coupled Model Intercomparison Project (CMIP5) idealized experiments; we investigate how our formal results can be used to derive unbiased feedback factors and how the conceptual linear model underlying these results could be made explicit. The abrupt 4×CO_{2} experiment is used to illustrate this issue and a comparison is made with the ramp forcing experiment.

So far, the formal TLS has been used to demonstrate the validity of classical methods. However, the limitations to filtered space and time scale using these methods call for more elaborate tools. For this purpose, we introduce a perturbation method based on the Gâteaux differential that enables the formally derived/ feedback characteristics to be obtained numerically in GCMs. A version of this method has been introduced in the Laboratoire de Météorologie Dynamique–Zoom, version 5 (LMDZ5), GCM and we give some of our first results in section 5. Finally, in the concluding section, we discuss the perspectives that our work has opened up and also explain the difficulties encountered in implementing the new method in LMDZ5.

## 2. Climate sensitivity and feedback

### a. Tangent linear system

**is the vector of the state variables. The diagonal matrix**

*σ*_{σ}represents the inertia, and

**classically represents temperature, specific humidity, wind field, etc. At this level of generality, the model can be full 3D or a single-column GCM as well as an mesoscale or global balance model. The tangent linear system of the climate model subjected to some forcing**

*σ*_{σ}is the Jacobian matrix ∂

_{σ}

**g**. This system gives the deviation Δ

**relative to the reference trajectory over time. On the integration time scale,**

*σ*_{σ}and

_{σ}are dependent on the state, and while we qualify this TLS as circulating along the trajectory (CTLS), it is still a linear system because the Jacobian matrix depends on the reference trajectory only and not on the variables in Δ

**. Section 5 will show how this CTLS can be implemented in GCM to analyze the climatic response Δ**

*σ***(**

*σ**t*) to

**(**

*σ**μ*) is the Laplace transform

^{2}of Δ

**(**

*σ**t*). Now, the complexity of dealing with integro-differential systems is avoided and instead simple algebraic treatment can be used. We can reduce the dimension of the system by algebraically eliminating all other variables except temperature. This yields the energy system asThe system obtained (the T system) incorporates all feedbacks from other variables in its reduced matrix

*μ*). In this system,

**represents temperatures.**

*η*Because climate analyses compare feedbacks with the specific Planck response,^{3} we separated matrix *μ*) into *μ*), where *μ* or on time.

### b. Feedback loop

**= Δ**

*η*

*η*^{r}+ Δ

*η*_{0}and

*θ*

_{s}was defined as the global response from which the surface-Planck response Δ

*θ*

_{0}is subtracted.

Let us describe how a feedback loop is implemented in the system. As sketched in Fig. 1 (left), the surface temperature response Δ*θ*_{s} is extracted from the state vector as ^{4} It is easily verified that when Δ*θ*_{s} is set to zero on the rhs of the first line of Eq. (6), it becomes equivalent to Eq. (5), the open-loop system. In Fig. 1, the surface forcing term Δ*θ*_{0} propagates into the troposphere, which responds, and the perturbation provides feedback to surface, increasing Δ*θ*_{s}. The feedback loop is closed when Δ*θ*_{s}(*t*) is added to the forcing term. The loop is resumed in the expression of the gain in Eq. (8); reading the factors from right to left:

*η*^{r}and Δ

*θ*

_{s}are additional (feedback) responses to the Planck response (Δ

*η*_{0}and Δ

*θ*

_{0}). The solutions of Eq. (6) iswhere the feedback gain is given byWe called ρ the effective response function. In the Laplace domain, it is the response to the impulse function Δ

*θ*

_{0}

*δ*(

*t*).

### c. Inclusive feedback components

*μ*) for clarity. In other words,

^{i}represents the whole system except that mechanism

*i*, identified by

*g**(

*μ*) is the (inclusive) gain where one mechanism is associated with the feedback loop

*within*the full system

Because all symbols represent functions of the Laplace variable *μ*, the preceding results are valid not only at equilibrium but also during transition, as soon as the time and spatial scales fit the relevance of the linear model. The analysis of the regression technique given below provides the criteria that have to be fulfilled.

### d. Radiative sensitivity

*λ*

_{i}to be introduced. We recall the notationwhich relates surface equations with the last row and column in matrices. With this notation, the feedback gain at equilibrium readsand the sensitivity factorsNoting

*λ*

_{0}the surface-Planck sensitivity and

*λ*

_{i}in Eq. (15) makes explicitly provides the feedback path from the atmosphere to the surface within a scalar expression: excitation of atmosphere from surface through

These are the main results we obtained showing that radiative sensitivity analysis and feedback analyses are consistent and share the same hypotheses. We finally justified the decomposition of temperature change into components (IFC) with the argument that the corresponding feedback loops can be built. This decomposition was already justified by Lu and Cai (2009) based on the fact that the top-of-the-atmosphere (TOA) budget can be decomposed into terms from different origins, in the same spirit as in Le Treut et al. (1994).

### e. Regression

*λ*

_{T}from the relation we called (L2Obs):where

*f*

_{0}was shown to be the surface forcing term and

*N*—the incoming TOA radiation imbalance—and Δ

*T*

_{s}over long time scales. We demonstrated that this relation was valid when the atmosphere was in quasi equilibrium with the surface. At these time scales, surface temperature drifts slowly and the atmosphere temperature follows:

*T*

_{s}. Even though the forcing applies to the whole atmosphere, the surface forcing is still sufficient to dominate the feedback system. The ensuing extension of our formalism follows.

## 3. Application to analysis with GCMs

### a. The attribution problem

So far, we have not treated in models, the explicit meaning of the forcing *attribution* problem) and under what circumstances their effects can be added (the *separation* problem).

Clearly, the attribution problem is dependent on the mode and the way climate mechanisms are mathematically represented. We have found no better way to trace the feedback mechanisms in GCMs than that proposed by Lu and Cai (2009): vector ** η** in our equations is the thermal flux budget of the GCM layers and is usually computed as a sum of separate parameterization contributions that give radiation budgets, turbulent and convective large-scale and mesoscale flows, their thermodynamic energy impacts, etc.

In such a sum of flux convergences: *φ*_{i} = _{i}Δ** η**, so that

_{i}, as part of the full matrix

The preceding algebraic developments are based on the CTLS and can be numerically obtained using a model that actually computes system Jacobian matrices (e.g., Lahellec et al. 2008; Giering et al. 2003). GCMs in general do not compute these matrices explicitly. We will show that the CTLS can be formally obtained using the Gâteaux differential, with mathematical rational explained in Cacuci (1981) and Cacuci and Hall (1984). We just need a finite counterpart of this approach, which we call the finite Gâteaux differential (FGD) method, to numerically determine the feedback elementary factors from GCMs.

### b. The method: Introducing finite Gâteaux differencing

Consider a model running along its reference trajectory, and in the same model we apply a perturbation *T*_{s}(*t*) + 1K *only when calling* a subprogram that takes a specific feedback mechanism into account. For instance, the LW radiation code is called with this perturbation, the rest of the code still seeing *T*_{s}(*t*). The result is a perturbation affecting the LW budget. The anomalies of this run are the solutions of a system analogous to the CTLS and should give the surface-Planck response Δ*θ*_{0}.

To be more specific, the perturbation amplitude should be large enough for its precision to overcome the numerical noise, but still small enough to preserve the linear relationship between perturbed variables. The linearity criterion is given mathematically in Cacuci (1981) as

(*σ**t*+*dt*) =[ (*σ**t*), Δ_{x}(*t*)] weakly Lipschitz (when close to equilibrium for instance);[ (*σ**t*) +*ϵ*Δ*σ*_{1}+*ϵ*Δ*σ*_{2}] −[ (*σ**t*) +*ϵ*Δ*σ*_{1}] −[ +*σ**ϵ*Δ*σ*_{2}] +[ (*σ**t*)] =*O*(*ϵ*) (linearity).

Originally, ^{5} that compute the incremental evolution of the prognostic variables ** σ**. The first criterion essentially concerns the uniqueness of the solution and its continuity when the initial conditions are moved. The linearity criterion allows anomalies to be determined and summed over different perturbation experiments. Neither of the criteria has any significant chance of being verified in the 3D fields. This is where the magic of averaging comes in: it allows a weak form of both criteria. In other words, if we apply the criteria to global averages of the anomalies, the linearity is known to apply, and can be checked. When the CTLS is integrated, that is, when the integration in time of the state variables Δ

*T*, Δ

*Q*for specific humidity, etc., are added to the standard calculation of the tendencies of

*T*,

*Q*, etc., the analysis of their global averaged trajectory should satisfy the required criteria, as long as climate bifurcation is not generated.

### c. Extending sensitivity analysis in GCMs

In section 2, we used _{i}. Perturbing the surface will immediately impact sensible and latent surface fluxes but, depending on the code, the impact is not obvious on other climatic mechanisms, such as deep convection or clouds. One can exemplify this facet by considering the vertical diffusion model in LMDZ, which represents all small-scale turbulent mechanisms: surface evaporation, surface-layer, boundary layer, and free-troposphere turbulent diffusion models (TRB; see section 5). The corresponding Jacobian matrix _{TRB} is the three-diagonal diffusion matrix with turbulent conduction. The rightmost column of _{TRB}, _{TRB} must be included in the feedback. This is general in GCMs: a mechanism is represented by the full matrix _{i} and we need to reformulate our results to take account of the direct impact of the forcing on the whole atmosphere, not only on the surface. Thanks to our Lyapunov argument concerning the slow climate response, after a transient evolution the troposphere is controlled by the drifting surface temperature. This provides an opportunity to extend our formalism to sensitivity analyses of climatic interest.

_{x}applies to the bulk of the atmosphere. Here, the open-loop matrix is the Planck matrix

^{b}and hence includes the other tropospheric feedbacks—giving what we named the “surface Planck” response. We also need to constrain the tropospheric response to the surface response:

^{6}for the slow Planck-response system).

^{7}In this way, mechanisms that are not immediately sensitive to the surface

_{i}(see appendix A). The new expressions of the extended factors

*λ*

_{T}. This redistribution of sensitivity factors applies to the new response functions

With this extension, only valid for slow time scales, we have finally reached our goal of developing a numerical framework that is not as strict as our formal introduction but gives the possibility of analyzing the climate sensitivity through the mechanisms represented in the GCM codes. All the methods we have already reviewed are valid, from the PRP-equilibrium method to the regression technique. This constitutes the body of methods we shall illustrate in the next sections.

## 4. Application to analysis of CMIP5 experiments

We use the following analysis of the abrupt CMIP5 experiment as a test case to show how a rigorous procedure can ensure that a global analysis leads to relevant climate feedback parameters.

### a. Dynamics

_{2}CMIP5 experiment (cf. Hourdin et al. 2013). Perturbations are computed as the difference between 140 yr of the perturbed trajectory and the average of the Pi-control trajectory with constant preindustrial CO

_{2}concentration. Figure 2 gives the evolution of Δ

*T*

_{s}and

*N*(vertically shifted by +5.35) that are used in the L2Obs relation. It can be observed that one exponential function alone cannot fit the data. This is well known as far from equilibrium, some heat leaks to the deeper ocean—see Wigley and Schlesinger (1985), who introduced the “ocean heat uptake” to account for this feature. It is the reason why the TOA budget is that far from equilibrium after running 140 yr. We therefore need to complete our formal model with a deep ocean layer—note that this does not change the radiation budget part of our analysis. Supplementing our surface model [Eq. (16)] with this deep ocean layer means adding one equation to our model:where we have chosen

*κ*as the opposite of the heat transfer coefficient for compatibility with the sign convention of the ocean heat uptake view (for brevity in this section,

*λ*stands for our global factor

*λ*

_{T}).

*e*-folding times can be found by eliminating each variable in the other equation, which giveswhere

We now have a modified *e*-folding time of *μτ*_{d} ≫ 1. A longer *e*-folding time is introduced as

We see that, as long as *μτ*_{d} ≫ 1, the surface temperature response tends to *μτ*_{d} → 0) *T*_{eq} in the following, meaning the asymptotic equilibrium temperature change is given by the two-layer model.

*T*

_{s}of Eq. (20) are given in appendix B—cf. Eq. (B2). It is found to bewith the coefficientsThis verifies that

*b*+

*c*= 1, and hence

*S*is the time for the ramp to reach the 4×CO

_{2}forcing at 1% yr

^{−1}, and we haveFor the radiation budget

The results of Fig. 2 will be shown to give a first ocean layer of ≃50 m and a deep layer of ≃600 m: the conditions are met for using the regression technique to determine the global sensitivity factor from the slope in Fig. 3. We find *λ*_{T} ≃ −0.85 W m^{−2} and *f*_{0} = 6.8 W m^{−2}.

### b. Regression

*e*-folding times differing only by one order of magnitude, and we can apply the fit with different strategies. Consider the two functions to fitwhere

*N*= −

*λ*

_{T}Δ

*T*

_{s}−

*f*

_{0}, and where we have introduced the two constants

*a*and

*d*to account for the fact that the L2Obs model might not verify Δ

*T*

_{s}(0) = 0 because of “fast mechanisms” occurring before it reaches the Lyapunov horizon. The model could be forced to reach equilibrium by taking 1 − (

*d*+

*e*+

*f*) = 0, which would lead to Eq. (26); equality could be imposed between the

*e*-folding times, etc. Based on our argument concerning equality between the characteristic times in

*N*(

*t*) and Δ

*T*

_{s}(

*t*), we fit

*N*+ Δ

*T*

_{s}(because |

*λ*| ≃ 1) to determine the

*e*-folding times and no other constraint. With these times, we fit the three other coefficients in each case (fit A). This provides the values seen in Table 1, where we also give the results for fit B, which, in addition, constrains the model to reach equilibrium [

*N*(∞) = 0].

Parameters from the two fitting procedures [*κ* is taken from *λ* in fit B is from the regression].

The return to equilibrium in fit B adds 1.5 K to Δ*T*_{∞} of fit A and artificially increases the second *e*-folding time. Note that, in both fits, we find *e*-folding times that are very comparable with those found by Jarvis and Li (2011) with a different method and with other atmosphere–ocean GCMs (AOGCMs).

We now discuss the more objective results of fit A. The value of *κ* is lower than the results of Raper et al. (2002) for nine AOGCMs in CMIP2, which ranged from 0.6 to 0.9 W m^{−2} K^{−1}. Geoffroy et al. (2013) developed an identical model to ours and applied it to 16 models involved in the CMIP5 abrupt and ramp experiments. Their fitting procedure is, however, different; closer to our B fit but without using *N*. It still gives parameter values in the same range as ours: multimodel mean for *f*_{0} = 6.9 ± 0.9 W m^{−2} (their 6.8) and *λ* = −1.13 ± 0.3 W m^{−2} K^{−1} (their −0.85) (see their Table 4). The two *e*-folding times are also given: the fast one is 4 yr and the slow one is 249 yr, where we find 10 and 100 yr—these parameters are very sensitive to the fitting procedure. Notice that our model [L’Institut Pierre-Simon Laplace Coupled Model, version 5, coupled with NEMO, low resolution (IPSL-CM5A-LR) in the paper] does not need the heat-uptake efficiency factor introduced by the same authors.

Now, if the two-ocean-layer TLS model were adequate, the two periods where the first or the second exponential function is overwhelming would give the same sensitivity factor: *e*/*b* = *f*/*c*. This is not the case and, in the second period, *N* is less sensitive to Δ*T*_{s}. By that time, Δ*T*_{s} > 5 K and, because the specific humidity responds exponentially to temperature, saturation of water vapor absorption bands would suffice to explain this nonlinearity, but clouds might also have an influence. Such behavior is also noticeable in Forster and Taylor (2006) for some GCMs (see their Fig. 2). Andrews et al. (2012) give the regression results for the CMIP5 abrupt experiment in their Fig. 1, in terms of global annual averages, which are noisy; still it allows us to notice an analogous behavior for most models. In particular, there is evidence of two periods: (i) below about 10 yr, a linear regression is aligned with the fast-forcing value (red cross on their figure), and (ii) the remaining period with a different slope and an eventual saturation (curvature).

This nonlinear behavior is presently seen in Fig. 3, where the regression between the fitted functions is superimposed in dark green. We see that the corresponding regression deviates from the straight line at the end of the run. Figure 3 also shows that the rightmost green point, which looks like a stopping point, is still far from equilibrium. In these conditions, one cannot extrapolate toward *N* = 0 to determine an equilibrium surface temperature; this justifies our fitting method and we differ on this point with the cited authors.

After 140 yr, the ocean might react with a new structure incompatible with the two-ocean-layer model. This is what Gregory (2000) claims from the analysis of a ramp experiment with the Hadley Centre Coupled Model, version 2 (HadCM2) AOGCM, the coupled model of the Hadley Center: a change in deep water formation and upwelling. Gregory (2000) also uses a two-layer ocean model to show how the heat flux to the deep ocean can become to have little correlation with the surface warming. Li et al. (2013) show evidence of hemisphere and latitude contrasts in SST warming in the long-term abrupt experiment, where surface temperature only reached equilibrium after 1200 yr.

### c. Checking the Lyapunov criteria

Following our methodology, we have to check that the troposphere temperature profile has reached the leading Lyapunov eigenvector *η*_{k}/Δ*T*_{s} for a few levels *k* as in Fig. 4. The asymptotic constancy is attained within a few years for the troposphere temperature but takes more than 5 yr in the stratosphere. Level 11 (400 hPa) in the mid–upper troposphere is the most sensitive level, 50% more than at the surface. This correctness criterion strongly constrains the regression method and, depending on the desired precision, excludes any attempt to link fast features with climate warming characteristics using this method; because fast features are incompatible with the constancy of

### d. CMIP5 ramp forcing experiment

_{2}experiment for which the CO

_{2}concentration is increased by 1% every 1 January. Such an exponential increase induces a nearly linear forcing increase. Compared to the response to the step increase, a time integration allows us to writeAt

*t*=

*S*, the quadrupled value is reached (

*S*= 140 yr =1680 months). The ramp TLS looks consistent with the fitted coefficients from the step experiment: in Fig. 5, where the step result from LMDZ and the fitted function are in red, the same function integrated over time (blue points) is consistent with the 1pctCO

_{2}result (blue line). The same consistency between the step and ramp forcing is also found with the HadCM3 AOGCM by Good et al. (2011). Because the ramp forcing applies a (small) step each 1 January, it is possible to check that it shows the same surface temperature warming patterns as the abrupt forcing, and also the same precipitation and cloud change patterns, beginning with strong perturbations at high latitudes. The amplitudes, however, obviously increase faster in the abrupt case (not shown).

_{2}regression curve in magenta in Fig. 6, we see that it does not follow a straight line. However, Gregory and Mitchell (1997) use this supposed linear regression to determine

*κ*, whereas we proposed Eqs. (24) and (25) as an alternative. If we examine the fraction

*b*+

*c*= 1 in the hypothesis that the model is valid up to equilibrium, we findWe see that the regression could provide an approximation to

*κ*or, in reality, to

*b*,

*τ*

_{1}, and

*τ*

_{2}, thanks to Eq. (25). It is the presence of

*τ*

_{1}in the numerator that limits the slope constancy and it can hardly be neglected for the fit with LMDZ5A. In addition, the criterion

*τ*

_{1}≪

*t*≪

*τ*

_{2}is hardly satisfied for the actual values of the two

*e*-folding times. An approximate fit here gives

*κ*ranging from 0.4 to 0.5, acceptable compared to the correct determination from

*λc*/(

*b*+

*a*) = 0.5 W m

^{−2}K

^{−1}. Another method adopted by Gregory and Mitchell (1997) takes

*κ*=

*N*(

*S*)/Δ

*T*

_{s}(

*S*), giving 0.56; it can be improved with our correction, then giving 0.5 again.

We apply the regression method to the net TOA budget with the ramp forcing added, as in Forster and Taylor (2006), to determine the climate sensitivity factor (Fig. 6, thin blue line). The slope giving *λ*_{T} is found to be 0.82 (0.85 in the first half and 0.71 thereafter), close to the step results. Unlike the step characterization, the ramp determination of *λ*_{T} does not stand alone; it relies on the forcing, which, as we emphasized, has to be the *y* intercept of the step regression straight line. The straight line in red is obtained from the fitted functions of time for the step experiment.

### e. On the fast responses

In the fitting procedures, we met two types of fast responses. When fitting the regression lines, the *y* intercept is currently interpreted as the forcing. In fact, it is a virtual forcing: it is the forcing that has to be applied to the L2Obs model, which, by its very essence, does not represent the fast responses as it does not include Δ*T*_{s} = 0. As a second type of fast features, departure from the straight line is seen along the regression line.

#### 1) Initial transient response

The *N* intercept cannot be the Planck forcing *f*_{0}, because fast responses of, say, clouds and precipitation participate in these, as is the case for LMDZ5A (not shown). Depending on the authors, initial fast features are called atmospheric adjustments or semidirect effects (see, in particular, Colman and McAvaney 2011; Gregory and Webb 2008; Andrews and Forster 2008).

In the fitting of exponential evolution functions, coefficients like *a* in Δ*T*_{s}(*t*) and *d* in *N*(*t*) [cf. Eq. (27)] also come from transient mechanisms, and depend on the smallest *e*-folding times of the TLS. They correspond to residuals of faster responses, when Δ*T*_{s}(0) ≠ 0. The quantity *N*(0) again cannot be interpreted as the Planck forcing: considering the value of *τ*_{1} ≃ 10 yr, all mechanisms faster than, say, 3 yr should be considered fast. These residuals change the equations for the temperatures slightly:

How these fast responses arise from our formal model can be seen in Part I [Eq. (39)]. The first order in *μ* forcing *N*^{L}. Once the transient responses are damped during the 1% forcing, they follow a ramp function in time that becomes a new long-lasting forcing—their asymptotic influence. This is true for the fast mechanisms as well. In Fig. 6, the straight green line gives the influence of these fast mechanisms from the fitted functions. In blue, we have added the contribution of the first ocean layer. We see that the fast mechanisms explain a large part of the global sensitivity; see also Fig. 5 where the magenta line gives the 1pctCO_{2} temperature anomaly without considering the fast mechanisms.

#### 2) Running fast response

Fast phenomena also arise along the path on the regression line. Consider, for instance, the first distinct spiral pattern in Fig. 3 at Δ*T*_{s} = 5 K. It corresponds to a global cooling of about 0.3 K. Such a feature is called “a hiatus period” in Meehl et al. (2011), like the one observed during the decade 2000–09. This hiatus was associated with a TOA radiative imbalance of about 1 W m^{−2}—very comparable to our 3-yr-average *N* swing of about 0.5 W m^{−2}. We could indeed verify that, in the GCM results, it is associated with a strong La Niña event occurring during year 50–51 of the step experiment, the very reason invoked by Meehl et al. to explain the hiatus period. This imbalance nullifies the validity of the L2Obs method at short time scales, as we claimed. The difficulty is also acknowledged by Gregory and Forster (2008, p. 68), who note that “episodic volcanic forcing cannot be described by the same relationship and merits further investigation.”

An undesirable consequence of fast-feedback mechanisms is that forcing cannot be determined by evaluating only the radiative impacts of greenhouse gases and aerosols as in Forster and Gregory (2006). This difficulty unfortunately hinders the use of remote sensing short-range observations to determine the climate sensitivity with the L2Obs model. An evidence of this is found in Good et al. (2011), when they compare the (implicit) use of the L2Obs relation to fit the response to a 70-yr ramp-up followed by another 70-yr ramp-down forcing. Our interpretation is that because they use yearly values in their reconstruction (using a convolution product) fast mechanisms that particularly show up at the transition period cannot explain the HadCM3 GCM run. This is conspicuous for precipitation, which is predominantly a fast mechanism (see their Fig. 2).

To sum up, the consequence of the constraints imposed by the relevance of the L2Obs model is to limit the number of climate sensitivity factors that can be determined from regression. Take, for example, the lapse-rate factor in any of its definitions. It is necessarily a component of fast mechanisms because the tropospheric temperature anomaly has to keep the same profile (the leading eigenvector). More elaborate analyses, using our FGD perturbation method for example, have to be foreseen for a detailed description of the influence of climate mechanisms related to their characteristic times *τ*. However, if atmospheric adjustments are taken altogether, a very adequate method can be used to evaluate the response to forcing while keeping the fixed SST. This is the clever way to determine *N* for Δ*T*_{s} = 0 in the L2Obs relation, as introduced by Hansen et al. (2005).

## 5. An implementation of feedback loops in LMDZ5

The preceding section has shown that a TLS at the global scale can be determined from a difference between two runs. On smaller regions and smaller time scales, this cannot work because climate trajectories are essentially stochastic. In contrast, if the CTLS [Eq. (2)] is integrated in time along a single trajectory, we have access to all spatial and time scales. We gave an example of such an implementation, concerning only radiation, in Part I. A version of the CTLS method has since been coded in the GCM and we report our first results in this section.

We take the analysis of the water-cycle (W cycle) feedback response as our test bed. In LMDZ5A, three models represent this mechanism: surface evaporation and vertical diffusion (TRB) (e.g., Deardorff 1974), deep convection (DC) from Emanuel (1991), and large-scale condensation (LSC; see Hourdin et al. 2006) in addition to the radiative shortwave (SW) and LW exchanges (RAD; cf. Fig. 7. These three mechanisms compose the total W-cycle feedback—evaporation, precipitation, and turbulent vertical exchanges. We recall that this feedback is not the PRP definition of the water vapor feedback (WV), which is only the greenhouse effect because of the added WV content, but here is the response of the mechanisms representing the water cycle in the GCM.

Perturbations are not propagated through cloud models or orographic effects. Sea ice and land ice change are also excluded from the CTLS—as they are prescribed in the forced mode. What about dynamics? Perturbations should be advected by the unperturbed wind field, without dynamical feedback. In other words, (**u** ⋅**∇**)Δ** η** is taken into account, not (Δ

**u**⋅

**∇**)

**.**

*η***, and total water**

*η***q**

_{T}=

**q**+

*q*_{l}(vapor + liquid). The trajectory is found by integrating the systemFour mechanisms are involved in the perturbation: TRB, DC, LSC, and RAD, so

*j*covers TRB, DC, and LSC. TRB also treats the subsurfaces (ocean, land, land ice, and sea ice) with snow cover and runoff; we will not give the details here.

^{8}The CTLS readswhere we use

*d*

_{t}as ∂

_{t}+

**V**⋅

**∇**. Equation (31) is approximated by application of the FGD:The atmospheric liquid water is computed in LSC, so, in addition,and additional equations for the soil model, snow cover, etc., are included in this definition of the W-cycle feedback.

The differences in square brackets appearing in the equations are the mathematical representation of the double call to each of the mechanisms in the model (cf. appendix C). At the limit of small perturbations, Eqs. (32) and (33) should approximate the CTLS. However, we chose the forcing as 2×CO_{2} for convenience, but a smaller forcing can be used as long as round-off errors do not make the response too noisy. This system, which gives the inclusive W-cycle response to 2×CO_{2}, can be schematized as in Fig. 7. The forcing (Frcg) is obtained from the integration of the Planck response as another CTLS—the one detailed in Part I.

Note that the system is structured as an exclusive feedback loop and some mechanisms can eventually be discarded to analyze stability. Another remark concerns the flexibility of building the CTLS: as our run is in the forced mode, SST is periodic from the data. The CTLS we built not only adds a perturbation to it, but also gives a certain inertia to the ocean surface—we chose the 50-m thickness found in the CMIP5 analysis, which applies to the Planck response as well. In this way, the CTLS includes a “slab ocean.”

The two CTLSs (Frcg and W cycle) are integrated in time along the trajectory following the same numerical method as the GCM for time-step integration except; of course, for the ocean surface. The perturbed fields (Δ** η**, converted into potential temperature, and Δ

**q**

_{T}) are given as passive tracers to the dynamical package. If advection were not taken into account, we would obtain for each column, the unstable response of the radiative–convective single-column model. With advection included, the so defined W-cycle response acts as forcing on the dynamic feedback.

### a. Regression and forcing

The first series of results concerns the L2Obs regression curve. To be clear about our numerical experiment, the GCM is run with fixed periodic SSTs and preindustrial CO_{2} concentration. The “physics” driver has additional code to compute in parallel, at each time step, the anomalies responding to the 2×CO_{2} abrupt perturbation. Clouds and sea and land ice cover are not sensitive, but the snow cover is part of all processes in Fig. 7. A second CTLS representing the Planck response is also integrated as part of the process.

Figure 8 shows the resulting regression curves (*N*, Δ*T*_{s}). Considering the monthly values (symbols and green line points), one can see that the warming buildup is not a trivial matter. The monthly values show the same complexity as in Fig. 3 with similar spiral and striation patterns but on a shorter time scale. However, there is no ENSO in our slab-ocean CTLS. Black lines are the 12-month averaged values that are both nearly straight lines. For the W-cycle curve (bottom right), the *λ* factor is found to be ~−1.64 W m^{−2} K^{−1}. To compare with the PRP WV feedback, despite the difference in definition, we can combine the WV and lapse-rate factors with the forcing: the multimodel mean value in Dufresne and Bony (2008) gives an *λ*_{p} + *λ*_{wv+lr} of −2.2 ± 0.6 W m^{−2} K^{−1}.

So even linear modeling of the perturbation brings out the fact that the L2Obs regression is not a straight line. To complement what we said in section 4, the underlying unperturbed model is strongly nonlinear; clouds, etc. vary quickly with deep-convection events during the run, and the CTLS inherits these changing meteorological scenes with a fluctuating atmosphere. The top-left curves give the Planck response regression, with the slope giving *λ*_{P} = −2 W m^{−2} K^{−1}, as already discussed in Part I.

We now examine the method of using the fixed SST perturbation to determine the initial forcing. We run the two CTLS with a quasi-infinite slab thickness, which gives the two red curves in the same figure—monthly values with blue crosses. This clearly shows that the method nearly determines that the integrated fast response is forcing the L2Obs straight line. However, this “initial” forcing of the L2Obs does not correspond to Δ*T*_{s} = 0: this would only be true for an ocean planet. Continental surfaces warm up and the L2Obs forcing is to be taken for the global Δ*T*_{s} obtained after the transient response of all surfaces other than the ocean. This is seen in Fig. 1 of Andrews et al. (2012) as well, where red crosses indicate the forcing found at fixed SST perturbation runs. The fixed SST method does not provide the exact fast-forcing constraint because of the slow-part response of continental surfaces. The fast response in the W-cycle surface temperature is found as 0.16 K, while it is 0.24 K for the Planck response; this gives a negative W-cycle effective response of −0.08 K and consequently, we can conclude that the fast W-cycle feedback is negative.

Figure 9 gives another look at that feature and shows the regression of total precipitation and evaporation with Δ*T*_{s} for the first 10 yr. The transient lasts about 30 months when precipitation decreases more than evaporation, leading to a negative fixed SST yearly steady-state value of −0.04 mm day^{−1} (green symbols). What happens is that the radiative warming of the troposphere makes relative humidity decrease and so precipitation decreases. Specific humidity hence increases and evaporation also decreases in turn. This mechanism explains the negative fast feedback found in Hallegatte et al. (2006) with a simple 1D model. Colman and McAvaney (2011) suggest the same conclusion. It can also be seen that, after that transition, water is conserved in the W-cycle CTLS.

As we claimed, atmospheric lapse rates form part of the fast responses. Figure 10 shows that they need a year to approach steady-state. The Planck response (Fig. 10, top panel) shows the initial strong cooling of the stratosphere and heating of the low troposphere. The curves then become smoother, from right to left, giving an almost vertical profile. This gives a justification of a classical method of using the troposphere +1-K perturbation on the vertical as the Planck response profile (cf. Part I). The W-cycle temperature response shows a more complicated evolution, starting with an S shape (rightmost curve) similar to the initial Planck forcing, and later showing an enhanced warming in the mid- and upper troposphere, which can be attributed to enhanced convection in some regions. The final difference between the Planck profile and the W-cycle response (leftmost curves) is part of the PRP lapse-rate feedback. One can notice the similar profile obtained with a fixed SST perturbation run in Fig. 3 of Colman and McAvaney (2011).

### b. Decadal term

After about 3 yr of abrupt forcing, the CTLS structure follows the seasonal variation. Figure 11 is the Hovmöller diagram of deep-convection precipitation anomaly in the tropical region during 2 yr. Convection is almost always enhanced by the W-cycle feedback (by 10%). This is a classical result (e.g., Dufresne et al. 2013). The Maritime Continent and the western Pacific (140°–170°E) constitute an exception; where deep-convection decreases, large-scale condensation takes over. This region was the most convective one in this run (*ω*_{500} ≃ −50 hPa day^{−1}). The W cycle brings more humidity, at higher temperature (about 2 K on the vertical), but convection does not increase as it should. This is clearly seen in Fig. 15 of Hourdin et al. (2006), where the Emanuel model begins to saturate when *ω*_{500} is lower than about −30 hPa day^{−1}. In that case, large-scale condensation removes the surplus humidity. Resulting total precipitation change is positive almost everywhere over the region and amounts to about 15% (not shown).

The atmospheric response to the Planck forcing (Frcg) and the induced W-cycle feedback response Δ*T*_{w} are shown in Fig. 12. The Planck response to forcing (SW + LW) can be compared with the immediate forcing found in Taylor et al. (2013, Fig. 2c). Even though, the National Center for Atmospheric Research (NCAR) Community Climate System Model, version 4 (CCSM4), is a coupled model, we choose comparison with it because, using the Climate Feedback Response Analysis Method (CFRAM) for diagnostics, the analysis follows the same logic as ours (see appendix B). Temperature anomalies in our results have smaller amplitudes, because our Planck response is relaxed after 12 yr; also zonal patterns are different.

As for the W-cycle response, we have a different definition, since Taylor et al. (2013) use the classical WV-content radiative feedback. Compared to their Fig. 3c, our results show that the equatorial temperature change maximum is shifted up, because of rising deep convection, as shown in Fig. 13. In fact, our results are closer to their Fig. 1c, which gives the global response in temperature that includes all feedbacks (WV, clouds, surface albedo, and dynamics). We obtain, however, the same north–south asymmetry. Our maxima are of the same order but greater (4–5 K compared to 3 K) and the W-cycle Δ*T* pattern is more concentrated in the equatorial region. Comparison with a full coupled GCM is difficult to justify, but knowing the relative importance of the WV feedback, the comparison is still meaningful. Recall that our results in particular do not include dynamic responses and thus represent only a forcing to the large-scale flow. This might be among the reasons of the stronger amplitude change in our results: this forcing would very probably be relaxed by the dynamics. Our results are particularly close to the lower left graph in Fig. 4 of Hansen et al. (1984), giving the global warming in response to 2×CO_{2}.

Figure 13 gives the change in the deep-convection-induced mass flux. In the model of Emanuel (1991), updrafts include the undiluted adiabatic core of the convective column and the mixing-induced updrafts, while downdrafts are composed of the unsaturated and saturated downwelling currents induced by precipitation. The net updrafts (shaded) show that DC influences the tropopause in the equatorial region while the net downdrafts (contours) display a systematic decrease in intensity. Deep convection is shifted upward all along the Hadley cell.

### c. Decomposing the W cycle in mechanisms

*η*_{p}(

*t*). The three feedback systems have the formwhere

*d*

_{t}= (∂

_{t}+

**V**·

**∇**).

So, for TRB-CTLS, for instance, the entries to the subroutine accounting for the mechanism are *T* + Δ*T* + Δ*T*_{frcg}, *q* + Δ*q*, and similar entries account for surface and soil temperature and humidity (cf. Fig. C1). For the same CTLS, the other mechanisms depend on *T* + Δ*T* and *q* + Δ*q* (no Δ*T*_{frcg}). The sum of these components and the forcing should, by linearity, give the full W-cycle CTLS. This can be mathematically checked by adding the four systems in Eq. (34), with *i* indicating TRB, DC, or LSC. This procedure was implemented using three runs.

The main test concerning the CTLS approach was to verify whether the sum of the three IFCs and the forcing gave the global response. We first explain the difficulty we met. During our numerous tests, we found that, below some threshold on the forcing amplitude, the response became independent of that amplitude; there was some pullback to an unexplained residual noise from DC that might be attributable to the triggering algorithm. This feature of the Emanuel model is obviously unpleasant and it can be expected that new triggering functions (using a statistical approach) could improve the model (cf. Rochetin et al. 2014). But the fact is that, meanwhile, we have to deal with this numerical problem. That problem, only detected when using the FGD method, shows the advantage this method may bring for testing parameterizations. We finally got rid of the problem by using a noise subtraction technique, which necessitates calling the DC code three times. The results are shown in Fig. 14 with the profiles of atmospheric temperature and specific humidity averaged over the tropical region—where the discrepancy between the sum of the IFCs and the full response is the most pronounced. Now, apart from a residual discrepancy in the boundary layer,^{9} both Δ*T* and Δ*Q* IFCs are additive. A probable origin of the discrepancy is the lack of a specific model for the boundary layer convection: without it, DC and TRB are used for this purpose, apparently in an excessive nonlinear behavior.

Concerning the IFCs obtained, it can be seen that DC has the opposite sign to Frcg in the lower and mid-atmosphere, unlike in the upper troposphere, and this is in agreement with the conclusion of Raymond and Herman (2011, p. 11) that “the convective quasi-equilibrium hypothesis appears to be valid in the lower troposphere but not in the upper troposphere.” Except in the boundary layer, LSC and TRB have similar responses, in both temperature and humidity,^{10} while DC is distinctive. The upward shift of deep convection is clearly seen with new heating near the tropopause.

We now conclude our illustration of the method: having a new technique in hand, some practice will be needed to learn how to use it properly. We hope other teams involved in GCM development will join us in investigating this new approach.

## 6. Concluding discussion

After having introduced a general formal framework in Part I and shown the equivalence between feedback and radiative sensitivity analyses in practice in the CMIP community, we have in this paper extended the formal analysis to render it more appropriate for numerical climate studies. The key point was to determine the rigorous dynamical conditions under which the regression technique provides relevant long-term climate sensitivity. These conditions are quite constraining, leading to time-scale low-pass filtering of the global climate evolution. Yet, at the same time, these constraints mitigate the limitations of the strict formal approach and extend the range of mechanisms that can be analyzed in GCMs with the same methodology. The same constraints also allow a regression of all available variables in the GCM with surface temperature anomaly because they are part of the same tangent linear system. This is of particular importance for CMIP analyses.

A major difference that we found with the standard analyses was that the lapse-rate feedback could not be chosen as a specific feedback and had to be distributed among the others, beginning with the Planck response, which already includes its own tropospheric lapse-rate response. We thus come to the same conclusion as Lu and Cai (2009) that the PRP lapse-rate feedback is irrelevant. Along the same line, Stocker et al. (2014) introduced the “effective radiative forcing” to account for that dynamical behavior.

Our main illustration concerned a global analysis of the CMIP5 quadrupling experiment with LMDZ. We were able to give details on the way fast mechanisms, not captured by the model underlying the regression method, explain much of the long-term evolution of the climate. The simplest dynamical model explaining the slow component of climate warming for the next century corresponds to the leading Lyapunov exponent of the 1D conceptual model underlying global mean analyses. It explains the PRP relation between the net TOA budget and the surface temperature anomalies in the regression method (the L2Obs relation). With only these two variables (*N* and Δ*T*_{s}), we found that two layers of the ocean (down to about 600 m) are *observable*—in the sense of the control theory—from the global trajectory of the difference between the perturbed experiment and the control one.

Adding the dynamical analysis to the regression method as we did has a double advantage: it ensures that the Lyapunov constraint is attained and also determines the “ocean heat uptake” efficiency in a coherent way. It furthermore shows that the response to the abrupt quadrupling and to the 1% ramp forcings share the same tangent linear model, thus providing cross proof of the results through the use of both experiments.

The results of fitting this model to the LMDZ5A trajectory (and also to other models that we have already analyzed)^{11} show that this two-layer model cannot reach equilibrium, suggesting that the oceanic warming structure could change dramatically after 100 yr of the quadrupled CO_{2} experiment. Following our formal constraint, the two-layer model is, by definition, not appropriate for studying fast mechanisms, which we defined as the mechanisms that rule out equality between the TOA radiative and surface energy budgets, a simple criterion of the troposphere being in equilibrium with the surface. This does not mean that fast mechanisms cannot be analyzed with the tangent linear system; it only shows that the two variables *N* and Δ*T*_{s} are not appropriate for this. But if the first time constant of about 10 yr found in section 4 is realistic, the forcing needs to include transient mechanisms that change the lapse rate, clouds, etc., in addition to the pure radiative forcing. This unfortunately makes it quite impossible to evaluate the real climate forcing needed in a remote sensing regression technique based on the L2Obs relation. Numerical models are essential for a detailed analysis of how the climate responds to perturbations, and this is why we need more efficient tools to meet the challenge of developing some understanding of how this complex natural system works.

We have proposed such a new method based on including the CTLS integration in GCMs. By introducing a finite version of the Gâteaux differential (FGD), we showed that each mechanism represented in a GCM can be differentially perturbed, the individual perturbation being or not propagated, through other chosen mechanisms, thus allowing the perturbed model behavior to be analyzed step by step from the exclusive to the inclusive approach. We used a first implementation of the method in LMDZ5 to analyze the water-cycle inclusive response to the CO_{2}-doubling experiment. In contrast to perturbed runs, the CTLS method provides results at given times, region by region, opening up the possibility of analyzing fast responses as well. We gave a short illustration of such analysis for the year following the abrupt forcing and a few more in the decadal term.

Concerning the Planck response, now that we know how to determine it, we have to find an agreement on its definition. As it represents the no feedback response, it includes all mechanisms that are not considered feedbacks. For example, we deliberately included surface thermal inertia in it, but what about soil models and deep ocean processes? In any case, advection by the unperturbed wind fields has to be included if the forcing response and the IFCs are to give the total response.

The 2×CO_{2} forcing was chosen for our purpose, but obviously any other forcing can be used. The amplitude can possibly be adjusted to save the linearity of the response if necessary. The FGD method is very flexible and allows us to analyze changes in parameters and basic physical laws as well as sensitivity to various forcings. This should make it easier to develop and evaluate subgrid parameterization, adjust results to experiments, and calibrate models, without the need to build adjoint models. All this might find applications in numerical weather prediction methods. In any case, it should help coping with the important issue of evaluating model uncertainty.

The coding of the method gave us some difficulty, essentially because of the way the interfaces to the diverse mechanisms are written in LMDZ5. The first step of implementation is to integrate a CTLS with no forcing, which should result in a null response, and also without changing the reference trajectory. Some variables in the arguments of subroutines have to be saved and restored when the subroutines are called again, etc. These difficulties are thus strongly dependent on the GCM. A question remains concerning the dynamic feedback: is it necessary to use a numerical method for the atmospheric circulation feedback or is the two-run method enough? The answer will hopefully come from practice.

The authors gratefully acknowledge useful discussions with J. Y. Grandpeix and also with Q. Libois during his Master internship on CMIP5 analyses. Special thanks to Professor M. Ghil and to our editor M. Cai for their suggestions to improve our manuscript. Two anonymous reviewers also gave us valuable advice. We wish to thank Susan Becker for careful proofreading of our first English version.

# APPENDIX A

## Extended Sensitivity Analysis

*x*to identify the extended factors). The new expressions come from the change in the forcing terms in Δ

*θ*

_{0}:and their analogs for matrix

*θ*

_{0}as the forcing term, but its new value now includes impact of troposphere temperature change on the surface. Note that the global factor

In consequence, the new sensitivity factors are just a redistribution of the previous ones.

*g*′)Δ

*T*

_{s}= Δ

*θ*

_{0}introducing a new feedback gainwhich looks fine: it corresponds to the new feedback loop in Fig. 1. The problem comes with the decomposition into partial gains, because we do not have the required equality

*θ*

_{s}+ Δ

*θ*

_{0}) on the rhs forcing term: see Eq. (10). With the new forms [Eq. (A5)], this is no longer possible; the normalized tropospheric profiles

^{A1}More explicitly, in the submatrix procedure, the troposphere responds with matrix

*λ*

_{l−r}: it plays the same role as the PRP lapse-rate sensitivity, which is the reason why we used that symbol. Here, it is the difference between the true Planck profile and other temperature profiles, and the feedback loop cannot be opened without changing the forcing.

_{i}—corresponding to a feedback loop would read

_{i}= (

^{−1}

_{i}as the extension of Eq. (9), with no evidence of the gain matrix

*φ*_{i}, attached to the mechanisms (the same as our separation into

_{i}). The IFCs are obtained at equilibrium with matrix

*l*th entry of vectors as

_{(i)}, we then haveHence we may writewhere matrix (

_{i})

_{d}is diagonal with entries:We conclude that the CFRAM decomposition leads to proper IFCs, see Figs. 3, 5, and 7 in Lu and Cai (2010). This and Eq. (A10) give Lu and Cai’s (2010) Eqs. (23) and (26).

^{A2}Accordingly, we share their conclusion that the PRP “lapse-rate feedback” is irrelevant.

These formal results also state that each IFC in CFRAM is associated with a feedback loop: this is what Cai and Lu verified numerically, successfully comparing the CFRAM decomposition and a special nonlinear suppression method (M. Cai and J. Lu 2012, personal communication), where the *n* inclusive feedback-loop responses were obtained from a difference between the unperturbed trajectory and *n* perturbed runs.

In summary, and because diagonal matrices commute in product, the extension of Eqs. (7), (9), and (12) are still valid at the cost of losing the simple scalar expressions: the (diagonal) exclusive gains become _{i} = _{i}/(

# APPENDIX B

## The Two-Layer Ocean Model

*T*

_{s}and thus no change to the troposphere eigenvectors. Solving Eq. (20) for Δ

*T*

_{s}gives

*e*-folding times. Before inverting to the time domain, we need the poles—the zeroes of the denominator—which depend on the discriminant:Using the smallness of the

*e*-folding time fractions, we obtain an approximation of the poles asCorresponding to the minus sign, we have the pole

*τ*

_{d}is replaced in coherence with

*τ*

_{2}:finally yieldingwith the coefficients

# APPENDIX C

## Implementing CTLS in GCMs

To explicit the logic of implementing the FGD scheme in a GCM, the diagram in Fig. C1 shows two columns. The left one corresponds to the main driver of physics in LMDZ. Without modifications (pink background), a subroutine SUB is called each time step with standard variables (Eta_seri) as input—which comprise atmospheric and surface temperature, humidity, and horizontal wind speed. The increment of these variables is output from SUB (d_Eta) and is used to update the standard variables.

The two rectangles contain the modifications, activated from the logical flag FGD. Basically, SUB is called with the perturbed variables as input: Eta_seri + DEta, where DEta are the CTLS-anomalies—Δ** σ** in Eq. (2), which may include the forcing [see Eq. (32)]. The FORTRAN variable dDEta is initialized from SUB output. After the second call to SUB (second rectangle), dDEta is set to its increment and added to DEta (time integration). When SUB has F90 in–out variables, these have to be saved before the first call, so that at the (standard) second call they are reset to their standard values.

The right column gives details concerning subroutines that are calling other processes—this is the case in LMDZ for the subsurface treatment (ocean, land, sea ice, and land ice), where yStateVar is extracted from the global grid to the subsurface grids then recast into the global grid. In this case, a new flag (mod_pert) must be added in SUB input and passed to SubSub. It is used to save and restore internally modified variables, some of which are input to SubSub. During the first call, when mod_pert is True, the in–out variables are saved and internal standard variables are perturbed—StateVar + DSVar, where DSVar are the internal anomalies—notice that in LMDZ, SubSub unfortunately does not compute increments. After the call to SubSub, DSVar is initialized. At the second call, internal variables are restored to their standard values before calling SubSub. Returning from SubSub, the internal anomalies DSVar are computed.

Hence, the favorable case for FGD implementation is that (i) there are no in–out F90 variables, and (ii) algorithms determine incremental values of the variables. This is, anyway, good practice in general for GCMs from numerical concerns.

Globally, the CTLS state variables DEta and DSVar are thus added to the standard variables and output in the history files. Potentially all other diagnostic anomalies can also be computed and stored, hence doubling the number of history variables.

## REFERENCES

Andrews, T., , and P. Forster, 2008: CO2 forcing induces semi-direct effects with consequences for climate feedback interpretations.

,*Geophys. Res. Lett.***35**, L04802, doi:10.1029/2007GL032273.Andrews, T., , J. Gregory, , M. Webb, , and K. Taylor, 2012: Forcing, feedbacks and climate sensitivity in CMIP5 coupled atmosphere–ocean climate models.

,*Geophys. Res. Lett.***39**, L09712, doi:10.1029/2012GL051607.Cacuci, D., 1981: Sensitivity theory for nonlinear system. I. Nonlinear functional analysis approach.

,*J. Math. Phys.***22**, 2794–2802, doi:10.1063/1.525186.Cacuci, D., , and M. Hall, 1984: Efficient estimation of feedback effects with application to climate models.

,*J. Atmos. Sci.***41**, 2063–2068, doi:10.1175/1520-0469(1984)041<2063:EEOFEW>2.0.CO;2.Colman, R., , and B. McAvaney, 2011: On tropospheric adjustment to forcing and climate feedbacks.

,*Climate Dyn.***36**, 1649–1658, doi:10.1007/s00382-011-1067-4.Cubasch, U., and Coauthors, 2001: Projections of future climate change.

*Climate Change 2001: The Scientific Basis,*J. T. Houghton et al., Eds., Cambridge University Press, 3–29.Deardorff, J. W., 1974: Three-dimensional numerical study of turbulence in an entraining mixed layer.

,*Bound.-Layer Meteor.***7**, 199–226, doi:10.1007/BF00227913.Dessler, A. E., 2010: A determination of the cloud feedback from climate variations over the past decade.

,*Science***330**, 1523–1527, doi:10.1126/science.1192546.Dufresne, J.-L., , and S. Bony, 2008: An assessment of the primary sources of spread of global warming estimates from coupled atmosphere–ocean models.

,*J. Climate***21**, 5135–5144, doi:10.1175/2008JCLI2239.1.Dufresne, J.-L., and Coauthors, 2013: Climate change projections using the IPSL-CM5 Earth System Model: From CMIP3 to CMIP5.

*Climate Dyn.,***40,**2123–2165, doi:10.1007/s00382-012-1636-1.Emanuel, K., 1991: A scheme for representing cumulus convection in large-scale models.

,*J. Atmos. Sci.***48**, 2313–2335, doi:10.1175/1520-0469(1991)048<2313:ASFRCC>2.0.CO;2.Forster, P., , and J. Gregory, 2006: The climate sensitivity and its components diagnosed from Earth Radiation Budget data.

,*J. Climate***19**, 39–52, doi:10.1175/JCLI3611.1.Forster, P., , and K. Taylor, 2006: Climate forcing and climate sensitivities diagnosed from coupled climate model integrations.

,*J. Climate***19**, 6181–6194, doi:10.1175/JCLI3974.1.Geoffroy, O., , D. Saint-Martin, , D. Olivié, , A. Voldoire, , G. Bellon, , and S. Tytéca, 2013: Transient climate response in a two-layer energy-balance model. Part I: Analytical solution and parameter calibration using CMIP5 AOGCM experiments.

,*J. Climate***26**, 1841–1857, doi:10.1175/JCLI-D-12-00195.1.Giering, R., , T. Kaminski, , R. Todling, , and S.-J. Lin, 2003: Generating the tangent linear and adjoint models of the DAO finite volume GCM’s dynamical core by means of TAF.

, Vol. 5, Abstract 11680. [Available online at http://www.cosis.net/abstracts/EAE03/11680/EAE03-J-11680.pdf.]*Geophysical Research Abstracts*Good, P., , J. Gregory, , and J. Lowe, 2011: A step-response simple climate model to reconstruct and interpret AOGCM projections.

,*Geophys. Res. Lett.***38**, L01703, doi:10.1029/2010GL045208.Gregory, J., 2000: Vertical heat transports in the ocean and their effect on time-dependent climate change.

,*Climate Dyn.***16**, 501–515, doi:10.1007/s003820000059.Gregory, J., , and J. Mitchell, 1997: The climate response to CO2 of the Hadley Centre coupled AOGCM with and without flux adjustment.

,*Geophys. Res. Lett.***24**, 1943–1946, doi:10.1029/97GL01930.Gregory, J., , and P. Forster, 2008: Transient climate response estimated from radiative forcing and observed temperature change.

*J. Geophys. Res.,***113,**D23105, doi:10.1029/2008JD010405.Gregory, J., , and M. Webb, 2008: Tropospheric adjustment induces a cloud component in CO2 forcing.

,*J. Climate***21**, 58–71, doi:10.1175/2007JCLI1834.1.Gregory, J., and Coauthors, 2004: A new method for diagnosing radiative forcing and climate sensitivity.

*Geophys. Res. Lett.,***31,**L03205, doi:10.1029/2003GL018747.Hall, A., 2004: The role of surface albedo feedback in climate.

,*J. Climate***17**, 1550–1568, doi:10.1175/1520-0442(2004)017<1550:TROSAF>2.0.CO;2.Hall, A., , and S. Manabe, 1999: The role of water vapor feedback in unperturbed climate variability and global warming.

,*J. Climate***12**, 2327–2346, doi:10.1175/1520-0442(1999)012<2327:TROWVF>2.0.CO;2.Hallegatte, S., , A. Lahellec, , and J.-Y. Grandpeix, 2006: An elicitation of the dynamic nature of water vapor feedback in climate change using a 1D model.

,*J. Atmos. Sci.***63**, 1878–1894, doi:10.1175/JAS3725.1.Hansen, J., , A. Lacis, , D. Rind, , G. Russell, , P. Stone, , I. Fung, , R. Ruedy, , and J. Lerner, 1984: Climate sensitivity: Analysis of feedback mechanisms.

*Climate Processes and Climate Sensitivity*.*Geophys. Monogr.,*Vol. 29, Amer. Geophys. Union, 130–162.Hansen, J., , M. Sato, , and R. Ruedy, 1997: Radiative forcing and climate response.

,*J. Geophys. Res.***102**, 6831–6864, doi:10.1029/96JD03436.Hansen, J., and Coauthors, 2005: Efficacy of climate forcings.

,*J. Geophys. Res.***110**, D18104, doi:10.1029/2005JD005776.Hourdin, F., and Coauthors, 2006: The LMDZ4 general circulation model: Climate performance and sensitivity to parametrized physics with emphasis on tropical convection.

,*Climate Dyn.***27**, 787–813, doi:10.1007/s00382-006-0158-0.Hourdin, F., and Coauthors, 2013: LMDZ5B: The atmospheric component of the IPSL climate model with revisited parameterizations for clouds and convection.

*Climate Dyn.,***40,**2193–2222, doi:10.1007/s00382-012-1343-y.Jarvis, A., , and S. Li, 2011: The contribution of timescales to the temperature response of climate models.

,*Climate Dyn.***36**, 523–531, doi:10.1007/s00382-010-0753-y.Lahellec, A., , and J.-L. Dufresne, 2013: A formal analysis of the feedback concept in climate models. Part I: Exclusive and inclusive feedback analyses.

,*J. Atmos. Sci.***70**, 3940–3958, doi:10.1175/JAS-D-12-0218.1.Lahellec, A., , S. Hallegatte, , J.-Y. Grandpeix, , P. Dumas, , and S. Blanco, 2008: Feedback characteristics of nonlinear dynamical systems.

,*Europhys. Lett.***81**, 60001, doi:10.1209/0295-5075/81/60001.Le Treut, H., , Z. Li, , and M. Forichon, 1994: Sensitivity of the LMD general circulation model to greenhouse forcing associated with two different cloud water parameterizations.

,*J. Climate***7**, 1827–1841, doi:10.1175/1520-0442(1994)007<1827:SOTLGC>2.0.CO;2.Li, C., , J.-S. von Storch, , and J. Marotzke, 2013: Deep-ocean heat uptake and equilibrium climate response.

,*Climate Dyn.***40**, 1071–1086, doi:10.1007/s00382-012-1350-z.Lindzen, R. S., , and Y.-S. Choi, 2009: On the determination of climate feedbacks from ERBE data.

,*Geophys. Res. Lett.***36,**L16705, doi:10.1029/2009GL039628.Lu, J., , and M. Cai, 2009: A new framework for isolating individual feedback processes in coupled general circulation climate models. Part I: Formulation.

,*Climate Dyn.***32,**873–885, doi:10.1007/s00382-008-0425-3.Lu, J., , and M. Cai, 2010: Quantifying contributions to polar warming amplification in an idealized coupled general circulation model.

,*Climate Dyn.***34**, 669–687, doi:10.1007/s00382-009-0673-x.Meehl, G., , J. Arblaster, , J. Fasullo, , A. Hu, , and K. Trenberth, 2011: Model-based evidence of deep-ocean heat uptake during surface-temperature hiatus periods.

,*Nat. Climate Change***1**, 360–364, doi:10.1038/nclimate1229.Raper, S. C. B., , J. M. Gregory, , and R. J. Stouffer, 2002: The role of climate sensitivity and ocean heat uptake on AOGCM transient temperature response.

,*J. Climate***15**, 124–130, doi:10.1175/1520-0442(2002)015<0124:TROCSA>2.0.CO;2.Raymond, D. J., , and M. J. Herman, 2011: Convective quasi-equilibrium reconsidered.

,*J. Adv. Model. Earth Syst.***3**, M08003, doi:10.1029/2011MS000079.Rochetin, N., , J.-Y. Grandpeix, , C. Rio, , and F. Couvreux, 2014: Deep convection triggering by boundary layer thermals. Part II: Stochastic triggering parameterization for the LMDZ GCM.

,*J. Atmos. Sci.***71,**515–538, doi:10.1175/JAS-D-12-0337.1.Stocker, T. F., and Coauthors, 2014:

*Climate Change 2013: The Physical Science Basis.*Cambridge University Press, 1535 pp. [Available online at http://www.climatechange2013.org/images/report/WG1AR5_ALL_FINAL.pdf.]Taylor, P., , M. Cai, , A. Hu, , J. Meehl, , W. Washington, , and G. Zhang, 2013: A decomposition of feedback contributions to polar warming amplification.

,*J. Climate***26,**7023–7043, doi:10.1175/JCLI-D-12-00696.1.Wetherald, R., , and S. Manabe, 1988: Cloud feedback processes in a general circulation model.

,*J. Atmos. Sci.***45**, 1397–1416, doi:10.1175/1520-0469(1988)045<1397:CFPIAG>2.0.CO;2.Wigley, T. M. L., , and M. Schlesinger, 1985: Analytical solution for the effect of increasing CO2 on global mean temperature.

,*Nature***315**, 649–652, doi:10.1038/315649a0.

^{1}

Compare Stocker et al (2014): chapters 7.1.3, 9.7.2.1, and 12.5.3.

^{2}

The Laplace variable *μ* has the dimension of per time.

^{3}

The “no feedback” response, called the Planck response, is the hypothetical response to an increase of CO_{2} concentration in the climate system when only radiation is involved; it serves as a gauge for other climatic feedbacks, such as water vapor and surface albedo (cf. Part I).

^{4}

If Δ*θ*_{s} is the last entry of Δ*η*^{r}, the *n*-dimensional row matrix

^{5}

FORTRAN code works more as an operator than a function, which is the reason why the Gâteaux differential has to be used; for instance, continuity in the variable is not required.

^{6}

It is the eigenvector associated with the leading Lyapunov exponent.

^{7}

The external product results in the null matrix with the exception of the last diagonal entry being unity.

^{8}

These are, however, the main source of difficulty for implementation.

^{9}

Only seen on oceanic surfaces (not shown).

^{10}

While having, however, very different meridional structures.

^{11}

Since this was written, Andrews et al. (2012) have given the regression results for 15 models involved in the CMIP5 abrupt experiment and proposed using a varying forcing or feedback for long term modeling.