WorksPhysics and Astronomy
Statistical Noise and Missing Forcing Limit Estimates of Earth's Feedback from Prescribed Sea-Surface Temperature Simulations
Abstract
Earth's feedback parameter measures how the Earth system responds to forcing and is inversely proportional to climate sensitivity. Sea-surface temperature (SST) patterns can modulate the value of the feedback parameter. Differences between observed and simulated SSTs have raised the question how the observed SSTs evolution impacts the global feedback. The standard method for estimating this effect uses observed SSTs prescribed to an atmospheric model with fixed pre-industrial atmospheric forcing. This method makes two assumptions: first, that the observed SSTs capture all relevant effects from the forcing, so that prescribing a time-varying forcing is unnecessary; second, that the temporal variations in the feedback parameter are driven by the evolving SST pattern and can be estimated via moving-window regressions. We test these assumptions by running controlled experiments in which SSTs from fully-coupled historical simulations are prescribed to an atmospheric model. We find that the prescribed-SST experiments fail to capture the coupled feedback evolution. This is explained by two effects: First, the absence of prescribed atmospheric forcing, and second, statistical noise arising from the computation of moving-window regressions. We find no evidence of any significant relationship between evolving SST patterns and changes in the feedback time series in a 4000-year pre-industrial control simulation. Any trends in the feedback parameter detected on timescales shorter than ~100 years are indistinguishable from statistical noise, making their attribution to the evolving SST pattern extremely difficult. Our results imply that prescribed SST simulations offer limited potential for inferring temporal changes in Earth's parameter over the observational period.
Introduction
The climate feedback parameter and its relation to future warming
Understanding and quantifying the sensitivity of Earth’s climate to anthropogenic greenhouse gas (GHG) emissions is one of climate science’s most fundamental questions. The Intergovernmental Panel on Climate Change’s (IPCC) 6th assessment report (AR6) estimates a "likely" range of 2.5–4 K for the equilibrium temperature after a -doubling, known as Equilibrium Climate Sensitivity (ECS). This assessment is based on combined evidence from climate feedbacks estimated from current observations and general circulation models (GCMs), historical trends in the global energy budget, and paleo-climatic evidence (Forster et al., 2021; Sherwood et al., 2020). In 1979, half a century earlier, the Charney Report provided a likely range of 1.5–4.5 K for ECS based on just two early-generation GCMs. This raises the question: why has the uncertainty in future warming remained so persistently, and where does it come from?
One can view the large uncertainty in future warming through the lens of a forcing-feedback framework (e.g., Gregory et al. (2004)): the energy imbalance of the Earth system—the difference between the outgoing long-wave (LW) radiation and net (incoming minus reflected) short-wave (SW) solar radiation at the top-of-atmosphere (TOA)—can be modeled as
The forcing-feedback framework assumes that , which corresponds to the heat uptake of the system (in watts per square meter), can be separated into a radiative forcing term and a response term . Importantly, the forcing term due to changes in the atmospheric composition is independent of the system state, while is typically modeled as being (linearly) dependent on the temperature of the system (Gregory et al. (2004); North et al. (1981); Wigley and Schlesinger (1985); Budyko (1969); Sellers (1969)):
The represents the change in the system’s temperature as a result of the imposed forcing , while is the linear feedback parameter:
One can then write
The feedback relates a given temperature change induced by the forcing to the radiative response , and is negative for a stable climate. It can therefore be interpreted as the restoring force of the system after a perturbation imposed by (Roe and Baker (2007)).
The importance of the feedback parameter lies in its relation to ECS: for a forcing resulting from a doubling of the atmospheric concentration, ECS is given by the temperature difference when the system reaches radiative equilibrium (), resulting in
The feedback parameter therefore directly provides information about how sensitive Earth’s climate is to perturbations: a strongly negative feedback means Earth is able to restore equilibrium after a forcing perturbation—such as an increase in concentration—with comparatively little increase in temperature, i.e., the system is efficient in restoring a balance resulting from the extra heat accumulated due to the forcing. A feedback parameter closer to zero (less negative), on the other hand, implies a system which needs to heat up more in order to restore radiative equilibrium. Due to its inverse proportionality to ECS, the feedback parameter is particularly relevant in determining future warming (Sherwood et al., 2020).
Existing methods to quantify the feedback parameter
There have been extensive efforts to quantify the feedback (i.e., ECS) in both models and observations. This endeavor is not straightforward, for multiple reasons. First, the linear feedback framework in Eq. (4) assumes a constant value for that is independent of timescale and type of forcing. Under this assumption, estimating can be done via the so-called ‘Gregory plot’ (Gregory et al., 2004): is retrieved as the slope of the regression of onto over the first 100-200 years in abrupt--quadrupling experiments. The linearity assumption of being constant is, however, only an approximation (Knutti and Rugenstein, 2015), and in reality has been shown to be forcing-dependent (e.g., Marvel et al., 2016; Yoshimori et al., 2011), state-dependent (e.g., [13] (Meraner et al. 2013)), and to vary due to evolving sea-surface temperature (SST) patterns via the so-called pattern effect (e.g., [8, 4] (Senior and Mitchell 2000; Rugenstein et al. 2016; Ceppi and Gregory 2017)). Early work on the pattern effect demonstrated that evolving SST patterns emerging when letting the GCM equilibrate in an abrupt--increase experiment result in a curvature of the Gregory plot towards a weakening of the feedback [4] (Rugenstein et al. 2016). Furthermore, several studies using linear response theory have shown that Earth’s feedback is best approximated not by a single , but rather three distinct but constant ’s, each governing a different component of the Earth system with its own characteristic timescale (e.g., Fredriksen et al., 2021, 2023; Proistosescu and Huybers, 2017).
The pattern effect has gained special attention in recent years, as the SST patterns modeled by state-of-the-art coupled atmosphere-ocean GCMs (AOGCMs) fail to reproduce some of the observed trends in SSTs (Seager et al., 2019; Olonscheck et al., 2020; Wills et al., 2022; Watanabe et al., 2021). Numerous studies suggest that the observed pattern of warming—with a strengthening of the East-West tropical Pacific SST gradient, for example—is one producing more negative feedbacks, similar to those during a La Niña phase [9] (Wills et al. 2022; Ceppi and Gregory 2017). In particular, idealized studies based on Green’s functions have demonstrated that cloud and lapse-rate feedbacks in the tropical Pacific are particularly sensitive to the pattern of warming (Zhou et al., 2017; Dong et al., 2019). This motivates the need to estimate the effect of the evolving observed SST pattern on the time-evolution of the global feedback parameter.
The second major difficulty in quantifying a time-varying global feedback parameter from observations is the need for accurate and long observational time series of , , and . However, uncertainties in estimates of are large, and satellite observations of exist only from 2001 onwards (Loeb et al., 2018), leading to highly uncertain observational estimates of the feedback (Meyssignac et al., 2023). For this reason, the state-of-the-art method to estimate the observed feedback evolution employs atmospheric GCMs forced by observed SSTs and sea-ice concentration (SIC), introduced formally as amip-piForcing runs by Gregory and Andrews (2016) (although SSTs have been prescribed to atmospheric models in many other earlier works). for different purposes, e.g., Gates et al. (1999); Boer et al. [12]; Cess et al. (1989); Andrews et al. [3]).
The idea behind the amip-piForcing setup for estimating the feedback evolution is that by prescribing the observed SST and SIC pattern as boundary conditions to the atmospheric model, one captures all relevant effects for the historical feedback evolution. The amip-piForcing method makes two important assumptions. First, atmospheric forcing is kept constant at preindustrial levels in amip-piForcing simulations, allowing to set in Eq. (4). This step implicitly assumes that all feedback-relevant processes of the true historical forcing are sufficiently well-captured in the observed SST and SIC evolution used as boundary conditions (Gregory and Andrews (2016)). The assumption allows for a simplification of the feedback calculation by avoiding the need of cumbersome estimations of . The feedback can now be estimated simply from time series of and , which can be directly retrieved from the amip-piForcing run output.
The second assumption lies in the way the feedback is estimated from and in practice: since the goal of the amip-piForcing runs is to quantify the effect of the evolving SST pattern on the evolving feedback , one wants to obtain a time series of which should capture possible SST-driven trends. In practice, this is done via 30-year moving window regressions of against , with being the slope of the regression:
By doing so, the implicit assumption made is that time variations in computed in this way are physically interpretable and result from SST variability.
Under these assumptions, numerous studies have used amip-piForcing runs to quantify the feedback evolution of the real climate system based on prescribed SST patterns ([5, 7] (Gregory and Andrews 2016; Zhou et al. 2021; Dong et al. 2021)). These studies all come to the conclusion that the feedback shows a strong trend towards more negative, stabilizing values in recent decades and attribute it to the specific pattern of SSTs. In particular, Dong et al. (2021) explicitly compare the observed feedback evolution as derived from amip-piForcing runs via Eq. (6) to the evolution of the feedback in coupled historical runs from CMIP6 (computed as in Eq. (6), but with in the denominator, using an estimate). They find that the historical simulations do not reproduce the stabilizing trend of the amip-piForcing simulations post-1980. They then attribute this effect to the difference in SST pattern between historical simulations and observations. This finding has wide-reaching implications, as a weaker or stronger feedback directly relates to a different future warming.
Given the importance of the amip-piForcing runs, it is crucial to test (I), whether they capture all relevant information to estimate the feedback when setting ; and (II), whether the variability found in using 30-year moving window regressions is driven by SST pattern variability. We test the first assumption by forcing the atmospheric component of the CESM2 model (Danabasoglu et al. (2020)) with evolving SSTs and SIC concentration in an amip-piForcing setup, but with SSTs and SIC taken from coupled historical simulations of CESM2 (§ 3.1). This allows for a direct comparison of the amip-piForcing feedback to the feedback estimated in the corresponding coupled historical run with identical SSTs and SIC boundary conditions. Our results uncover the inability of the amip-piForcing method to recover the feedback of the corresponding historical run with identical SSTs and SIC. We next show that this is in large part due to the omission of atmospheric forcing in the amip-piForcing setup, thus challenging the first assumption underlying the amip-piForcing setup.
Next, we study sources of variability in the feedback time series of a coupled pre-industrial control (piControl) simulation (§ 3.2). Here, by definition and if the second main assumption holds, variability in should be driven by the SST pattern. We find that variability on timescales shorter than years is primarily driven by an artificial statistical effect arising from applying a moving window regression over and . Our findings imply that even if SSTs are driving variability in , the causal effect is difficult if not impossible to separate from the statistical variability generation of the moving window process. Furthermore, we find no evidence of a link between specific SST patterns and the feedback evolution. Our results imply that a) trends found in from amip-piForcing runs cannot be clearly attributed to evolving SST patterns, and b) that any trends measured in on timescales shorter than years are within the typical variability timescale of and therefore potentially indistinguishable from internal variability.
Data & Methods
Experimental setup
We conduct atmosphere-only runs with the Community Atmosphere Model Version 6 (CAM6), which is the atmospheric component of the Community Earth System Model Version 2 (CESM2) ((Danabasoglu et al., 2020)). We conduct three distinct types of experiments with CAM6: amip-piForcing-style runs and amip-histForcing-style runs are outlined below and form the basis for our analysis. Additionally, we perform piClim-style runs required for estimating the time-varying forcing ((Pincus et al., 2016)), which are detailed in Appendix A. The individual simulations are summarized in Table 1.
| Name | Members | Simulation length / period | Forcing | SST & SIC boundary conditions |
| amip-piForcing-histSST | 11 | 1850–2014 | 1850 (fixed) | CESM2 historical SSTs and SICs from members , in the CMIP6 ensemble (Danabasoglu et al. (2020)) |
| amip-piForcing-ghgSST | 11 | 1850–2014 | 1850 (fixed) | CESM2 GHG2 SSTs and SIC from members 1–11 of the CESM2 single-forcing large ensemble (Simpson et al. (2023)) |
| amip-piForcing-piSST | 1 | 405 years | 1850 (fixed) | CESM2 piControl SSTs and SIC (Danabasoglu et al. (2020)) |
| amip-histForcing-histSST | 5 | 1850–2014 | historical | CESM2 historical SSTs and SICs from members , in the CMIP6 ensemble (Danabasoglu et al. (2020)) |
| piClim-control | 1 | 200 years | 1850 (fixed) | CESM2 piControl SSTs and SICs (climatology) (Danabasoglu et al. (2020)) |
| piClim-histall | 1 | 1850–2014 | historical | CESM2 piControl SSTs and SICs (climatology) (Danabasoglu et al. (2020)) |
Table 1. AMIP-style simulations performed with the CAM6 model. piClim-style runs are detailed in Appendix A.
“amip-piForcing”-style runs
These simulations are identical to the amip-piForcing setup ((Gregory and Andrews, 2016; Webb et al., 2017)), except having different (i.e., simulated rather than observed) SST and SIC boundary conditions. In all amip-piForcing-style simulations we fix atmospheric forcing at preindustrial (1850) levels. We call these runs amip-piForcing-[...] in Table 1. The [...] stands for the different SST and SIC used as boundary conditions. We conduct three distinct types of these amip-piForcing-style experiments, differing only in the prescribed SST and SIC boundary conditions:
amip-piForcing-histSST: For these 11 runs we prescribe SST and SIC from from 11 fully coupled historical initial condition members of the CESM2 model performed as part of CMIP6 ((Danabasoglu et al., 2020)). The forcing in the coupled historical runs is prescribed according to the CMIP6 protocol with historically evolving best-estimates of all
known forcing agents from 1850–2014 (Eyring et al., 2016). The SSTs taken from those runs therefore exhibit a warming over the historical period.
amip-piForcing-ghgSST: For these 11 runs we prescribe SST and SIC from 11 fully coupled initial condition members from the CESM2 single-forcing large ensemble which have historically evolving GHG-concentrations as forcing (Simpson et al., 2023). The coupled runs from which we take the boundary conditions are called GHG2, following the convention of Simpson et al. (2023). The coupled GHG2 runs therefore have all other forcings fixed at 1850 levels, apart from GHG forcing, which evolves as in the historical simulation (Simpson et al., 2023). The SSTs of these simulations also exhibit the GHG-induced warming over the historical period.
amip-piForcing-piSST: For this run we prescribe SSTs and SIC from the fully coupled CESM2 piControl simulation produced as part of CMIP6 (Danabasoglu et al., 2020). Forcing in the coupled piControl simulation is held fixed at 1850 levels (Danabasoglu et al., 2020). We run the amip-piForcing-piSST simulation for 410 years.
“amip-histForcing”-style runs
These runs differ from the standard amip-piForcing, because apart from prescribing time-varying SST and SIC, we also prescribe time-varying historical atmospheric forcing. One ensemble of such runs is performed (Table 1):
amip-histForcing-histSST: These 11 runs have SSTs and SIC identical to the 11 members of the amip-piForcing-histSST ensemble (and therefore, identical to the fully coupled historical runs), but also time-varying historical forcing prescribed to the atmosphere. CAM6 in these runs therefore experiences an almost identical forcing from ocean and atmosphere as
the corresponding 11 members from the coupled historical CESM2 runs (up to variability from the initial conditions and numerical differences).
Calculation of the feedback parameter from numerical simulations
The feedback is generally estimated by Eq. (3): (e.g., (Gregory et al., 2004; Gregory and Andrews, 2016)):
In general, the radiative response is given by .
Each experimental setup implies a radiative response, which is calculated according to the and present in the corresponding experimental setup. All amip-piForcing-style experiments and the piControl simulation imply no forcing by definition, i.e., (Gregory and Andrews, 2016). In all other coupled and atmosphere-only runs with time-varying forcing, needs to be estimated separately (see Sec. A).
In order to obtain time series of , we follow the moving-window-regression method used in literature (e.g., [5, 7] (Gregory and Andrews 2016; Dong et al. 2021)). This consists of taking 30-year moving windows, for which the numerator is regressed onto the denominator , as in Eq. (6):
The time series of then results from the time series of the slopes of these moving-window regressions. is calculated from the CMIP6 variables as the difference in TOA downward shortwave and upward short-and longwave fluxes, rsdt - rsut - rlut. For the simulations done with CESM2, we calculate the net TOA imbalance from the CESM2 output variables FSNTOA - FLUT, being the difference of net shortwave and outgoing longwave radiation. For we use the variable tas for CMIP6 output and the variable TREFHT for CESM2 raw output.
We estimate the forcing using three different methods, in order to test the sensitivity of our results to the estimate. These are outlined in detail in Appendix § 6. For all figures of the main text, we display results where has been estimated following the Radiative Forcing Model Intercomparison Project (RFMIP) method (Pincus et al., 2016). This method involves subtracting an atmosphere-only control run with fixed climatological SSTs, from an atmosphere-only run where the time-varying forcing is prescribed but SSTs are also held climatologically fixed (see Appendix § 6).
Monte-Carlo simulations for variability analysis
In § 3.2 we turn to idealized coupled piControl and atmosphere-only piClim-control simulations in order to analyze variability in . For this purpose, we attempt to model and from piControl and piClim-control statistically: Our goal is to find a good statistical model of and in these simulations, so that we can generate Monte-Carlo time series of and with similar statistical properties as the true time series.
Modeling and in piClim-control statistically
We hypothesize that and in the piClim-control simulation can be modeled as jointly normally distributed. This is based on the fact that piClim-control is forced by climatologically fixed, monthly-varying SSTs and SIC. We therefore expect to have no SST-driven variability in and on annual timescales. Since ocean memory is erased, we also do not expect to have any autocorrelation in , , or jointly in and . To test the normality hypothesis, we perform first a uni-variate normality test for and separately based on D’Agostino and Pearson (1973). We find p-values of 0.99 and 0.82 for and , respectively, meaning that there is insufficient evidence to reject the null-hypothesis that and are normal. Similarly, the p-value for the multivariate-normality test based on Henze and Zirkler (1990) is 0.49, meaning that we can justify the assumption of multi-variate normality of and .
In order to generate Monte-Carlo simulations of and with the statistical properties of piClim-control, we generate one million samples from a bi-variate Gaussian distribution with mean vector and covariance matrix estimated from the sample statistics of and .
Modeling and in piControl statistically
In the coupled piControl simulation we expect memory from ocean variability to lead to strong autocorrelation in (e.g., (Hasselmann, 1976; Frankignoul and Hasselmann, 1977)). Since coupling between and is now free, we also expect there to possibly be cross-(auto)correlations between and . For this reason we choose to model the piControl time series of and via a vector autoregressive model (VAR) of order (Lütkepohl, 2005): Let , then we can model via
The matrices contain autoregression coefficients for lags , and the noise term comes from a bivariate normal distribution with zero mean. We fit the VAR process with automatic optimal lag-order selection via the Akaike information criterion. This leads to an optimal lag order of . We then generate one million years of and from the VAR(5) model fitted on piControl data.
Spectral densities of the time series were computed with the multi-tapering method (Huybers and Curry, 2006). The computation was done over the whole length of the respective simulations, after linearly de-trending the time series of the variables , , and .
Bootstrapping of periods in piControl for variability analysis
In § 3.4 we investigate whether periods of weakening or strengthening feedback can be associated to particular SST patterns. For this analysis, we use a 4000-year long fully-coupled pre-industrial control simulation performed with CESM2, which we term piControl-long. It has identical boundary conditions as the CMIP6-version of the CESM2 piControl introduced earlier (Danabasoglu et al., 2020).
We compute the feedback in the piControl-long simulation from 30-year moving window regressions as in Eq. (7). We then seek periods during which the feedback weakens (i.e., increases towards zero), or strengthens (i.e., becomes more negative). To select such periods we first smooth the feedback time series using a Savitzky-Golay filter with window length 81 and polynomial order 4 (Savitzky and Golay, 1964). We then select periods where the derivative of the smoothed time series is positive or negative for at least twenty consecutive years. In total, we identify 49 periods of feedback strengthening and 49 of feedback weakening in the 4000-year long piControl simulation. The time series of the feedback in the piControl simulation together with the selected periods of weakening and strengthening is displayed in Figure C1. Our results remain largely unchanged if the minimum period length is shortened to 10 or increased to 30 years.
Figure C1. (a): The raw (thin grey line) and smoothed (thick black line) feedback time series in the piControl-long simulation. Red and blue periods indicate regions of at least 20 consecutive years identified as having an increasing (weakening) or decreasing (strengthening) feedback according to the methodology described in § 2.7. (b): Histogram of the average local slopes of the trends in the periods identified in (a).
We then compute the surface air temperature trend for each grid cell during each period of weakening and strengthening. We then average over all periods of strengthening or weakening feedback, in order to obtain the composite of the surface air temperature trends. We also compute the agreement fraction on the sign of the trend between the samples that make up the composite. The agreement fraction is the number of samples in the composite which belong to the majority group (i.e., if the majority of samples show a positive trend in a grid cell, then we count the number of samples which show this positive trend), divided by the total number of samples in the composite (here, 49 for both weakening and strengthening feedback periods).
To test the significance of the composite surface air temperature trends and their agreement, we employ bootstrapping on the 4000-year long time series. We sample a set of non-overlapping periods of the same length and number as the ones used to compute the composites, 10 000 times in total. This gives us a null-distribution of randomly sampled periods in piControl-long during which any trends in the composites occur due to random noise. We would expect the average agreement on the sign of trends between periods of any bootstrap sample to be close to 0.5 (half of sampled periods show a positive, and half a negative trend). For the significance of the trend in a given grid cell, we compute the two-sided p-value as
where is the empirical quantile of the composite trend with respect to the bootstrapped distribution of composites. For the significance of the agreement fraction we compute the one-sided p-value based on the empirical quantile of the agreement between the samples of the composite as
We use a one-sided p-value because the agreement fraction is by definition, and exactly 0.5 when there is no agreement (i.e., when exactly half of the composites show a positive and the other half a negative trend). We therefore only test whether the agreement in our feedback strengthening or weakening composites is larger than that of the bootstrapped distribution.
With the above procedure, we perform a hypothesis test on each grid cell in order to test its significance, either for the trend of the composite, or for the agreement fraction among the composite. If the p-value for a given grid cell is smaller than our chosen significance level of , we reject the null hypothesis (the bootstrapped distribution) and term the values ‘significant’, in the sense of being different from random noise. However, this procedure quickly leads to the problem of multiple testing, according to which we naturally expect to find around of our samples to be significant, purely by construction of the statistical test ((Livezey and Chen, 1983; Wilks, 2016)). The problem is further aggravated by strong spatial correlations in climatic fields ((Wilks, 2016)). We therefore employ the correction by Benjamini and Hochberg [11] which controls for the false discovery rate at the 10% level and is the recommended correction for use cases in climate science in the presence of spatial (auto)correlation ((Wilks, 2016)).
Results
Comparing amip-piForcing-style and coupled runs with identical SSTs & SIC
We compare global time series of surface air temperature and net, SW, and LW fluxes at the TOA in the fully coupled historical simulations and their counterpart amip-piForcing-histSST simulations, with pairwise identical SSTs (Figure 1a–Figure 1d). The coupled and amip-piForcing runs differ only little in their global surface temperature evolution (a), but substantially in the radiative fluxes (b–d). The coupled runs exhibit a slightly higher surface temperature and warming, which has also been found by Andrews et al. [7] and is well-understood: in the absence of increasing atmospheric concentrations, the land temperature, which is not prescribed in amip-piForcing runs, is cooler than in the coupled run with forcing. It warms only indirectly due to the warming of the SSTs.
Figure 1. Results from amip-piForcing-histSST simulations (blue) compared to the CESM2 historical simulations (orange) with pairwise identical SSTs (i.e., amip-piForcing-histSST are forced with SSTs from CESM2 historical fully coupled simulations) for (a) global surface air temperature, (b) net TOA imbalance, (c) net TOA LW imbalance, (d) net TOA SW imbalance, and (e) radiative response . For amip-piForcing-histSST, is identical to , the net TOA imbalance shown in (b). For historical runs, is calculated as (net TOA imbalance as in (b), minus a forcing estimate). is derived using the RFMIP method (see Appendix § 6).
Radiative fluxes exhibit differences between the coupled and amip-piForcing setup, which are indicative of the absence of the time-evolving forcing in the amip-piForcing runs. In the fully coupled runs, as in the real world ((Forster et al., 2025)), the net radiative flux at the TOA becomes increasingly positive in recent decades (Figure 1b). In contrast, the amip-piForcing trend shows a decrease in the net flux, which was also identified by Andrews et al. [3]: this is owed to progressively warmer SSTs, which lead to increased outgoing LW radiation. Since there is no increasing concentration in amip-piForcing, there is also no additional greenhouse effect to compensate the increased outgoing LW radiation. This is confirmed by the increasingly more negative LW flux in the amip-piForcing runs (Figure 1c). In contrast, the LW flux of the coupled run first increases, owing to the trapping of LW radiation from increased , and only starts to decrease again once the Planck feedback starts to dominate due to increasingly warmer surface temperatures ((Myhre et al., 2025)).
In the SW flux, the main difference between amip-piForcing and coupled runs hints to the effect of aerosols and volcanoes. The coupled runs exhibit the characteristic dimming and brightening ((Wild et al., 2026)) along with spikes of lower SW flux following volcanic eruptions. The amip-piForcing runs, on the other hand, only exhibit a gradual increase in the incoming SW flux in the last years of the simulation. This increase could be owed to both the decreasing SIC, as well as positive cloud feedbacks which are more prevalent with increasing SSTs in both the coupled and amip-piForcing simulations.
The computation of the feedback according to Eq. (3) requires the regression of the radiative response , shown in Figure 1e, onto the surface air temperature in Figure 1a. At first sight, the differences in shown in panel e) do not show a large systematic difference between the coupled setup () and the amip-piForcing setup (). The radiative response in the amip-piForcing setup lies slightly above the one from the coupled setup, although the difference is small. Combined with the the comparatively small difference in the time series (panel a), one would expect a similar feedback time series (computed as 30-year moving window regressions of against ) between the coupled and amip-piForcing runs.
The ensemble mean of the feedback time series for amip-piForcing and historical coupled runs with pairwise identical SSTs is shown in Figure 2a. The feedback is systematically different in the amip-piForcing-histSST and coupled historical runs, with the coupled feedback being consistently weaker than the feedback derived from amip-piForcing runs (Figure 2a). The average correlation between the coupled and amip-piForcing feedback across the ensemble is low (0.23), and the mean squared error (MSE) 0.9 (Table 2). The bias in the feedback estimation is particularly pronounced in the historical setup (Figure 2a,Figure 2c).
Figure 2. (a) The feedback parameter in amip-piForcing-histSST runs (blue) compared to CESM2 historical runs with pairwise identical SSTs. Dotted lines show the feedback calculated on ensemble means of and (also in (b)). (b) The feedback parameter in amip-piForcing-ghgSST (blue) compared to the CESM2 single-forcing GHG2 runs (orange), with pairwise identical SSTs. Panel (c) and (d): black line shows the difference in the feedback between the coupled and atmospheric setup (blue minus orange lines in (a) and (b)) for historical and GHG2 SSTs, respectively. The purple line shows the average MSE across the 11 members of the ensemble for the simulations with historical SSTs (c), and GHG2 SSTs (d).
| AMIP experiment | Target coupled simulation | Ensemble average MSE | Pearson Correlation |
| amip-piForcing-histSST | historical | 0.90 | 0.23 |
| amip-piForcing-ghgSST | GHG2 | 0.59 | 0.39 |
| amip-histForcing-histSST | historical | 0.48 | 0.64 |
| amip-piForcing-piSST | piControl | 0.27 | 0.49 |
Table 2. MSE and Pearson correlation of feedback time series for paired simulations with identical SSTs but different setup. The table shows how feedback time series of the AMIP-type experiment compare to the feedback time series of the corresponding coupled simulations with identical SSTs and SIC. The MSE and Pearson correlation are averaged across all members and model years.
For the period 1940–2000 the ensemble mean of the coupled feedback shows a pronounced weakening and strengthening, deviating from the amip-piForcing feedback by >1 around 1970. This difference is particularly striking when looking at the similar and time series for the amip-piForcing and coupled simulations (Figure 1a,Figure 1e). The mean squared error between the coupled and amip-piForcing feedback across the eleven members also peaks in this period (Figure 2c). The period of largest deviation coincides with the period of strongest aerosol forcing (Forster et al., 2021). This could indicate that the missing aerosol forcing in the amip-piForcing causes this discrepancy.
We therefore also compare the difference in amip-piForcing and coupled feedback in simulations with historical GHG forcing for the coupled runs, and pairwise identical SSTs in the amip-piForcing runs. In the GHG-only setup, the strong systematic deviation in the feedback time series during the 1940–2000 period disappears (Figure 2b,Figure 2d). However, the consistent negative bias of the amip-piForcing feedback remains throughout the entire period (Figure 2b,Figure 2d). The average Pearson correlation of the coupled and amip-piForcing-ghgSST feedback time series increases to 0.39 compared to the setup with historical forcing (Pearson corr.=0.23), and the MSE drops from 0.9 to 0.59 (Table 2). The improvement in the GHG-only setup supports the hypothesis that the missing aerosol forcing in amip-piForcing-histSST runs is responsible for the large bias around 1940–2000.
We test the sensitivity of our results to the way the feedback parameter is calculated. First, the results remain qualitatively the same if the feedback time series is directly computed on ensemble mean quantities of and (dotted lines in Figure 2a,Figure 2b), rather than averaging the individual feedback time series of individual members. Second, we test the robustness of our results with respect to the forcing estimate used for computation of the coupled feedback. The results remain largely unchanged if the forcing is either i) corrected for the land surface response using the Hansen et al. (2005) method, or ii) computed via the AerChemMIP method (Collins et al., 2017) (Figure A1 and Figure A2). This consolidates our findings that the bias in the feedback between coupled and amip-piForcing-style runs with identical SSTs and SIC is owed to an inherent difference in the experimental setup.
Figure A1. Three different forcing estimates, obtained via the methods described above in Appendix A. The blue shading shows the ensemble spread from the eleven members used to derive the estimate.
Figure A2. Identical to Figure 2, but with two alternative estimates of the feedback in the coupled runs calculated with the different estimates (see Appendix § 6): the first correction adds the Hansen et al. (2005) land correction to account for warming land masses (brown line). The second alternative estimate is based on the Collins et al. (2017) method (pink line). The correction can only be estimated for the historical setup, as the necessary runs for the GHG-only setup do not exist.
The results from Figure 2 demonstrate that the amip-piForcing setup fails to capture the evolution of the feedback in the coupled setting, despite identical SSTs. The particularly large difference in the feedback during the period of strong aerosol forcing suggests that the lack of atmospheric forcing in the amip-piForcing simulations might be causing the mis-estimation of the true coupled feedback.
Zooming into individual ensemble members of the amip-piForcing-histSST and corresponding historical runs reveals large differences between the best- and worst-performing members in terms of mean squared error (MSE) (Figure 3, a, d, g). In the member with the worst MSE (a), the feedback time series happens to be anti-correlated, despite identical SSTs. During the 1960–1980 period the feedback differs by up to 2 between amip-piForcing and historical member (Figure 3a). However, even the member with the lowest MSE (g) shows large discrepancies in the computed feedback time series between amip-piForcing and historical: for example, the negative trend in the amip-piForcing feedback from 1975 onwards, which is similar to the pronounced negative trend found in many amip-piForcing experiments using observed SSTs (e.g., Andrews and Webb, [2]; Andrews et al., [7]; Zhou et al., 2021; Dong et al., 2021), is completely absent in the corresponding historical fully coupled run, despite identical SSTs and SIC. This implies that the conclusion made in the cited literature above, namely that the negative trend in the feedback is owed to the specific observed SST pattern driving negative feedbacks, is not robust.
Figure 3. Feedback parameter for three individual members (columns) with pairwise identical SSTs: amip-piForcing-histSST simulation (blue) and the corresponding coupled historical simulation (orange). Left column (a), (d), (g): the raw feedback parameter time series. Middle column (b), (e), (h): high-pass filter applied to the corresponding raw feedback parameter from the left column. The high-pass filter (Butterworth filter) has a cutoff frequency of 0.05, eliminating variability occurring on timescales >20 years. Right column (c), (f), (i): low-pass filtered feedback (the residual resulting from the subtraction of the high-pass filter from the original feedback parameter time-series). The three members shown are those with the "worst" (top row), "median" (middle row), and "best" (bottom row) MSE between coupled and corresponding atmosphere-only run across the 11-member ensemble.
We decompose the variability in the feedback time series using a high-pass Butterworth filter (Butterworth, 1930), which filters any variability occurring on timescales years (Figure 3b,Figure 3e,Figure 3h). On short timescales, the coupled and amip-piForcing feedback time series exhibit much better agreement, with an average Pearson correlation across members of 0.54 and an average MSE of 0.04. Contrarily, the low-pass residual, obtained by subtracting the high-pass filtered time series from the original feedback time series, shows even worse agreement than the original time series. Our results show that the short-term (interannual and decadal) radiative and temperature fluctuations, which are largely governed by SST variability, are mostly reproduced correctly in the amip-piForcing runs. This suggests that a slower driver, rather than interannual SST variability, is driving the discrepancy between the coupled and amip-piForcing feedbacks. This is unexpected, as long-term variability in is assumed to come from SST variability, which is, however, the same in both runs. The missing forcing in amip-piForcing runs therefore remains a plausible candidate for explaining the discrepancy.
To gain insight into the spatial structure of differences between amip-piForcing-histSST and the coupled historical runs we compute temporal trends of surface air temperature (), net TOA flux (), LW TOA flux, and SW TOA flux for the period 1981–2014, for each grid cell (Figure 4). Spatial temperature trends show that amip-piForcing simulations warm less than the coupled run during this period, particularly over Northern Hemisphere land masses (Figure 4a–Figure 4c). An exception is South-East Asia and the Indian peninsula in particular. This pattern again hints at the effects of aerosol forcing in the coupled simulation, which are absent in the amip-piForcing simulations. Aerosol forcing decreases in the 1981–2014 period in most of the Northern Hemisphere apart from South-East Asia, leading to an increased warming signal (Forster et al., 2021). The lower warming in amip-piForcing overall and over land is also expected due to the lack of GHG forcing in the amip-piForcing runs and the fact that land temperatures are not prescribed in amip-piForcing.
Figure 4. Ensemble mean trends for the period 1980-2014 for the variables surface air temperature (a–c), net TOA flux (d–f), long-wave TOA flux (g–i), and short-wave TOA flux (j–l). The first column shows the trends in the amip-piForcing-histSST experiments. The middle column shows the trends in the CESM2 historical members whose SSTs are prescribed in the amip-piForcing-histSST simulation. The third column shows the difference in the trends computed as amip-piForcing-histSST minus historical.
The 1981–2014 trends in net TOA radiation differ in sign almost everywhere on the globe (Figure 4d–Figure 4f), as expected from the diverging time series in Figure 1b. Again, Northern Hemisphere land masses stand out as areas where the amip-piForcing trends underestimate the historical trends most (Figure 4f). The decreasing aerosol forcing in historical runs—which is absent from the amip-piForcing runs—could explain the stronger net radiation trends in the Northern Hemisphere during this period. This hypothesis is further supported by the SW trends (Figure 4j–Figure 4l), which dominate the net trends: much stronger positive SW TOA imbalance trends are found in Europe and North America, whereas South-East Asia—which does not experience aerosol reductions in the 1981–2014 period but an increase (Yang et al., 2024)—even shows slightly negative SW EEI trends for the same period. The LW imbalance (Figure 4g–Figure 4i) also aligns with this interpretation: while the sign of trends is negative almost everywhere on the globe in both simulation setups, the amip-piForcing simulations show a less negative LW trend over the Northern Hemisphere continents, with the exception of South-East Asia. Again, if decreasing aerosols lead to a stronger warming trend over the NH land masses in the coupled runs, this could explain the stronger negative LW trends there compared to the amip-piForcing trends.
Our results remain qualitatively the same if we only consider trends over the boreal summer months or for clear-sky fluxes (Figure B1, Figure B2). If snow albedo effects were the main contributor to the differences, we would not expect to see this effect in the boreal summer months. If the differences were mainly owed to cloud effects, we should not expect to see the same signature also in clear-sky fluxes. This consolidates the hypothesis that the missing in amip-piForcing runs explains most of the spatial differences in Figure 4.
Figure B1. As Figure 4 of the main text, but with SW and LW flux trends computed on clear-sky fluxes.
Figure B2. As Figure 4 of the main text, but with trends computed only for the summer season (June, July, August).
We now test the hypothesis that the absence of time-varying atmospheric forcing in amip-piForcing simulations can explain the bias in the feedback calculation. The feedback calculated from amip-histForcing-histSST runs—which are identical to the amip-piForcing-histSST runs but now have time-varying prescribed historical radiative forcing—is shown in Figure 5 (dark blue lines). On average, the Pearson correlation between the feedback in amip-histForcing and coupled historical runs shows a large improvement to 0.64, compared to 0.23 for the amip-piForcing runs (Table 2), while the MSE is reduced from 0.90 WmK to 0.48 WmK. However, there remains a negative bias of the amip-histForcing-histSST ensemble mean compared to the coupled ensemble mean during the 1940–2000 period, albeit reduced compared to the piForcing setup (Figure 5a). For some members (e.g., Figure 5b, Figure 5d, Figure 5e, Figure 5k), the amip-histForcing feedback indeed tracks the coupled feedback much better (orange lines), however this is not the case for others (e.g., Figure 5c, Figure 5g, Figure 5h, Figure 5i).
Figure 5. (a): like Figure 2a, but with the ensemble mean and spread from amip-histForcing-histSST included in dark blue. (b-l): feedback parameter for individual members, all members in one panel have identical SSTs. Light blue lines show the amip-piForcing-histSST feedback, as in Figure 3a,Figure 3d,Figure 3g. Orange line shows the feedback in the corresponding coupled historical run, and the dark blue line shows the feedback estimated from amip-histForcing-histSST runs, recovered via Eq. (7) with the forcing estimated as .
The strong improvement in the average agreement between coupled and amip-histForcing feedback (compared to coupled and amip-piForcing feedback) suggests that the assumption made in amip-piForcing runs leads to a substantial bias in the feedback estimation (Table 2). In particular, the spatial analysis in Figure 4, the characteristic increase and decrease in the mean bias of the feedback during 1940–2000 (Figure 2a,Figure 2c), and the reduced bias in the GHG-only setup (Figure 2b,Figure 2d) strongly hint to aerosols being particularly relevant in causing the discrepancy. Aerosols are a) fairly short-lived and b) are generated over land: this leads to the fact that only a small part of the aerosol forcing is reflected in the SSTs. GHGs, in contrast, are
well-mixed, long-lived, and have been increasing over the entire historical period, so their radiative forcing has had time to be reflected in the SSTs. Because only SSTs—but not land temperatures—are prescribed in the amip-piForcing setup, forcing effects which are not reflected in the coupled runs’ SSTs will not end up in the amip-piForcing run.
Nonetheless, it is remarkable to see that even with prescribed identically as in the coupled historical runs, there remain substantial differences in the feedback time series (e.g., Figure 5a, Figure 5b, Figure 5f, Figure 5g). In these runs there are periods where the feedback time series diverge substantially despite identical prescribed SSTs and atmospheric forcing. This suggests that something other than SSTs or forcing is driving these differences. In particular, it suggests that SST and SIC alone are unable to fully constrain the feedback response in simulations with identical SSTs and SIC. This finding implies that inferring Earth’s feedback from amip-piForcing simulations with prescribed observed SSTs, SIC and atmospheric forcing may still not provide an unbiased estimate of Earth’s observed feedback evolution.
The remaining differences between amip-histForcing-histSST and coupled historical simulations could arise for two reasons. First, it is possible that there are sources of variability in the feedback time series, which are unrelated to the SST or SIC pattern. Such sources could be, for example, physical variability originating over land or the atmosphere and unconstrained by SSTs, or statistically-induced variability due to the computation method of via moving-window regressions. In the following section, we turn to an idealized pre-industrial control setting without potentially confounding radiative forcing in order to investigate such sources of variability in the feedback estimates. Second, it is possible that the amip-type simulation does not capture the response of the coupled system correctly. This has been demonstrated in a simple stochastic model and verified in early generation GCMs already by Barsugli and Battisti [10] and the follow-up work of Bretherton and Battisti (2000). In these works, the authors show theoretically and with evidence from GCMs that the severed coupling between atmosphere and ocean leads to unphysically enhanced atmospheric variance and surface heat fluxes. Other works have expanded on these findings and shown that ocean-atmosphere coupling matters for accurate representation of ENSO-teleconnections, for example (Alexander et al., [1]). Whether the bias between amip-type and coupled simulations seen in Figure 5a is a result of the missing atmosphere-ocean coupling in amip, or is merely an effect of the small sample size, remains unclear and is beyond the scope of this study.
Sources of variability in the feedback parameter
We now turn to idealized piControl simulations in order to investigate sources of variability in the feedback parameter time series which could potentially explain the differences between coupled and amip-type simulations despite identical SST, SIC and forcing. Figure 6a-Figure 6b show the surface air temperature () and net TOA radiation () in the coupled CESM2 piControl simulation (orange) and its amip-piForcing-piSST counterpart (blue), which has been forced with the SSTs and SIC of the coupled run. The time series of and are strongly correlated between amip-piForcing and coupled run (Pearson corr.: 0.93 for and 0.74 for ). When the feedback is computed from the time series of and , however (Figure 6c), the correlation drops to 0.49. Although the Pearson correlation of 0.49 between the two feedback time series is stronger than the correlation found between the amip-piForcing-histSST and corresponding historical runs (0.23), there nonetheless remain substantial differences in the estimated feedback. Notably, there are periods—such as years 30–80, or years 200–240 of the simulation—during which the feedback time series diverge substantially. How can this be the case, when the AMIP version of and seems to follow the
Figure 6. Feedback calculated in the amip-piForcing-piSST run (blue) compared to the coupled piControl run with identical SSTs (orange).
coupled and so closely, with no apparent bias by visual inspection?
Our results suggest that minor variations in and between the coupled and amip-style simulations, which are not obvious from visual inspection (Figure 6a, Figure 6b), become amplified through the feedback calculation and result in feedback discrepancies over large periods. This must be the case, because if there were substantial differences in the and between the two setups, we should have been able to see them in Figure 6a, Figure 6b. In addition, AMIP simulations have generally been routinely tested in terms of their capability to reproduce global variables like net TOA radiation and temperature for a long time (e.g., Gates et al. (1999)). Therefore, we hypothesize that what we are seeing is a statistical artifact of the moving window regressions, where small deviations in the and time series—for instance due to unconstrained variability over land regions and in the atmosphere—might lead to large periods in the moving-window feedback time series, where the amip-piForcing and coupled simulations diverge.
The Yule-Slutzky effect: how random noise generates multi-decadal variability
To test this hypothesis, we turn to the piClim-control simulation, where SSTs and SIC are held climatologically fixed and vary only with the seasonal cycle (Figure 6d–Figure 6f). First, the output of and (Figure 6e, Figure 6f) reveals that and have a much smaller variance than in the coupled piControl simulation, and in particular, do not seem to exhibit long-term variability. The joint autocorrelation function of and is not significant at any lag for any variable, exactly as expected from a prescribed climatological SST-setup, where memory from the ocean is erased. Also, the distributions of and (and their joint distribution) are both consistent with white noise, which we tested via uni-variate (D’Agostino and Pearson (1973)) and multi-variate (Henze and Zirkler (1990)) normality tests (see § 2.4).
However, the feedback time series computed from and in piClim-control still exhibits substantial interannual variability (Figure 6d), with periods of lower or higher feedback, similar to the piControl feedback time series (Figure 6c). We illustrate this further by looking at the spectral densities of , , and the feedback in piClim-control (Figure 7a): the spectral density of and is rather flat, as expected for white noise time series, but the power spectral density of the feedback time series increases substantially towards lower frequencies. This is noteworthy, as the feedback has been computed from two time series which both exhibit no significant autocorrelation and are consistent with white noise.
Figure 7. Spectral densities of the variables (surface air temperature), (net TOA imbalance) and the feedback , computed via 30-year moving window regressions of the and timeseries (Eq. (6)), (a): for the 200-year long piClim-control simulation, (b): for synthetic timeseries generated via a Monte-Carlo simulation. and were fitted on data from piClim-control as jointly normal with no autocorrelation (see § 2.4); (c): for the 1200-year long coupled piControl simulation, (d): for synthetic time series generated via a Monte-Carlo simulation. and were fitted on data from piControl as a VAR(5) process (see § 2.4).
We want to point out that observing variability in the piClim-control feedback of a magnitude similar to the piControl feedback is not surprising per se: since the feedback is computed as a ratio between and , it is independent of the absolute magnitude of variability in and . What matters for the feedback is only the variability of and relative to one another. In other words, even small variations in and could result in the same feedback variance, if the relative variability of and is the same. In fact, we would even expect a higher variance in the feedback in an unforced (such as piControl) or even fixed-SST setup (such as piClim-control): the absence of forcing implies that over the 30 years used for the computation of there is a weak signal in the denominator (), which, if comes close to zero, inflates the variance of the feedback timeseries. This is indeed reflected in the high variance of the piClim-control feedback (Figure 6d). It also provides a reason why the correlation between coupled and amip feedback is worse for the piControl-case than for the amip-histForcing-histSST case, as the latter has a historical prescribed forcing causing a potentially stronger signal in . What is striking, however, is that the variability spectrum of the piClim-control feedback exhibits long-term variability with substantial memory (as seen from the increasing power spectral density towards lower frequencies in Figure 7a), or, as seen in Figure 6d from pure visual inspection: sustained periods of weaker or stronger feedback despite there being no memory in the generating time series and .
What we observe in Figure 6d–Figure 6f and Figure 7a—namely, two white noise time series generating low-frequency variability after applying a moving-window regression—is an example of what is known as the Yule-Slutzky effect (Slutzky, 1937; Yule, 1926; Gershunov et al., 2001; Kelly and Gráda, 2014): when a moving window is applied over a random noise process, it generates long-term variability which does not stem from actual low-frequency variability of the underlying data. The effect of applying moving window averages on noisy climate data has been documented and studied before (e.g., Gershunov et al., 2001; Kelly and Gráda, 2014; Li et al., 2024), yet moving averages nonetheless remain a common and ubiquitous (but often problematic) practice in climate science.
We perform a Monte-Carlo simulation to show how the Yule-Slutzky effect operates in piClim-control and to confirm that the effect is not a result of the relatively short piClim-control simulation. For this, we simulate a 1000000-year long and time series from a joint normal distribution fitted to the piClim-control and output (see § 2.4). We then compute the resulting feedback time series as in 30-year moving window regressions. The resulting spectra of , , and the feedback are shown in Figure 7b. These confirm the results from Figure 7a: despite the flat white-noise spectra of and , there is substantial variability of the feedback parameter time series increasing towards longer timescales. The feedback spectrum flattens out at a decorrelation timescale of roughly 60–100 years. This means that there are substantial variations in the piClim-control feedback time series up to 60–100 year timescales. Importantly, this multi-decadal variability is purely stochastically generated (in the case of Figure 7b), or results from short-term white-noise-like internal variability in and (in the case of Figure 7a). It therefore does not originate from any physical long-term variability or memory in or .
The idealized piClim-control setup helped to identify a purely statistical source of variability in the feedback time series. However, the real world and the fully coupled piControl simulation are different from the idealized piClim-control simulations, because the freely-evolving ocean-atmosphere system in piControl harbors substantial memory and cannot be modeled as a white noise process. This can be seen in the power spectral densities of , , and the feedback from the 1200-year long coupled piControl simulation (Figure 7c). Here, the spectrum of shows increasing variance up to centennial timescales. This is the expected effect of slow ocean variability leading to strong memory in the system and is typical for time series of climatic variables ([10] (Hasselmann 1976; Frankignoul and Hasselmann 1977; Roe 2009)). The time series of , on the other hand, shows a power spectrum which peaks at a frequency of ∼ 2–5 years, which is expected given this roughly corresponds to the ENSO timescale. The power spectral density then decreases with period length, indicating a process with maximal variability on the shortest timescales and highly stable at long timescales. This spectrum of is also to be expected, and follows directly from a linear coupled system with stochastic forcing and a restoring, negative feedback: at long timescales, the restoring feedback ensures that the system restores equilibrium (), meaning that long periods with higher or lower do not occur internally and hence, on long timescales, is constant ([10] (Bretherton and Battisti 2000)).
The feedback spectrum of the piControl simulation (Figure 7c) shows a similar increase at long frequencies like the one of the piClim-control simulation (Figure 7a). We again verify this effect via a Monte-Carlo simulation. This time, we model the joint evolution of the coupled piControl and time series via a vector autoregressive (VAR) process of order 5, and again simulate one million years of and from it (see § 2.4). The spectra of the simulated and , as well as the spectrum of the resulting feedback, are shown in Figure 7d. The results reproduce the spectrum obtained from the raw piControl simulation. The variance of the feedback calculated from the 30-year moving window regressions peaks at periods of 60–100 years and longer. This means that also in the coupled simulation, any trends in the feedback detected on timescales shorter than years could be owed to statistical artifacts, making their attribution to changing physical mechanisms (such as the evolving SST pattern) extremely difficult.
That being said, the Monte-Carlo simulation using the VAR(5) process fitted on coupled piControl data does not say anything about the origin of the statistically found (auto-)correlations in and . It only demonstrates that the Yule-Slutzky effect could be generating substantial variability also in the non-white-noise setting of the coupled simulation, accounting for some of the feedback variance therein. The results in Figure 7c–Figure 7d do not, however, exclude that the specific autocorrelation structure of and found via the VAR(5) process arises from specific, persistent or evolving, patterns of SSTs, which then also drive and fluctuations. In other words, our Monte-Carlo simulation is unable to attribute the amount of variance found in the feedback to a specific cause: pure statistical Yule-Slutzky effect, versus a physically meaningful evolution of SST patterns, for instance.
Is there an unforced SST pattern effect in piControl?
In order to investigate whether some of the variations in the coupled piControl feedback arise from evolving SST patterns, we investigate periods of strengthening and weakening of the feedback time series in the 4000-year long, piControl-long simulation (see § 2.7). If the SST pattern was driving these fluctuations, we would expect that SST patterns share certain similarities during periods of feedback strengthening or weakening. For instance, a strong East-West Pacific SST gradient has been found to induce more negative feedbacks due to direct effects on lapse rate, inversion strength and clouds (e.g., [e.g. 2] (Ceppi and Gregory 2017; Dong et al. 2019)). We therefore identify periods of at least 20 consecutive years in the piControl-long run, during which the feedback is strengthening or weakening (the first 1200 years of piControl-long are shown in Figure 8a). The resulting composite maps of surface temperature trends during periods of feedback strengthening and weakening are displayed in Figure 8b and Figure 8c.
Figure 8. Surface air temperature trend composites in the fully coupled piControl run during periods of feedback strengthening and weakening. (a): the raw feedback time series (thin grey line), together with a smoothed feedback time series (thick black line), smoothed using a Savitzky-Golay filter with window length 81 and polynomial order 4. Blue and red shaded regions indicate consecutive periods of length years, during which the derivative of the smoothed feedback time series is negative or positive, respectively. A positive derivative means that the feedback is weakening, a negative derivative, that the feedback is strengthening. Panels (b) and (c) show the surface air temperature trend composites computed over the blue (b) and red (c) shaded regions in (a). Panels (d) and (e) show the fraction of feedback strengthening (d) or feedback weakening (e) periods which agree on the sign of the surface air temperature trend in a given location. Black hatching indicates regions where the composite trend (b–c) or agreement (d–e) is significant at the 95% confidence level. Turquoise areas in (b)–(e) indicate regions significant at the 95% confidence level after accounting for multiple testing (see § 2.7).
The surface temperature trend composite maps do not exhibit the patterns that would be expected from literature: the tropics and the tropical Pacific in particular do not stand out with strong trends, although they are believed to be a main driver of feedback variations (e.g., [e.g. 2] (Zhou et al. 2016; Ceppi and Gregory 2017; Dong et al. 2019)). In fact, only very few of the trends in the composite maps are significant at the 5% level (black hatching), and they are not in the tropics. In particular, there is no significant trend in the Pacific SST gradient. This leads to the question, are any of these trends physically meaningful, and how can our results be reconciled with previous literature?
First, we find that when accounting for multiple testing [11], none of the computed trends are significant at the 5% level apart from a small region in the South Pacific during periods of feedback strengthening (turquoise blob in Figure 8b). This implies that we are unable to identify a characteristic trend pattern in surface air temperature emerging during periods of feedback strengthening or weakening. We confirm this finding by looking at the agreement on the sign of the trends among periods of weakening or strengthening. We find that the vast majority show very weak agreement (Figure 8d, Figure 8e): the agreement fraction rests below 0.6 almost everywhere, for both the feedback weakening and feedback strengthening periods. This means that in roughly half of the red or blue periods in Figure 8a, the trend in a region is positive, and in the other half of the periods, it is negative. In other words, the individual periods making up the composite do not agree on the sign of the trend, almost anywhere on the globe. In fact, also for the agreement fraction, we find that almost none of the hatched regions are indeed significant if the multiple testing problem is accounted for: the areas which show an agreement on the sign of the SST trend which is significantly different from random noise at the 95% level amount to only 6% of the total number of grid cells for periods of feedback strengthening, for instance—close to the amount expected from a multiple testing error at the 95% significance level.
This finding is particularly important when considering the discrepancy between observed and modeled SST patterns in the Pacific (e.g., (Olonscheck et al., 2020; Wills et al., 2022)). The prevailing consensus in literature is that the strengthening SST gradient in observations is the cause of the stabilizing feedback in recent decades found via amip-piForcing simulations ([5, 7] (Zhou et al. 2017; Dong et al. 2021)). However, our results suggest that—at least in the unforced piControl setting in this climate model—trends in a specific region of the globe cannot be uniquely mapped to a specific feedback evolution. This could either be because (i) the variance in the feedback computed from 30-year moving window regressions is to a large degree statistical noise from the Yule-Slutzky effect, or (ii) it could be because the co-evolution of the SST field over the entire globe is never pronounced enough to lead to detectable changes in the feedback, or (iii) in an unforced setting it is hard to detect the effect of the SST pattern due to less-pronounced signals in and . In other words, even if a certain region exhibits feedbacks which are known to be stabilizing—such as the recent SST evolution in the tropical Pacific—, this effect could be compensated by opposite effects in other areas of the globe. This would make sense from a statistical point of view, given the multidimensionality of the spatial field and the multitude of feedbacks acting everywhere simultaneously. Therefore, unless the SST pattern in observations is really exceptionally different from what would be expected ‘naturally’, and acts in a direction which affects the feedback positively or negatively in several regions at the same time, our results suggest that there is not enough evidence to claim an effect of the observed SST pattern on the evolution of the observed feedback time series.
Discussion
Some of our results may at first appear contradictory to previous literature. Numerous studies show that there are key regions on the globe, which drive important cloud, lapse-rate and water vapour feedbacks, particularly the tropical and subtropical Pacific Ocean [2, 5] (Zhou et al. 2016; Ceppi and Gregory 2017; Dong et al. 2019). In addition, it has been shown that in abrupt -increase experiments, the shift of the average pattern of warming leads to a change in the feedback towards less stabilizing values [e.g. 4] (Senior and Mitchell 2000; Rugenstein et al. 2016; Dong et al. 2021). Why do we then not see any of these patterns emerging in the variability of the feedback in the fully coupled piControl simulation? We attempt a systematic reconciliation, which is mainly related to the signal-to-noise ratio of the experimental setups.
Transient feedback estimates from amip-piForcing runs are likely not robust
Our results suggest that the findings of numerous studies using 30-year moving window regressions to compute the time-evolution of the global feedback based on amip-piForcing runs [e.g. 5, 7] (Gregory and Andrews 2016; Zhou et al. 2016; Dong et al. 2021) may not be robust, for two reasons: first, the lack of prescribed historical atmospheric forcing may induce a large bias and poor correlation with respect to the true feedback evolution. Second, the recent period of the past hundred years over which the feedback evolution is evaluated is too short to separate any changes in the feedback from the timescale of statistical noise induced by the Yule-Slutsky effect. Our results suggest that the stabilization of the historical feedback found in these experiments cannot be robustly attributed to changing SST patterns, and that this stabilization may in fact be purely statistical noise. This problem is further aggravated by the fact that amip-piForcing runs with observed SSTs and SIC prescribe model-foreign SST boundary conditions to the atmospheric model. This in itself might introduce a bias of unknown magnitude whose quantification is beyond the scope of this paper. The fact that even in our controlled setup—where model-internal SSTs and SIC are prescribed to the same atmospheric model—the feedback of the amip setup does not match that of the coupled setup, raises serious concerns about any conclusions on the feedback inferred from amip-piForcing with observed SSTs, which (due to structural model biases) will differ from those "preferred" by (i.e., internally consistent with) the atmospheric model.
The signal-to-noise problem in quantifying a transient, time-evolving pattern effect
Our findings do not challenge fundamental theoretical arguments and model results on how specific SST patterns in specific locations affect cloud, lapse-rate and water vapour feedbacks [e.g. 4, 2] (Senior and Mitchell 2000; Rugenstein et al. 2016). Rather, our results point to a signal-to-noise problem: In studies where a strong, persistent, idealized perturbation is imposed over a sustained period—such as abrupt--quadrupling experiments [e.g. 4], AMIP-4K-type experiments [e.g. 2], or idealized prescribed heat-flux experiments (e.g., Rugenstein et al., 2016)—there is a much higher signal-to-noise ratio. In these cases, one considers either a long-term mean response (such as the -quadrupling), or a response to a fixed SST pattern which is not allowed to (co)vary in time.
In both cases, there is much less noise generated, because there is no need to compute moving-window regressions in order to estimate a transient feedback response. The feedback can be directly estimated via the Gregory-method. These sustained SST patterns then may very well impact the feedback, as has been shown. The fact that we do not find said relationships in the transient piControl setup is likely due to the covariance of the entire SST field, where compensating effects may occur. In other words, the signal-to-noise ratio in an unforced piControl setup is too low to detect the expected physical mechanisms (or, the statistical noise in the transient moving-window calculation is too big).
Also related to the signal-to-noise problem are studies based on Green’s function: here, individual patches are systematically subjected to SST perturbations and their effect is therefore spatially isolated, assuming linearity of the total response with respect to the individual geographic responses (Zhou et al., 2017; Dong et al., 2019). Green’s functions can very well be used to identify the effects of individual warming patterns on the global feedback. However, unless the entire spatially-varying field is convolved with the complete Green’s function, it is impossible to quantify the effect of the global observed SST pattern on the feedback. Since such a convolution would again lead to noisy timeseries of and and the subsequent need to compute a transient feedback via 30-year moving window regressions, Green’s function approaches will likely also lead to strong statistical noise. Our results do not necessarily challenge the physically-consistent results form the Green’s functions. Rather, it is possible that the spatial field of SSTs in an unforced piControl simulation covaries in ways such that strong global fluctuations in the feedback are largely balanced out. It is therefore possible that a warming or cooling in a specific region in the Pacific does have the physically expected effect, however this effect is globally compensated by other radiative effects accumulating from elsewhere on the globe. In other words, the (univariate) signal from one specific region does not say anything about the simultaneous other signals coming from the multivariate pattern of surface warming.
In summary, a persistent, strong and externally prescribed SST anomaly will produce a radiative response that is sufficient to infer the magnitude of the feedback parameter, consistent with decades of AMIP type experiments successfully used in model development and tuning (e.g., of cloud parameters) and Greens-function or pacemaker type experiments. Inferring changes in the feedback parameter from internally produced patterns, however, is hard or impossible.
Conclusions and Outlook
Our work sheds light on two important limitations of using the previously established amip-piForcing method to estimate the variability of Earth’s feedback from observations. First, we have shown that the classical amip-piForcing setup fails to capture the evolution of the true coupled system feedback, with a very low correlation between the two (Pearson corr. = 0.23). We also show that the amip-piForcing runs exhibit a systematic negative bias in the feedback, in particular during the period 1940–2000. We next identify a major driver of this bias, namely the omission of prescribing atmospheric forcing (and in particular, aerosol forcing) in the amip-piForcing setup. We then show that when prescribing the correct atmospheric forcing, the agreement between coupled and amip-style feedback improves substantially (Pearson corr. = 0.68). We therefore strongly recommend that any amip-piForcing-type simulations ran to estimate the observed feedback parameter evolution be changed to amip-histForcing-type runs, where the correct atmospheric forcing is also prescribed, as in the classical AMIP protocol (Eyring et al., 2016). However, individual members still fail to reproduce the coupled feedback evolution, and there can be substantial differences in the sign of the feedback trends, despite identical SSTs, SIC, and correctly prescribed atmospheric forcing (Figure 5). In particular, the average ensemble bias towards an overall more stabilizing feedback in the amip-style simulations remains, albeit reduced, even if the correct atmospheric forcing is prescribed. This result might imply that the true Earth feedback is in reality less stabilizing than inferred from amip-style simulations. Our findings thus undermine the conclusions made on the basis of amip-piForcing runs. Prescribing observed SSTs, SIC, and forcing to an atmospheric model appears to still be insufficient to constrain the coupled feedback evolution.
Our results are particularly worrying when considering that the experimental setup we provide is using model-internal SSTs and SIC as boundary conditions. In this very idealized case we would expect the amip-setup—especially with prescribed atmospheric forcing as in the coupled setup—to reproduce the feedback variability of the coupled run, because everything happens in a "perfect model" world. In the amip-piForcing runs with observations, however, SSTs and SIC with completely model-foreign mean-state patterns are fed to the model, trying to force the model to produce a pattern that it potentially internally would never do. The effect of this foreign SST and SIC prescription is not taken into account in our idealized setup and might bias the estimate even further. Our results are based on a single GCM and should therefore be validated with other models before confirming all our findings, but it is hard to imagine a method to produce a robust relationship across many models if it does not work in a single perfect model setup. We provide the experimental setup to do this validation and propose that it is done in order to assess the effectiveness of amip-piForcing runs in future use. However, we point out that precisely because the amip-piForcing setup fails even when prescribing model-internal SSTs and SIC to the same model, the leap of faith when prescribing observations to a model is even larger.
Second, we show that estimating the feedback time series via 30-year moving window regressions introduces substantial statistical noise via the Yule-Slutsky effect. This variability makes the feedback vary from values of -3 to above 0 in piControl simulations. This low-frequency variability in the feedback could be purely statistical, as it occurs also in Monte-Carlo simulations with pure white noise, and also in the simulations with a covariance and autocorrelation structure of and as that in the fully-coupled piControl. The timescale of this statistically-induced variability is 60–100 years. This finding alone implies that trends identified in the feedback timeseries on timescales shorter than 100 years could be purely statistical artifacts. This result helps explain how the amip-style and coupled runs can show differences in the feedback despite identical SSTs, SIC, and atmospheric forcing. Our results are also in line with the findings of Modak and Mauritsen (2023), who show that just changing the observational SST and SIC dataset used for forcing the amip-piForcing simulations leads to substantially different feedback evolutions. We then demonstrate that periods of weakening or strengthening feedback cannot be statistically linked to specific trends in SST patterns in a fully-coupled 4000-year piControl simulation. We find no significant SST trends emerging during periods of feedback strengthening or weakening, and virtually no agreement in SST trends between different periods of similar feedback change. We thus find no evidence of an unforced SST pattern effect in the CESM2 piControl simulation.
Based on these results, we believe that trends detected based on amip-piForcing runs over the past decades a) can not reproduce the corresponding feedback in the coupled system correctly, b) can not be causally linked to the evolving SST pattern, and c) may be entirely unphysical and a result of the amplification of statistical noise via the Yule-Slutzky effect. Since the statistically-induced variations peak at a timescale of 60–100 years, any variations detected in the feedback time series on periods shorter than this may be purely statistical and their attribution to SSTs, at least to our current knowledge and based on standard methods, is impossible. The problem is further complicated by the fact that signal-to-noise is comparatively weak in the recent period, as the real world is still far away from an idealized, strong forcing setup like a CO-quadrupling. While the substantial differences between the CMIP6-simulated and the observed Pacific SST trends are undoubtedly real (e.g., (Wills et al., 2022; Olonscheck et al., 2020)), the argument of a trend towards a more stabilizing, SST-driven feedback in recent decades based on amip-piForcing simulations and found in a multitude of studies (e.g., [5, 7] (Gregory and Andrews 2016; Zhou et al. 2017; Dong et al. 2021; Zhou et al. 2021)) might be a) an effect of the misrepresentation of the coupled feedback in the atmosphere-only setup, b) an effect of the lack of prescribing atmospheric forcing, or c) a statistical effect. Based on the broad spectrum of strong limitations of both the current amip-piForcing setup, but also on the feedback estimation via 30-year moving window regressions, we believe that this method provides little potential to constrain changes in the Earth’s feedback over the historical period.
Code and data availability. Climate model simulations from CMIP6 used to generate the boundary conditions for our simulations are available at https://esgf-node.llnl.gov/projects/cmip6/ and were accessed through the ETH Zurich Next-Generation Archive Brunner et al. (2020). We acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modelling, coordinated and promoted CMIP6. We thank the climate modeling groups for producing and making available their model output, the Earth System Grid Federation (ESGF) for archiving the data and providing access, and the multiple funding agencies who support CMIP6 and ESGF. Climate model simulations from the CESM2 single-forcing large ensemble used to generate the boundary conditions for our simulations are available via https://gdex.ucar.edu/datasets/d651055/dataaccess/.
The code used for reproducing the analysis and all figures in the study will be made available at the time of publication at https://doi.org/10.5281/zenodo.21937291 ((Gyuleva et al., 2026)).
Methods for estimating the forcing
Estimating the feedback parameter requires estimates of the forcing in simulations with prescribed time-varying atmospheric forcing. We employ three different methods of estimating the forcing in order to test the sensitivity of our results to the forcing estimate: the RFMIP method ((Pincus et al., 2016)), the RFMIP method with a correction proposed by Hansen et al. (2005), and the AerChemMIP method ((Collins et al., 2017)). The three different estimates for the historical forcing are shown in
The RFMIP method for calculating the time-varying time series involves running fixed-SST experiments with time-varying forcing: the time-varying radiative forcing is estimated as the difference in net TOA imbalance between an atmosphere-only run with fixed climatological SSTs but time-varying atmospheric forcing, and a control run with climatological SSTs and fixed pre-industrial atmospheric forcing (piClim-control). We perform a 75-year long piClim-control simulation with monthly climatological SSTs taken from the coupled CESM2 piControl simulation. For the scope of the forcing estimation 30 years would be sufficient (Forster et al., 2016), but we use the piClim-control also in our analysis of feedbacks (§ 3.2). We then also perform the piClim-histall and piClim-ghg simulations with the identical climatological SSTs as the piClim-control, but time-varying atmospheric forcing (historical for the piClim-histall and only historical GHG for the piClim-histghg). We thereby obtain forcing time series for evolving historical and historical GHG-only forcing (orange line in Figure A1). These forcing time series then correspond to our estimates of the forcing for the coupled historical runs and the GHG2 runs, from which SST and SIC boundary conditions were obtained.
The RFMIP method has the advantage of being computationally relatively simple and requiring only roughly 30 years of piClim-control simulation, in addition to the time-varying experiment. In addition, due to the climatological prescribed SSTs, internal variability is largely suppressed, leading to a more accurate estimate (Forster et al., 2016). However, the RFMIP method also exhibits two potential caveats. First, as only ocean temperatures are held fixed but the land is allowed to respond to the prescribed atmospheric forcing, part of the net TOA imbalance in the piClim-histall and piClim-histghg experiments is a response of the land to the forcing, rather than the forcing itself. The magnitude of this effect has only been quantified for forcing estimates of abrupt CO quadrupling runs, where it has been shown to lead to an underestimation of the forcing by up to 14% (although this result is based on a single GCM) [6].
One way to correct for this approach is to employ the method proposed by Hansen et al. (2005) (from now on denoted as H05). The net TOA response is corrected by an estimate of the response of the land given by :
Since is negative and the change in land temperature is positive in runs with historical or historical GHG-only atmospheric forcing, the H05 correction results in an overall upward correction of the forcing. Andrews et al. [6] find that this correction still underestimates the true estimate, yet their results are based on abrupt 4xCO forcing, and for just one model. Due to the simplicity of the correction and its low computational cost, we decide to use the H05 correction to test the sensitivity of our results, despite its possibly limited accuracy (green line in Figure A1).
An alternative method to diagnose in transient runs is employed as part of the Aerosols and Chemistry Model Intercomparison Project (AerChemMIP) (Collins et al., 2017) and was first proposed by Andrews et al. [3]. In this method, is diagnosed as the difference between an atmospheric simulation with time-evolving prescribed SSTs, SIC, and atmospheric forcing (such as a standard AMIP simulation Eyring et al. (2016), or the amip-histForcing-histSST simulations), and a simulation with identical time-evolving SSTs and SIC, but no atmospheric forcing (such as amip-piForcing-style runs). The benefit of this approach is that potential interactions of the forcing agents with the transient SST and SIC change are accounted for (Forster et al., 2016; Collins et al., 2017). However, the noise in the transient estimate due to variability is larger, calling for the need of an ensemble of simulations to derive a reliable estimate Forster et al. (2016). Here, we have such an ensemble: we compute the time series as the difference between the amip-histForcing-histSST and the corresponding amip-piForcing-histSST runs with pairwise identical SSTs. The resulting transient estimate, denoted as is then taken as the ensemble mean of the eleven estimates from the ensemble (blue line in Figure A1).
Robustness of spatial trends
Perios of feedback weakening and strengthening in -long
Author contributions. GG devised the experiments, analyzed the data and wrote the manuscript. RN, RK, and SS developed ideas for the analysis. RK and SS provided supervision. RN, EF, RN, and SS contributed to the manuscript. UB ran the experiments and generated the data.
Competing interests. The authors declare that they have no conflict of interest.
Acknowledgements
Acknowledgements. GG thanks Patricia Helpap, Florian Römer, Iris de Vries, and Konstantin Weber for the insightful discussions. GG, SS and RK acknowledge the project ‘Constraints on near-term warming projections via distributionally robust statistical and machine learning’ (COPE; grant agreement C22-02, funded by the Swiss Data Science Center). SS acknowledges funding provided by the German Research Foundation through the Heinz Maier-Leibnitz Prize 2024 and the European Union’s Horizon Europe programme via the project ‘Artificial Intelligence for Enhanced Representation of Processes and Extremes in Earth System Models’ (AI4PEX; grant agreement 101137682). GG and RK acknowledge funding by the European Union’s Horizon 2020 research and innovation programme under grant agreement No. 101003536 (ESM2025–Earth System Models for the Future). RN has received funding from the Swiss National Science Foundation (SNSF) via grant no. 216710.
References
References
- [1]Alexander, M. A., Bladé, I., Newman, M., Lanzante, J. R., Lau, N.-C., and Scott, J. D.: The Atmospheric Bridge: The Influence of ENSO Teleconnections on Air–Sea Interaction over the Global Oceans, Journal of Climate, 15, 2205–2231, https://doi.org/10.1175/1520-0442(2002)015<2205:TABTIO>2.0.CO;2, 2002.
- [2]Andrews, T. and Webb, M. J.: The dependence of global cloud and lapse rate feedbacks on the spatial structure of tropical Pacific warming, Journal of Climate, 31, 641–654, 2018.DOI
- [3]Andrews, T., Gregory, J., and Webb, M.: Using an AGCM to Diagnose Historical Effective Radiative Forcing and Mechanisms of Recent Decadal Climate Change, Journal of Climate, 27, 1193–1209, https://doi.org/10.1175/JCLI-D-13-00336.1, 2014.
- [4]Andrews, T., Gregory, J. M., and Webb, M. J.: The Dependence of Radiative Forcing and Feedback on Evolving Patterns of Surface Temperature Change in Climate Models, Journal of Climate, 28, 1630–1648, https://doi.org/10.1175/JCLI-D-14-00545.1, 2015.
- [5]Andrews, T., Gregory, J. M., Paynter, D., Silvers, L. G., Zhou, C., Mauritsen, T., Webb, M. J., Armour, K. C., Forster, P. M., and Titchner, H.: Accounting for changing temperature patterns increases historical estimates of climate sensitivity, Geophysical Research Letters, 45, 8490–8499, 2018.DOI
- [6]Andrews, T., Smith, C. J., Myhre, G., Forster, P. M., Chadwick, R., and Ackerley, D.: Effective Radiative Forcing in a GCM With Fixed Surface Temperatures, Journal of Geophysical Research: Atmospheres, 126, e2020JD033880, https://doi.org/10.1029/2020JD033880, _eprint: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/2020JD033880, 2021.
- [7]Andrews, T., Bodas-Salcedo, A., Gregory, J. M., Dong, Y., Armour, K. C., Paynter, D., Lin, P., Modak, A., Mauritsen, T., Cole, J. N. S., Medeiros, B., Benedict, J. J., Douville, H., Roehrig, R., Koshiro, T., Kawai, H., Ogura, T., Dufresne, J.-L., Allan, R. P., and Liu, C.: On the Effect of Historical SST Patterns on Radiative Feedback, Journal of Geophysical Research: Atmospheres, 127, e2022JD036675, https://doi.org/10.1029/2022JD036675, _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1029/2022JD036675, 2022.
- [8]Armour, K. C., Bitz, C. M., and Roe, G. H.: Time-varying climate sensitivity from regional feedbacks, Journal of Climate, 26, 4518–4534, 2013.
- [9]Armour, K. C., Proistosescu, C., Dong, Y., Hahn, L. C., Blanchard-Wrigglesworth, E., Pauling, A. G., Inglin Wills, R. C., Andrews, T., Stuecker, M. F., Po-Chedley, S., Mitevski, I., Forster, P. M., and Gregory, J. M.: Sea-surface temperature pattern effects have slowed global warming and biased warming-based constraints on climate sensitivity, Proceedings of the National Academy of Sciences, 121, e2312093121, https://doi.org/10.1073/pnas.2312093121, 2024.
- [10]Barsugli, J. J. and Battisti, D. S.: The Basic Effects of Atmosphere–Ocean Thermal Coupling on Midlatitude Variability, Journal of the Atmospheric Sciences, 55, 477–493, https://doi.org/10.1175/1520-0469(1998)055<0477:TBEOAO>2.0.CO;2, 1998.
- [11]Benjamini, Y. and Hochberg, Y.: Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing, Journal of the Royal Statistical Society. Series B (Methodological), 57, 289–300, https://www.jstor.org/stable/2346101, 1995.DOI
- [12]Boer, G. J., Arpe, K., Blackburn, M., Déqué, M., Gates, W. L., Hart, T. L., le Treut, H., Roeckner, E., Sheinin, D. A., Simmonds, I., Smith, R. N. B., Tokioka, T., Wetherald, R. T., and Williamson, D.: Some results from an intercomparison of the climates simulated by 14 atmospheric general circulation models, Journal of Geophysical Research: Atmospheres, 97, 12 771–12 786, https://doi.org/10.1029/92JD00722, _eprint: https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/92JD00722, 1992.
- [13]Boer, G. J., Hamilton, K., and Zhu, W.: Climate sensitivity and climate change under strong forcing, Climate Dynamics, 24, 685–700, https://doi.org/10.1007/s00382-004-0500-3, 2005.












