THIS POST IS A CRITICAL REVIEW OF AN UNPUBLISHED MANUSCRIPT ON ECS UNCERTAINTY BY STEVEN SHERWOOD AND HIS LARGE TEAM OF CLIMATE SCIENTISTS AT THE UNIVERSITY OF NEW SOUTH WALES. THE FULL TEXT OF THE UNPUBLISHED MANUSCRIPT IS AVAILABLE FOR DOWNLOAD IN PDF FORMAT  [LINK] . THE WORK IS BASED ON WAYS OF REDUCING ECS UNCERTAINTY SUGGESTED BY BJORN STEVENS OF THE MAX PLANCK INSTITUTE OF METEOROLOGY IN GERMANY. 

STEVEN SHERWOOD           BJORN STEVENS

PART-1: WHAT THE MANUSCRIPT SAYS

1. ABSTRACT: We assess evidence relevant to Earth’s equilibrium climate sensitivity per doubling of atmospheric CO2, characterized by an effective sensitivity=S. This evidence includes feedback process understanding, the historical climate record, and the paleoclimate record. An S value lower than 2K is difficult to reconcile with any of the three lines of evidence. The amount of cooling during the Last Glacial Maximum provides strong evidence against values of S greater than 4.5 K. Other lines of evidence in combination also show that this is relatively unlikely. We use a Bayesian approach to produce a probability density function for S given all the evidence, including tests of robustness to difficult-to-quantify uncertainties and different priors. The 66% range is 2.6-3.9 K for our
   Baseline calculation, and remains within 2.3-4.5 K under the robustness tests;
   corresponding 5-95% ranges are 2.3-4.7 K. This indicates a stronger constraint on S than reported in past assessments, by lifting the low end of the range. This
   narrowing occurs because the three lines of evidence agree and are judged to be largely independent, and because of greater confidence in understanding feedback processes and in combining evidence. We identify promising avenues for further narrowing the range in S, in particular using comprehensive models and process understanding to address limitations in the traditional forcing-feedback paradigm for interpreting past changes.
2. PLAIN LANGUAGE SUMMARY:  Earth’s global climate sensitivity is a fundamental quantitative measure of the susceptibility of Earth’s climate to human influence. A landmark report in 1979 (Jules Charney) concluded that it probably lies
   between 1.5-4.5℃ per doubling of atmospheric carbon dioxide, assuming that other influences on climate remain unchanged. In the 40 years since, it has appeared difficult to reduce this uncertainty range. In this report we thoroughly assess all lines of evidence including some new developments. We find that a large volume of consistent evidence now points to a more confident view of a climate sensitivity near the middle or upper part of this range. In particular, it now appears extremely unlikely that the climate sensitivity could be low enough to avoid substantial climate change well in excess of 2℃ warming under a high-emissions future scenario. We remain unable to rule out that the sensitivity could be above 4.5℃ per doubling of carbon dioxide levels although this is not likely. Continued research is needed to further reduce the uncertainty and we identify some of the more promising possibilities in this regard.
3. INTRODUCTION: The ECS, defined as the steady-state global temperature increase for a doubling of CO2, has long been taken as the starting point for understanding global climate changes. It was quantified specifically by Charney in 1979 as the equilibrium warming as seen in a model with ice sheets and vegetation fixed at present-day values and with a proposed range of 1.5-4.5 K based on the information at the time, but did not attempt to quantify the probability that the sensitivity was inside or outside this range. The IPCC 2013 report asserted the same now-familiar range, but more precisely dubbed it a >66% likely credible interval, implying an up to one in three chance of being outside that range. It has been estimated that, in an ideal world where the information would lead to optimal policy responses, halving the uncertainty in a measure of climate sensitivity would lead to an average savings of US$10 trillion in today’s dollars. Apart from this, the sensitivity of the world’s climate to external influence is a key piece of knowledge that humanity should have at its fingertips. So how can we narrow this range? Quantifying ECS is challenging because the available evidence consists of diverse strands, none of which is conclusive by itself. This requires that the strands be combined in some way. Yet, because the underlying science spans many disciplines within the Earth Sciences, individual scientists generally only fully understand one or a few of the strands. Moreover, the interpretation of each strand requires structural assumptions that cannot be proven, and sometimes ECS measures have been estimated from each strand that are not fully equivalent. This complexity and uncertainty thwarts rigorous, definitive calculations and gives expert judgment and assumptions a potentially large role. Our assessment was undertaken under the auspices of the World Climate Research Programme’s Grand Science Challenge on Clouds, Circulation and Climate Sensitivity  {2015workshop at Ringberg Castle in Germany}. It tackles the above issues, addressing three questions: (1) Given all the information we now have, acknowledging and respecting the uncertainties, how likely are very high or very low climate sensitivities outside the presently accepted likely range of 1.5-4.5 K: (2) What is the strongest evidence against very high or very low values?: (3) Where is there potential to reduce the uncertainty?  In addressing these questions, we follow Stevens et al. (2016, hereafter SSBW16) who laid out a strategy for combining lines of evidence and transparently considering uncertainties. The lines of evidence we consider, as in SSBW16, are modern observations and models of system variability and feedback processes; the rate and trajectory of historical warming, and the paleoclimate record. The core of the combination strategy is to lay out all the circumstances that would have to hold for the climate sensitivity to be very low or high given all the evidence (which SSBW16 call “storylines”). A formal assessment enables quantitative probability statements given all evidence and a prior distribution, but the “storyline” approach allows readers to draw their own conclusions about how likely the storylines are, and points naturally to areas with
   226 greatest potential for further progress. Recognizing that expert judgment is unavoidable, we attempt to incorporate it in a transparent and consistent way. Combining multiple lines of evidence will increase our confidence and tighten the range of likely ECS if the lines of evidence are broadly consistent. If uncertainty is underestimated in any individual line of evidence, inappropriately ruling out or discounting part of the ECS range—this will make an important difference to the final outcome (see example in Knutti et al., 2017). {Blogger’s note: There are two citations for Knutti etal 2017 in the list of citations}. Therefore it is vital to seek a comprehensive estimate of the uncertainty of each line of evidence that accounts for the risk of unexpected errors or influences on the evidence. This must ultimately be done subjectively. We will therefore explore the uncertainty via sensitivity tests and by considering ‘what if’ cases in the sense of Bjorn Stevens, including what happens if an entire line of evidence is dismissed. The most recent reviews (Collins et al., 2013, Knutti et al., 2017 (which one?) have considered the same three main lines of evidence considered here, and have noted they are broadly consistent with one another, but did not attempt a formal quantification of the probability distribution function of ECS. Formal Bayesian quantifications have been done based on the historical warming record (see Bodman and Jones 2016 for a recent review), the paleoclimate record (PALAEOSENS, 2012), a combination of historical and last millennium records (Hegerl et al., 2006), and multiple lines of evidence from instrumental and paleo records (Annan and Hargreaves, 2006). An assessment based only on a subset of the evidence will yield too wide a range if the excluded evidence is consistent (e.g. Annan and Hargreaves, 2006), but if both subsets rely on similar information or assumptions, this co-dependence must be considered when combining them (Knutti and Hegerl 2008). Therefore, an important aspect of our assessment is to explicitly assess how uncertainties could affect more than one line of evidence and to assess the sensitivity of calculated PDFs to reasonable allowance for interdependencies of the evidence {blogger’s note: i.e. violation of the independence assumption}. Another key aspect of our assessment is that we explicitly consider process understanding via modern observations and process models as a newly robust line of evidence. Such knowledge has occasionally been incorporated implicitly (via the prior on ECS) based on the sample distribution of ECS in available climate models (Annan and Hargreaves, 2006) or expert judgments (Forest et al., 2002), but climate models and expert judgments do not fully represent existing knowledge or uncertainty relevant to climate feedbacks, nor are they fully independent of other evidence (in particular that from the historical temperature record, see Kiehl, 2007). Process understanding has recently blossomed, however, to the point where substantial statements can be made without simply relying on climate model representations of feedback processes, creating a new opportunity exploited here. Climate models (GCMs) nonetheless play an increasing role in calculating what our observational data would look like under various hypothetical ECS values in effect translating from evidence to ECS. Their use in this role is now challenging long held assumptions, for example showing that 20th-century warming could have been relatively weak even if ECS were high, that paleoclimate changes are strongly affected by factors other than CO2, and that climate may become more sensitive to greenhouse gases in warmer states. GCMs are also crucial for confirming how modern observations of feedback processes are related to ECS. Accordingly, another novel feature of this assessment will be to use GCMs to refine our expectations of what observations should accompany any given value of ECS and thereby avoid biases now evident in some estimates of ECS based on the historical record using simple energy budget or energy balance model arguments. GCMs are also275 used to link global feedback strengths to observable phenomena. However, for reasons noted above, we avoid relying on GCMs to tell us what values to expect for key feedbacks except where 277 the feedback mechanisms can be calibrated against other evidence. Since we use GCMs in some way to help interpret all lines of evidence, we must be mindful that any errors in doing this could reinforce across lines. We emphasize that this assessment begins with the evidence on which previous studies were based, including new evidence not used previously, and aims to comprehensively synthesize the implications for climate sensitivity both by drawing on key literature and by doing new calculations. In doing this, we will identify structural uncertainties that have caused previous studies to report different ranges of ECS from (essentially) the same evidence, and account for this when assessing what that underlying evidence can tell us. An issue with past studies is that different or vague definitions of ECS may have led to perceived, un-physical discrepancies in estimates of ECS that hampered abilities to constrain its range and progress understanding. Bringing all the evidence to bear in a consistent way requires using a specific measure of ECS, so that all lines of evidence are linked to the same underlying quantity. We denote this quantity as S. The implications for S of the three strands of evidence are examined separately in sections 3-5, and anticipated dependencies between them are discussed in section 6. To obtain a quantitative probability distribution function of S, we follow Bjorn Stevens and many other studies by adopting a Bayesian formalism, which is outlined in sections 2.2-2.6. The results of applying this to the evidence are presented in section 7, along with the implications of our results for other measures of climate sensitivity and for future warming. The overall conclusions of our assessment are presented in section 8.
4. SECTION 8: PREAMBLE TO CONCLUSIONS:  There are subjective elements in this study but there are also objective ones, in particular, enforcing mathematical rules of probability to ensure that our beliefs about climate sensitivity are internally consistent and consistent with our beliefs about the individual pieces of evidence. All observational evidence must be interpreted using some type of model that relates underlying quantities to the data, hence there is no such thing as a purely observational estimate of climate sensitivity. Uncertainty associated with any evidence therefore comes from three sources: observational uncertainty, potential model error, and unknown influences on the evidence such as unpredictable variability. By comparing past studies that used different models for interpreting similar evidence we find that the additional uncertainty associated with the model itself is considerable compared with the stated uncertainties typically obtained in such studies assuming one particular model. When numerical Global Climate Models (GCMs) {blogger’s note: The acronym GCM stands for General Circulation Model} are used to interpret evidence, they reveal deficiencies in the much simpler models used traditionally—in particular the failure of these models to adequately account for the effects of non-homogeneous warming. This insight is particularly important for the historical temperature record, which is revealed by GCMs to be compatible with higher climate sensitivities than previously inferred using simple models. In general, many published studies appear to have overestimated the ability of a particular line of evidence to constrain sensitivity, leading to contradictory conclusions. When additional uncertainties are accounted for, single lines of evidence can sometimes offer only relatively weak constraints on the sensitivity. The effective sensitivity S analyzed here is defined based on the behavior during the first 150 years after a step change in forcing, which is chosen for several practical reasons. While our study also addresses other measures of sensitivity (the Transient Climate Response TCR) and long-term equilibrium sensitivity, the calculations of these were not optimal and future studies could apply a methodology similar to that used here to quantify them, or other quantities perhaps more relevant to medium-term warming, more rigorously. After extensively examining the evidence qualitatively and quantitatively we followed a number of past studies and used Bayesian methods to attempt to quantify the
   implications and probability distribution function for S. It must be remembered that every step of this process involves judgments or models, and results will depend on assumptions and assessments of structural uncertainties that are hard to quantify hard to quantify. Thus we emphasize that a solid qualitative understanding of how the evidence stacks up is at least as important as any probabilities we assign. Nonetheless, sensitivity tests suggest that our results are not very sensitive to reasonable assumptions in the statistical approach.
5. SECTION 8: THE CONCLUSIONS:   (1) Each line of evidence considered here—process knowledge, the historical warming record, and the paleoclimate record—accords poorly with values outside the traditional “Charney” range of 1.5-4.5 K for climate sensitivity. (2) But when these lines of evidence are taken together, because of their mutual reinforcement, we find the “outside” possibilities for S to be substantially reduced. Whatever the true value of S is, it must be reconcilable with all pieces of evidence; if any one piece of evidence effectively rules out a particular value of S, that value does not become likely again just because it is consistent with some other, weaker, piece of evidence as long as there are other S values consistent with all the evidence. If on the other hand every value of S appeared inconsistent with at least one piece of evidence, the evidence would need reviewing to look for mistakes. But we do not find this situation. Instead we find that the lines are broadly consistent in the sense that there is plenty of overlap between the ranges of S each supports. This strongly affects our judgment of S: if the true S were 1 K, it would be highly unlikely for each of several lines of evidence to independently point toward values around 3 K. And this statement holds even when each of the individual lines of evidence is thought to be prone to errors. We asked the following question (following Bjorn Stevens): what would it take, in terms of errors or unaccounted-for factors, to reconcile an outside value of S with the totality of the evidence? A very low sensitivity (S ~ 1.5 K or less) would require all of the following: Negative low-cloud feedback. This is not indicated by evidence from satellite or process model studies and would require emergent constraints on GCMs to be wrong. Or, a strong and unanticipated negative feedback from another cloud type such as cirrus, which is possible due to poor understanding of these clouds but is neither credibly suggested by any model, nor by physical principles, nor by observations. Cooling of climate by anthropogenic aerosols over the instrumental period at the extreme weak end of the plausible range (near zero or slight warming) based both on direct estimates and attribution results using warming patterns. Or, that forced ocean surface warming will be much more heterogeneous than expected and cooling by anthropogenic aerosols is from weak to middle of the assessed range.  Warming during the mid-Pliocene Warm Period well below the low end of the range inferred from observations, and cooling during the Last Glacial Maximum also below the range inferred from observations. Or, that S is much more state-dependent than expected in warmer climates and forcing during these periods was higher than estimated. In other words, each of the three lines of evidence strongly discounts the possibility of S around 1.5 K or below: the required negative feedbacks do not appear achievable, the industrial-era global
   warming of nearly 1 K could not be fully accounted for, and large global temperature changes through Earth history would also be inexplicable. A very high sensitivity (S > 4.5 K) would require all of the following to be true: Total cloud feedback stronger than suggested by process-model and satellite studies, Cooling by anthropogenic aerosols near the upper end of the plausible range. Or, that 4184 future feedbacks will be much more positive than they appear from this historical record because the mitigating effect of recent SST patterns on planetary albedo has been at the high end of expectations, Much weaker-than-expected negative forcing from dust and ice sheets during the Last Glacial Maximum Or, a strong asymmetry in feedback state-dependence, significantly less positive feedback in cold climates than in the present, but relatively little difference in warmer paleoclimates). Thus, each of the three lines of evidence also argues against very high S, although not as strongly as they do against low S. This is mainly because of uncertainty in how strongly “pattern effects” may have postponed the warming from historical forcing, which makes it difficult to rule out the possibility of warming accelerating in the future based on what has happened so far. Indeed, we find that the paleoclimate record (in particular, the Last Glacial Maximum) now provides the strongest evidence against very high S, while all lines provide more similar constraints against low S (paleo slightly less than the others). An important question governing the probability of low or high S is whether the lines of evidence are independent, such that multiple chance coincidences would be necessary for each of them to be wrong in the same direction. For the most part, the various elements in low- and high-S scenarios do appear superficially independent. For example, while possible model errors are identified that (if they occurred) could affect historical or paleo evidence, they mostly appear unrelated to each other or to global cloud feedback or model-predicted S. Some key unknowns act in a compensating fashion, i.e., where an unexpected factor would oppositely affect two lines of evidence, effectively cancelling out most of its contributed uncertainty. Even in the one identified possibility where an unknown could affect more than one line of evidence in the same direction, modelling indicates a relatively modest impact on the probability distribution function. The IPCC AR5 concluded that climate sensitivity is likely (≥ 66% probability) in the range 1.5-4.5 K. The probability of S being in this range is 93% in our Baseline calculation, and is no less than 82% in all other “plausible” calculations considered as indicators of reasonable structural uncertainty. Although consistent with IPCC’s “likely” statement, this indicates considerably more confidence than the minimum implied by the statement. We also find asymmetric probabilities outside this range, with negligible probability below 1.5 K but up to an 18% chance of being above 4.5 K . This is consistent with all three lines of evidence arguing against low sensitivity fairly confidently, which strengthens in combination. Given this consensus, we do not see how any reasonable interpretation of the evidence could assign a significant chance to S < 1.5 K. Moreover our plausible sensitivity experiments indicate a less-than 5% chance that S is below 2 K: our Baseline 5-95% range is 2.3-4.7 K and remains within 2.0 and 5.7 K under reasonable structural changes. Since the extreme tails of the probability distribution function of S are more uncertain and possibly sensitive to “unknown unknowns” and mathematical choices, it may be safer to focus on 66% ranges (the minimum for what the IPCC terms “likely”). This range in our Baseline case is 2.6-3.9 K, a span less than half that of AR5’s likely range, and is bounded by 2.3 and 4.5 K in all plausible alternative calculations considered. Although we are more confident in the central part of the distribution, the upper tail is important for quantifying the overall risk associated with climate change and so does need to be considered. We also note that allowing for “surprises” in individual lines of evidence via “fat-tailed” likelihoods had little effect on results, as long as such surprises affect the evidence lines independently. Our S is not the true equilibrium sensitivity ECS, which is expected to be somewhat higher than S due to slowly emerging positive feedback. Values are similar, however, because we define S for a quadrupling of CO2 while ECS is defined for a doubling, which cancels out most of the expected effect of these feedbacks .We find that the 66% ECS range, at 2.6-4.1 K bounded by 2.4 and 4.6 K, is not very different from that of S, though slightly higher. Thus, our constraint on the upper bound of the ‘likely’ range for ECS is close to that of the IPCC AR5 and previous assessments, which formally adopt an equilibrium definition. The constraint on the lower bound of the “likely” range is substantially stronger than that of AR5 regardless of the measure used. The uncertainties in ECS and S assessed here are similar because each is somewhat better constrained than the other by some subset of the evidence. Among the plausible alternate calculations, the one producing the weakest high end constraint on S uses a uniform-S-inducing prior, which shifts the ranges upward to 2.8-4.5 K (66%) and 2.4-5.7 K (90%). Our Baseline calculation assumes feedbacks are independent (or that dependence is unknown), which predicts a non-uniform prior probability distribution function for S; to predict a uniform one requires instead assuming a known, prior dependence structure among the feedbacks. Although lack of consensus on priors remains a leading-order source of spread in possible results, we still find that sensitivity to this is sufficiently modest that strong constraints are possible, especially at the low end of the S range. The main reason for the stronger constraints seen here in contrast to past assessments is that new analysis and understanding has led us to combine lines of evidence in a way the community was not ready to do previously. We also find that the three main lines of evidence are more consistent than would be expected were the true uncertainty to be as large as in previous assessments. While some individual past studies have assigned even narrower ranges, as discussed above, past studies have often been overconfident in assigning uncertainty so not too much weight should be given to any single study. We note that although we did not use GCM “emergent constraint” studies using present-day climate system variables in our base results, our results are nonetheless similar to what those studies suggest in the aggregate. New models run for CMIP6 are showing a broader range of S than previous iterations of CMIP.  Our findings are not sensitive to GCM S distributions since we do not directly rely on them. The highest and lowest CMIP6 S values are much less consistent with evidence analyzed here than those near the middle of the range. Some of the  effects quantified in this paper with the help of GCMs were looked at only with pre-CMIP6 models, and interpretations of evidence might therefore shift in the future upon further analysis of newer models, but we would not expect such shifts to be noteworthy.

PART-2; CRITICAL COMMENTARY

1. The extremely verbose and confused rant of statements about climate sensitivity intermingled with a multiplicity of interpretations and disclaimers along with a duality in the definition of climate sensitivity as ECS and as S, does not provide useful information on the subject.
2. Also a research question stated as what the width of the climate sensitivity confidence interval should be and whether it can be reduced from the width suggested by Charney is inappropriate in an unbiased and objective scientific inquiry. The issue is not the width of the confidence interval or what the probability in the confidence interval should be but only what the mean and variance of the estimate are. Confidence intervals are simply a way of expressing mean and variance. An extreme form of bias is contained in the research question stated as “We identify promising avenues for further narrowing the range in S“.
3. The Equilibrium Climate Sensitivity described by Charney is derived from climate model simulations of CO2 forcing only that refers specifically to the correlation between temperature and the natural logarithm of atmospheric CO2 concentration in the absence of other forcings. However, temperature forecasts are made with a portfolio of forcings that includes but is not restricted to CO2 forcing. Some of the complexity of the presentation appears to derive from the lack of clarity in this distinction.
4. As suggested at the end of the paper, but not used in the analysis that precedes it, the understanding of warming and the forecast of future warming should be based on the complete portfolio of forcings that includes ECS CO2 forcing. The forcings portfolio can then be tested against observed temperatures and evaluated according to the fit as demonstrated in related posts at this site: [LINK] [LINK] .
5. The analysis also displays the oddity in climate science of understanding  variance not as degradation of the information content of the mean but as how extreme the the values are that define the confidence interval such that the low information content of large variances is understood not as uncertainty but as the certainty of how extreme the the values COULD be. Such odd interpretations of variance likely derives from confirmation bias in climate science that looks at confidence intervals not as measures of uncertainty (not knowing) but measures of knowing how extreme it COULD be [LINK] .
6. The extensive research efforts and their interpretation presented in the manuscript appear to be products of inappropriate research questions that derive from a flawed interpretation of a confidence interval and the confirmation bias of the researchers expressed as a research objective not of discovering an unbiased estimate of the mean and variance of climate sensitivity to atmospheric CO2 but of finding ways to reduce the width of the confidence interval from the large interval proposed by Jules Charney.
7. The authors mention the relevance of the TCRE (transient climate response to cumulative emissions) but do not address its many anomalous interpretations. For example, the TCRE shows that cumulative emissions of one teratonne will cause 1.5C of warming within a small uncertainty band. The corresponding increase in atmospheric CO2 implies a climate sensitivity and the corresponding uncertainty in the TCRE implies a climate sensitivity uncertainty and its 95% confidence interval. A study of climate sensitivity and its uncertainty should be able to explain the TCRE.
8. The authors do not do that writing only that “The Transient Climate Response (TCR, or warming at the time of CO2 doubling in an idealized 1% per year increase scenario), has been proposed as a better measure of warming over the near- to
   medium-term; it may be more generally related to peak warming, and better constrained by historical warming, than S. It may also be better at predicting high-latitude warming. But 21st-century global-mean trends under high emissions are better predicted by S than by TCR perhaps because of non-linearities in forcing or response or because TCR estimates are affected by noise. TCR is less directly related to the other lines of evidence than is S” 
9. With this brief and mysterious assessment, the authors dismiss the topic altogether. In fact TCRE is not any of these things and TCRE is not affected by noise. In fact the TCRE coefficient is derived from a near perfect correlation between temperature and cumulative emissions. The authors cite the Knutti 2017 paper in which Reto Knutti and co-authors extol the virtues of the TCRE and propose that the TCRE should replace climate sensitivity as our way of understanding the warming effect of fossil fuel emissions. It is clear from the authors’ language that they either did not study the TCRE sufficiently or chose to dismiss it without a sufficient explanation of why it was dismissed. 
10. In summary, we find that this study derives from biased research questions and a poor understanding of variance as a measure of uncertainty. It does not present a useful analysis of the climate sensitivity issue in climate science specifically having to do with the understanding of climate sensitivity in the context of all forcings and of being able to relate the sensitivity issue to the TCRE.