Roy Spencer on feedback in climate models

Climatologist Roy Spencer posted a very interesting entry this week at the Climate Science blog, entitled “Positive Feedback: Have We Been Fooling Ourselves?“, in which he focuses on the following concern:

The traditional way in which feedbacks have been diagnosed from observational data has very likely misled us about the existence of positive feedbacks in the climate system.

Looking at it from an epistemological view, the concern is: given a particular dataset of certain observables of a natural system, are the correct causal relationships within the system being inferred from the dataset? Spencer proceeds to point out multiple problems in current methodology relating to this concern.

 

First, are the mathematical structures typically being used actually models or are they simulacra? In Rosennean terms, a model creates a synonymy of entailment relationships in the model with causal entailments in the natural system under study, and as a result of this synonymy, both the causal interactions and the resultant behavior of the natural system are captured by the model. By contrast, a simulacrum has no such requirement on synonymy of entailment; instead, a simulacrum merely has to mimic the behavior of the system study. Models provide scientific understanding: we can ask ‘why?’ questions about the natural system  by asking ‘why?’ questions about the models. A simulacra has no similar usefulness, since the underlying entailment structure in a simulacra need not have any relationship to that in the natural system. Further, by inducing variations (within certain limits) in certain values of observables in a model, the model’s change in behavior will reflect the corresponding change in behavior of the natural system under those same changes in conditions. A simulacra offers no such capacity, again since the underlying entailment structure in a simulacra need not have any relationship to that in the natural system, there is no reason to expect synonymous variations in behaviors.

Spencer describes the apparently typical attitude toward modelling climate:

Feedbacks are at the heart of most disagreements over how serious man-induced global warming and climate change will be. To the climate community, a feedback is by definition a RESULT of surface temperature change. For instance, low cloud cover decreasing with surface warming would be a positive feedback on the temperature change by letting more shortwave solar radiation in.

But what never seems to be addressed is the question: What caused the temperature change in the first place? How do we know that the low cloud cover decreased as a response to the surface warming, rather than the other way around?

For awhile, a few people had me convinced that this question doesn’t really matter. After all, cause and effect are all jumbled up in the climate system, so what’s the point of trying to separate them? Just build the climate models, and see if they behave the way we observe in nature, right? [bold added]

 

I find the described attitude utterly remarkable, and I wonder how prevalent and pervasive it is in climate science. If 1) cause and effect are not separated, and 2) the criteria for their correctness or usefulness is instead only whether “they behave the way we observe in nature”, then in what sense are the resultant mathematical structures “models”? Surely, they cannot be models in the aforementioned sense; instead, by definition they must be simulacra. As a consequence, 1) we cannot generally expect to increase scientific understanding by probing these simulacra, and 2)  perturbating values of certain observables in the simulacrum will generate variations in the behavior of the simulacrum which will have no necessary correspondence to a similar perturbation in the natural system — simulacra have essentially no credible predictive value for altered scenarios. This point is obviously critical in determining the credibility of the resultant predictions of any given climate “model” generated using variations  in certain observables.

 

 Second, under what circumstances can one observable replace or serve as a proxy for another observable in a model?  Spencer continues:

Feedbacks from observational data have traditionally been diagnosed by plotting the co-variability between top-of-atmosphere radiation budget changes and surface temperature changes, after the data have been averaged to monthly, seasonal, or annual time scales. The justification for this averaging has always remained a little muddy, but from what I can gather, researchers think that it helps to approach a quasi-equilibrium state in the climate system.

The trouble with this approach, though, is that when we average observational data to seasonal or annual time scales in our attempts to diagnose feedbacks, it turns out that there are a variety of very different physical ways to get the very same statistical relationships. (Be patient with me here, I’ll demonstrate this below).

In particular, ANY non-feedback cloud variations that cause surface temperature to change will, necessarily, look like a positive feedback – even if no feedback exists. And the time averaging that everyone employs actually destroys all evidence that could have indicated to us that we were misinterpreting the data. [bold added]

 

Epistemologically, causal relationships are not in themselves observables in nature. Instead, causal relationships are inferred from relationships between observables. The upshot here is that substitution for one set of observables (e.g., daily temperatures) with another set of synthetically generated observables (e.g., monthly, seasonal or annual averaged temperatures) will almost certainly alter what kinds of relationships between observables are evident, and as a result, which causal relationships can be modeled. In particular, Spencer argues that such substitution destroys the evidence for relationships occurring on shorter timeframes than the averaged data. Indeed, I think there exists a further question here: do the relationships identified when using synthetically generated observables represent actual causal relationships in the natural system, or are the relationships merely mathematical artefacts, or some combination of the two? It would seem that a justification for using averaged data as a substitute or proxy can only occur after the underlying causal relationships in a system have been identified, since only then can it be argued from evidence that some given averaging method will continue to reliably maintain the synonymy of the inferential entailments in the model to the causal entailments in the natural system.

 

Finally, Spencer notes in his conclusion:

I think it is time to provoke some serious discussion and reconsideration regarding what we think we know about feedbacks in the real climate system, and therefore about climate sensitivity. While I’ve used the example of low cloud SW feedback, the potential problem exists with any kind of feedback.

For instance, everyone believes that water vapor feedback is positive, and conceptually justifies this by saying that a warmer surface causes more water to evaporate. But evaporation is only half the story in explaining the equilibrium concentration of atmospheric water vapor; precipitation is the other half. What if a decrease in precipitation efficiency is, instead, the cause of the surface warming, by not removing as much water vapor from the atmosphere? Then, it would be the water vapor increase driving the surface temperature change, and this would push the (unknown) diagnosed water vapor feedback in the positive direction.

And from a previous quote above:

But what never seems to be addressed is the question: What caused the temperature change in the first place? How do we know that the low cloud cover decreased as a response to the surface warming, rather than the other way around?

 

I wonder whether these must be either/or questions. It would certainly be simpler if we could identify the (many) relationships as being of the form ‘X drives a feedback Y in system S’. But perhaps some (or many) of the relationships are impredicative, whereby in effect X and Y both drive each other and are feedbacks to each other, simultaneously. This would render the system complex, in the Rosennean sense. Such systems are more difficult to model, insofar as computer models require fractionating such impredicative loops, and thereby lose the behavior in the system due to such causal relationships. Simple (i.e., non-complex) models can be made of complex systems, but each such model will apply only locally and temporarily, as the entailments and behaviors of the complex system will outstrip the simple models[1].  There is, of course, nothing about nature generally, or the climate system particularly, which would disallow such causal relationships from existing, and perhaps they are even prevalent in the climate system.

 

References:

[1] Rosen, R. 2000. Essays on Life Itself. Columbia Univ. Press.

Bookmark the permalink.

Comments are closed.