Quote of the Week – models, climate sensitivity, the pause, and psychology

Unfortunately, numbers given as the ECS and TCR (TCS) and also given on their presumed forcing elements rely mostly on judgment in the process to be used to ‘estimate’ them.
Fortunately, on the other hand regarding measured climate focused ‘observational’ data . . . . well . . . nature is honest because there are no neat little sticky notes attached to any elemental piece of surface temp data in any of the GASTA or LTTA datasets saying “I was forced to change by changes in CO2 from fossil fuel”.
John

Steve Mosher,
Well, this thread has partly turned into something like a vote for highschool senior prom king and queen.
I rather enjoy your intellectual input, whether I agree with it or not. I find myself agreeing with the avenue your thought fairly often when even not agreeing with your premise or conclusion.
Limerick for you:

Limerick Title: ‘Is Occam Alive in San Francisco?’
There is a man nicknamed Moshpit
Whose meanings were oft sparse wit
Then he commented one day
In an unusual illuminated way
Which kept me thinking all night

John

“If we believe that model based approaches are best, then we need to have our heads examined.”
That’s swell.
But climate modelers are not going to voluntarily have their heads examined.
Will Mosher hold an intervention at his place?
Or is his furniture too non-confrontational?
Seinfeld Quotes. … Jerry and Elaine, in “The Pez Dispenser”; “I don’t have a good apartment for an intervention. The furniture, it’s very non-confrontational.

I don’t find it particularly surprising that Steven said this, because it is true and he’s not stupid. Especially the final comment. This is where the action is going to be over the next few years, assuming that warming doesn’t resume and pop the temperature up a half a degree in two years (which it could, I don’t pretend to understand the climate well enough to solve for its future behavior in my head either). Or if it actually cools a tenth of a degree or two. People are finally twigging to just how hard it is to computationally model the climate in a quantitatively meaningful manner out decades.
Here’s a nice metaphor. At state fairs in the fall, they have an event called a “turkey shoot”. Anybody can play. You take a shotgun with basically no choke, loaded with tiny shot, e.g. number 8 or so. A fairly long ways away (25-35 yards), you have a paper target with a small dot in the middle. You point the gun in that general direction, pull the trigger, and a few zillion small holes appear in the paper. You measure the distance from the closest hole to the target, per shooter. The one who comes closest wins a frozen turkey.
The point of this game is that there is basically very little skill involved. A half-blind teenager in a wheelchair can, if they can manage to point the gun downrange in the general direction of the target, get at least some pellets onto the paper, and if they the paper at all, they might come closest to the turkey’s-eye in the middle. An Olympic gold medal sharpshooter can’t do better than center their spread on the target and have more pellets that hit the paper, but they have no control over the individual pellets and the distribution of pellets is very broad, a meter or so wide and nearly flat within its width. I could aim at a point that would miss the target paper altogether with a rifle and still get almost as many pellets onto the paper as a central shot, in other words.
Modern models are basically a turkey shoot. Instead of relying on turbulence and chaos to spread out the tiny pellets of shot, they use the “perturbed parameter ensemble” — different random number seeds, small perturbations of the initial conditions and/or variable settings drawn from presumed distribution around equally presumed central values. The nonlinear chaotic dynamics inherent in the models does the rest. The result is that each model produces a staggeringly wide “bundle” of possible future climates from what should be more or less equivalent starting points.
This really does make them Anthony’s blind squirrels (one of my favorite metaphors as well) as much as a turkey shoot. Every blind squirrel finds an acorn, every blind squirrel in a turkey shoot gets at least a few pellets on the target paper and who knows, might even have “a winnah” and take home a frozen turkey to bowl with.
Now, how can we really compare the olympic sharpshooter to the 20/200 vision teenager with palsy? If we had a really big piece of paper, we could do things like find the centroid of the pellet distribution and the width of that distribution. One would expect that the sharpshooter would allow for ballistics and fairly consistently place their centroid dead on the target spot, while the poor teenager, who can barely lift the gun, might shoot consistently a half meter low and absolutely anywhere side to side. Either one could get a respectable number of pellets on the target paper per se, but a broader examination of the distribution of results would reveal that one of them truly is a metaphorical blind squirrel while the other one has some skill, even if the skill is largely defeated by the crude imprecision of the shotgun spread.
It is this latter analysis that the climate modelling community has steadfastly refused to do. Well over half of the models in CMIP5 as reported in AR5 are no better than our blind palsied squirrel participating in a turkey shoot. Sure, they get enough pellets onto the actual target that the participants can feel good about themselves and say “We could have won the turkey, c’mon Dad, buy me another ticket so I can try again, look, I had one run that was the third closest, c’mon Dad tickets are only five million dollars for three more years puh-leeze?” But even a cursory examination of the centroid of their “shots” compared to the actual target indicates that they are aiming consistently well above the target and it’s only because they are firing a bloody shotgun with a huge spread that they got any pellets at all near the turkey-eye.
Other models — a very few — actually look like might be produced by somebody developing some skill at shooting. I don’t think any of the models qualify as “Olympic sharpshooting” — all of the guns seem to shoot somewhat high and none of the shooters have figured out how to correct the sights to compensate — but some centroids aren’t absurdly high relative to the target and they consistently get pellets to cover the target zone pretty well.
If “Dad” really wants to win the turkey, it’s time to send the really bad shooters off to try their luck at ring the bottle or whack-a-mole, and concentrate the remainder of the shots he funds on the small handful of shooters with respectable skill, admonishing them to analyze their misses and get their sights adjusted before trying again.
The one thing that the shooters cannot at this time accomplish, however, is to do the one thing that really needs to be done. Tighten the choke right on down to where the pattern is no longer a meter plus wide. As it is, it is difficult to reject a poor shooter because the blindest of squirrels still can get 5% of their pellets onto the target, and we have this silly notion that 5% is “the” critical number for determining shooter incompetence instead of just ranking the shooters and sending the bottom half or bottom two thirds on (which would take around five minutes to do once the data were loaded into R and a suitable R routine written — it is the latter that are the “barrier” to the process such as it is).
If we tightened the choke, however, the patterns might end up no wider than the target sheet. A miss would be a miss, plain and simple, and any sort of accurate hit would drop many, many pellets very close to the turkey-eye of climate reality.
Sadly, there are very good reasons to think that we cannot expect to tighten the choke much, if at all. For one thing, there is a certain amount of irreducible spread because the dynamics are truly chaotic. For another, we are many, many orders of magnitude shy of being able to actually compute the models at the granularity of the fundamental true dynamics of the system being modelled. For a third, we do not know the actual initial state of the system at all accurately. For a fourth, we do not know the actual parameters input to the system dynamics at all accurately. For a fifth, in order to be able to do the computation at all, even at the terrible granularity we can currently afford, we have to use coarse grained approximations to the dynamics per cell. These approximations are supposedly representing the “mean” behavior of entire Navier-Stokes solutions for the cell on all smaller spatiotemporal scales.
They are therefore themselves at best turkey shoots within the turkey shoot, only they don’t even produce a turkey shoot with the range of internal noise and variability of nature at scale, they feed the centroid of a turkey shoot with some presumed dynamic into the next step of each cell’s computation at even coarser dynamics. My metaphor fails me — describing coarse grain averaging and renormalization in stat mech as a stack of sequentially reaveraged turkey shoots, each layer used to do supposedly physics based dynamics at a coarser layer above it even though all that gets passed up are the crudest of semi-empirical averages with a few wild-guess parameters controlling the interaction both ways — it just doesn’t do it justice. A Monte Carlo of Monte Carlos, a turkey shoot of turkey shoots, models of squirrel-blindness in particular acorn-laden forests used to model blind squirrel based acorn finding worldwide.
But the real sin, the unforgivable statistical sin, is when they take all of the results of these turkey shoots of turkey shoots, most of which are made by palsied teenager blind squirrels who always shoot high, and then average them all together as blindly as any squirrel might to claim that reality is really where the collective centroid of this bloody turkey shoot of turkey shoots of turkey shoots lies. The multimodel ensemble mean as an evil of the Universe that needs to simply disappear forever and without a trace, never to be referred to again in any assessment report.
We can, with some care, assess the skill of individuals even when they are participating in a turkey shoot with an unchoked shotgun with tiny pellets on a breezy October afternoon, especially if we get to look at the distribution of holes they produce in the sheet behind the target and not just at how many pellets they get “close” to the target. We have no good axiomatic statistical basis for expecting the collective to have any skill at all, and empirically it is clear at a glance that the collective centroid lies well above the target and has no skill, reflecting the lack of skill of most of the shooters.
Metaphorically yours,
rgb (who is at worst a presbyopic squirrel who is actually a pretty good shot and has been known to find the occasional acorn:-)