Feedback, climate, algebra and circuitry.

I've been arguing again at WUWT ( (more here)). It is the fourth of a series by Lord Monckton, claiming to have found a grave error in climate science, so it is now game over. My summary after three posts is here.

The claim is, of course, nonsense, and based on bad interpretation of notions of feedback. But I want to deal here with the general use of feedback theory in climate, and the mystery that electrical engineers who comment on this stuff like to make of it. The maths of feedback is trivial; just simple linear equations. And it is best to keep it that way.

A point I often make in commentary is that climate science really doesn't make much use of feedback theory at all. Critics invoke it a lot more. I continually encounter people who think that feedback is the basis of GCMs. I have to explain that, no, they do not form any part of the structure of GCMs, and cannot. A GCM is a solver for partial differential equations. That means it creates for each step a huge array of linear equations relating variables from neighboring cells. That isn't always obvious in the explicit methods they tend to use, but there is still an underlying matrix of coefficients. And because each row just related a few neighboring values, the matrix is sparse. This is an essential feature, because of the number of cells. But global averages, such as would come from a feedback expression, are not sparse. They connect everything. So they cannot fit within the discretised pde framework.

Linear equations and feedback

Problems described as feedback are really just linear equations, or systems of a few linear equations; usually one less equations than unknowns, so on elimination, one variable is expressed as a multiple of another. I described here how a feedback circuit could be analysed simply by writing linear current balance (Kirchhoff rule) equations at a few nodes. In climate, the same is done by balancing global and time average heat fluxes, usually at TOA.

The paper of Roe 2009 is often cited as the most completely feedback oriented analysis. I'll show its presentation table here:

It gives the appearance that ΔR is both input and output, because it is a flux that is conserrved. But the more conventional feedback view is that ΔT is the output. If we take the multi-feedback version of (c)
ΔT = λ₀(ΔR + Σc_iΔT )
which I can rewrite setting c₀=-1/λ₀ as just
ΔR + c₀ΔT + Σc_iΔT = 0

This is just the equilibrium heat flux balance at TOA since each of the c_iΔT is a temperature-responsive flux. I have given the c₀ΔT special status, because it is the Planck term, representing radiation guaranteed by the Stefan-Boltzmann law (c₀ = -4ΔT).

Feedback reasoning and linear equations

Just resolving a linear equation is not a mathematical difficulty. So what is all the feedback talk about? Mainly, it is trying to see the equation as built up in parts. There is no math reason to do that, but people seem to want to do it. The process can be described thus:

• Select (as in Roe above) a subset to refer to as the reference system. A logical set is the forcing and the necessary Planck response.
  ΔR + c₀ΔT + ΣcₖΔT = 0
  This is like a finite gain amplifier (c₀)
• Express the other terms as feedbacks relative to c₀:
  ΔR + c₀ΔT *(1 - Σfₖ) = 0, fₖ = -cₖ/c₀
  The f's are then called the feedback coefficients. For stability (see next) they should sum to less than 1. Negative values make this more likely, and so are stabilising. As the coefficient of ΔT, diminishes, it increases the amount by which ΔT would have to change to keep balance. That is said to increase the gain, and creates a singular situation (of high gain) approaching zero.

Stability

If the singularity is passed (Σfₖ>1), and the coefficient of ΔT becomes positive, the system is unstable. The reason involves an extra bit of physics. Suppose total flux is out of balance. Then the region into which it flows will cool or heat. The coefficient here is, for a uniform material, called the heat capacity H, and is positive. For a complex region like the Earth surface, that is hard to quantify, but will still be positive. That is, heat added will make it warmer, not cooler. So the equation for temperature change following imbalance is
ΔR + cΔT = H*dΔT/dt
If c is positive, this has exponentially growing solutions, and so is unstable. For c negative, the solutions decay, and lead toward equilibrium.

It's often said that positive feedback is impossible, because it would mean instability. But in the above algebra, that is not true; the requirement is that Σfₖ>1. It is true if you choose a different reference system - just the forcing. That can only work in conjunction with a c₀ΔT where c is negative. Electrically, the reference system is then like an operational amplifier.

Summary so far

Systems often described using feedback terminology are really just linear equations (or systems). Feedback arguments do not yield anything beyond what elementary linear solving can do, including a stability criterion. But with linear algebra, you can identify the various steps of feedback reasoning if you want to.

Systems are not exactly linear

Roe points out that linear feedback is just the use of a first order Taylor Series expansion of a nonlinear relation. This is very direct seen as a linear system. If the forcing R is to be balanced by a flux F which is a function of T and variables u,v which depend on T, then to first order

dR = (∂F/∂T) dT + (∂F/∂u du/dT) dT + (∂F/∂v dv/dT) dT

each partial holding the other variables (from T,u,v) fixed. This gives the required linear relation with the bracketed terms becoming the c coefficients (but negative).

More advanced

There is a lot of approximation here. Not only linearity (usually OK) but also in the use of global averaging. But that doesn't mean linear analysis has to be discarded if you want to take account of these things. You can extend using an inexact Newton's method. Suppose we have the base system

R = F(u,v,T)

where again u and v are variables (like humidity) that depend on T. Suppose we have an initial state subscripted 0, and a perturbed state subscripted 1, of which R₁ is known. Then to first order

F(u₁,v₁,T₁) - R₁ = F(u₀,v₀,T₀) - R₁ + (∂F/∂T)₁ dT + (∂F/∂u du/dT)₁ dT + (∂F/∂v dv/dT)₁ dT = 0

This can be solved as before as a linear equation in dT. Then updating

T = T + dt, u = u + du/dT)₁ dT etc, we can solve again

F(u,v,T) - R₁ + (∂F/∂T)₁ dT + (∂F/∂u du/dT)₁ dT + (∂F/∂v dv/dT)₁ dT = 0

and iterating until F(u,v,T) - R₁. Note that I have not updated the partial derivatives, which are the feedback coefficients. That is what makes it an inexact Newton; convergence is a bit slower, but we probably don't have information to do that update.

So non-linearity is not a show-stopper; it just takes a little longer. This also allows you to work out a more complicated version of F, with, say, latitude variation. You can still use the simpler global feedback coefficients, so the extra trouble is only in the evaluation of F. The penalty will again be slower convergence, and it may even fail. But it gives a way to progress.