This will probably be my final post on this question - however I may append updates if other issues come up. In particular this post will look first at whether the fitted parameter values for two-time-constant fits to temperature vs forcing data for Earth's climate system have a collection of underlying physical two-box models that satisfy the basic physical constraints on such systems, and then whether the range of physical parameters in these matching models appears to correspond roughly with appropriate associated physical properties of the real Earth climate system.
CORRECTION: The following text has been modified significantly due to errors in the preceding post that nullify most of the original discussion here:
The most recent post using a perturbative approach provides some initial guidance on these issues, but we'll return to the full model shortly since it's not really much more complicated to solve. From the perturbative solution, for small values of the thermal coupling (β, i.e. γ's small compared to 1/τ+), the constraints of Eq. 36A and 36B of the previous post limit the allowed values for the fast-box areal heat capacity Cs given a heat capacity ratio λ (and the fitted parameters for which we're trying to develop a solution). For the "Tamino" case, if λ is in the range 1 to 100, the excluded range for Cs can range from around 10^7 to over 10^10 J/Km^2, which includes the physical range of the atmosphere plus ocean system. That means there is a large subspace of choices in the physically realistic region for Cs and Co that are excluded (at least for small values of the γs heat transfer rate) because the fit cannot lead to a real value for the y parameter. Even for those with real y, we may not get a result within the allowed range 0 to 1. So for any given case we'll need to look at the particular constraints in some detail to find solutions that are allowed by the math of the problem.
The upper bound of Eq. 40 provides a simpler constraint on the fast-time-constant heat capacity Cs:
END OF CORRECTION
Cs < τ-/a3
since y must be at most 1. From the earlier analysis, smaller assumed τ- values led to smaller fitted coefficients a3, so the fit is probably associating a physical heat-capacity-related constraint of this sort on the resulting a3 values. For the sample case of Tamino's τ- = 1 year and translating to SI units as above we get
Cs < 3.16x10^7 J/Km^2 /a3SI
On the other hand, Cs should be at least as large as the heat capacity of the well-mixed atmosphere (the troposphere which contains the bulk of the atmosphere's mass, and heat capacity). Lucia's value of about 1.34x10^7 J/Km^2 seems reasonable, so from both these constraints we should expect a reasonable short-time-constant heat capacity to be on the order of a few times 10^7 J /K m^2, certainly not much less, and not a lot more than that.
Now, returning to the solution space, as noted at the end of the perturbative discussion, the choice of αs and αo as two of the three free parameters was perhaps not so wise. In the following the three free parameters in our two-box system will be the two heat capacities Cs and Co (therefore determining their ratio λ = Cs/Co) whose physical values we have just considered and somewhat constrained, and the heat transfer coefficient γs (which then determines γo = λ γs - and of course the original β = Cs γs). The perturbative approach works when γs is small compared to 1/τ+ (and γs must be positive) so the range for γs to look at for the full solution is from close to zero to on the order of 1/τ+ or perhaps higher. Higher values for the heat transfer physically imply that the two boxes are not really so isolated from one another.
The solution process of the earlier post can still be followed, but we need to rearrange Eq. 22 and 23 to solve for αo and αs in terms of γs and λ now, instead of the other way around. The best starting point is actually the pair of equations (with (+) or (-)) represented by the unnumbered expression just before Eq. 22 in that post. Divide that expression by γs and simplify to dimensionless variables by defining:
φs = αs/γs
φo = αo/γs
ξ+ = 1/(γsτ+)
ξ- = 1/(γsτ-)
(note that all of these become large in the perturbative limit γs close to 0) we then get:
Eq. 41: ξ+ - φo + ξ+ φs - φoφs - λ φs - ξ+2 + ξ+φo + λ ξ+ = 0
together with the identical equation replacing ξ+ with ξ-.
Rearranging Eq. 41 to solve for φo in terms of the others, we find:
φo(1 + φs - ξ+) = ξ+ + ξ+ φs - λ φs - ξ+2 + λ ξ+
Eq. 42: φo = (ξ+ + ξ+ φs - λ φs - ξ+2 + λ ξ+)/(1 + φs - ξ+)
and the same equation should hold for the (-) version, which we can combine on the way to finding an expression for φs:
Eq. 43: (1 + φs - ξ+)(ξ- + ξ- φs - λ φs - ξ-2 + λ ξ-) = (1 + φs - ξ-)(ξ+ + ξ+ φs - λ φs - ξ+2 + λ ξ+)
Bringing the right-hand expression in Eq. 43 to the left, canceling out the terms symmetric in ξ+ and ξ-, organizing in powers of φs and dividing by (ξ- - ξ+) we are left with:
Eq. 44: φs^2 + (2 - ξ- - ξ+)φs + 1 + λ - ξ- - ξ+ + ξ+ξ- = 0
which is a standard quadratic equation with two solutions for φs:
φs = (-b ± sqrt(b^2 - 4c))/2
where -b = ξ+ + ξ- - 2 and c = 1 + λ - ξ- - ξ+ + ξ+ξ- so the condition b^2 - 4c > 0 becomes:
Eq. 45: (ξ- - ξ+)^2 > 4 λ
With τ- the short time constant, ξ- > ξ+, so going back to our definitions for the ξ's, this constraint becomes:
Eq. 46: (1/τ- - 1/τ+) > 2 sqrt(λ) γs
which gives a limit on the magnitude of the heat transfer coefficient γs. If γs becomes too large relative to the τ's and λ as indicated here, then the corresponding time-constant becomes complex, and the parameter space is no longer suitable for the original physical interpretation of the two-box model.
Note that if we want αs to be closer to the inverse of the short time constant τ- then we need to take the (+) sign in Eq. 44, which will be assumed going forward.
Eq. 42 then gives φo and hence αo, and we can continue to find the remaining parameters as in the earlier post.
Let's plug in values appropriate to Tamino's time-constants and see how they do. Set τ+ to 30 years, τ- to 1 year. The fitted parameters are a2SI = 0.739 and a3SI = 0.038. Assume the short-time-constant-box heat capacity Cs is just that of the atmosphere, 1.34x10^7 J/Km^2. Give the long-time-constant-box Co a heat capacity of 1% of the full ocean, so 1.06x10^8 J/Km^2. For these time-constants and heat capacities the associated constraints listed earlier (associated with Eq. 36A, B and 40) are all satisfied.
This leaves the heat transfer rate γs as a free parameter. We'll start with γs = 3.17x10^-11 s^-1 which is roughly a factor of 30 smaller than 1/τ+ = 1.06x10^-9 s^-1.
With the input time constants and coefficients set, and the free heat capacities and transfer rate given these physically reasonable values, we find for this "Tamino" set of parameters:
αs = 3.17x10^-8 s^-1 (very close to 1/τ-)
αo = 1.05x10^-9 s^-1 (very close to 1/τ+)
r+ = 967
r- = -1.31x10^-4
(a quick check shows eq. 17 is satisfied by these when the precise values for the α's are used).
Then Eq. 26 allows for two complementary values for y, one in which the measured temperatures are close to those of the fast box, and one where they are close to the slow box:
Case 1 (+ sign in solution from Eq. 26):
y = 0.917
w+s = 9.08x10^-3 K/(W/m^2)
w-s = 4.15x10^-2 K/(W/m^2)
x = 0.0177
Case 2 (- sign in solution from Eq. 26):
y = 0.0177
w+s = 7.79x10^-4 K/(W/m^2)
w-s = 2.16 K/(W/m^2)
x = 0.917
Neither of these are actually very good solutions because the split of forcings that results (x) is far too asymmetrical - in case 1 the "slow" box heats up significantly while the fast box does little, but most of the weight in measured temperature comes from the fast box, so neither one makes sense as a model of a part of Earth's climate. In case 2 the reverse happens with the same result.
So, let's try a few more examples for the "Tamino" parametrization. Set Co 10 times larger, to 1.06x10^9 J/Km^2 (about 10% of the ocean), and increase γs to 1.0x10^-8 s^-1 (300 times larger than before, so about 10 times bigger than 1/τ+) with the remaining input parameters the same. That gives us
αs = 2.16x10^-8 s^-1
αo = 9.71x10^-10 s^-1
r+ = 3.06
r- = -4.13x10^-3
and two more cases to look at -
Case 3 (+ solution for y):
y = 0.156
w+s = 0.270 K/(W/m^2)
w-s = 0.249 K/(W/m^2)
x = 0.110
Case 4 (- solution for y):
y = 0.110
w+s = 0.261 K/(W/m^2)
w-s = 0.358 K/(W/m^2)
x = 0.156
The x values and response numbers now do look somewhat sensible - perhaps a little too sensitive to the short time constant in the second case, but not far off.
Note: the following text is obsolete, and has been replaced by the above discussion.
Case 1 (+ sign in solution from Eq. 26):
y = 0.910
w+s = 8.40x10^-3 K/(W/m^2)
w-s = 4.18x10^-2 K/(W/m^2)
x = 0.0179
Case 2 (- sign in solution from Eq. 26):
y = 0.0912
w+s = 8.42x10^-4 K/(W/m^2)
w-s = 0.417 K/(W/m^2)
x = 0.177
As far as I can tell, both the above sets of parameters constitute a physically reasonable realization of a two-box model for Tamino's original fit. And these are just two of an infinite collection that are both physical in principle and in practice have values (for heat capacity, time constants, etc.) that are in the reasonable range for our planet.
Increasing the value of γs by a factor of 100, so that it is actually larger than 1/τ+, changes these values enough that the (+) solution for y is larger than 1, and so probably implausible. However, the (-) solution is still viable, and has:
αs = 2.85x10^-8 s^-1
αo = 6.97x10^-10 s^-1
r+ = 9.65
r- = -1.31x10^-2
y = 0.106
w+s = 8.46x10^-2 K/(W/m^2)
w-s = 0.403 K/(W/m^2)
x = 0.172
Quintupling Cs to 6.7x10^7 J/K m^2 changed almost none of these parameters - the only ones substantially affected were r- and x, which were both made larger by roughly a factor of 5 (so unphysical values of x, larger than 1, would obtain if Cs was much larger than this).
Quintupling Co however was sufficient to violate the Eq. 36 constraint, so that the fraction y became a complex number and the remainder of the solution made no sense. So the results here make plenty of sense, but only if you choose reasonable values (that can be justified relative to Earth's climate system) for the heat capacities of the two boxes in question.
In short, there certainly do exist (an infinite number of) two-box models with physical parameters that resemble components of Earth's climate system, which correspond to the fitting parameters found in Tamino's elegant demonstration of simple ways to estimate climate sensitivity from the measured data.
UPDATE: I've posted a Google Docs Spreadsheet that does the math on these and other examples. The above-mentioned results are from rows 14 through 17.
UPDATE 2 - Sep 8, 2009: It looks like I made a boo-boo somewhere. Lucia pointed out a problem here (her Eq 1N, 2N) which indicate my solutions don't satisfy my own Eq. 11 and 14. I've added checks to the above spreadsheet (columns Y through AF) to compare the solutions to the various equations stated so far, and indeed there are some problems. Well, better to have algebra checked by blog post than to go too far with wrong-headed results!
UPDATE 3 - Sep 8, 2009: See corrections to the previous posts - the problem was in the original Eq. 26, which led to the solutions for y; the original post here had extensive discussion on plausible values for Co based on an associated relationship, but that turns out to be wrong - the constraint is instead on the combination of Cs and the heat capacity ratio λ, and isn't nearly as simple as I thought. I have simply removed most of the original discussion of this since in the correct context it is unjustified. Thanks to Lucia for suggesting some likely sources of problems.