Ah, so I'm working at a level of generality that applies to all sorts of dynamical systems, including ones with no well-defined volume. As long as there's a conserved quantity Q, we can define the entropy S(Q) as the log of the number of states with that value of Q. This is a univariate function of Q, and temperature can be defined as the multiplicative inverse of the derivative dSdQ.
You still in general to specify which macroscopic variables are being held fixed when taking partial derivatives. Taking a derivative with volume held constant is different from one with pressure held constant, etc. It's not a universal fact that all such derivatives give temperature. The fact that we're talking about a thermodynamic system with some macroscopic quantities requires us to specify this, and we have various types of energy functions, related by Legendre transformations, defined based off which conjugate pairs of thermodynamic quantities they are functions.
By
dQ1dt∝1T1−1T2
I mean
dQ1dt(t)=a(1T1(t)−1T2(t))
for some constant a that doesn't vary with time. So it's incompatible with Newton's law.
And I don't believe this proportionality holds, given what I demonstrated between the forms for what you get when applying this ansatz with T versus with Q. Can you demonstrate, for example, that the two different proportionalities you get between T and Q are consistent in the case of an ideal gas law, given that the two should differ only by a constant independent of thermodynamic quantities in that case?
Oh, the asymmetric formula relies on the assumption I made that subsystem 2 is so much bigger than subsystem 1 that its temperature doesn't change appreciably during the cooling process. I wasn't clear about that, sorry.
Since it seems like the non-idealized symmetric form would multiply one term by T21 and the other term by T22, can you explain why the non-idealized version doesn't just reduce to something like Newton's law of cooling, then?
Here is some further discussion on issues with the 1T law.
For an ideal gas, the root mean square velocity vrms is proportional to √T. Scaling temperature up by a factor of 4 scales up all the velocities by a factor of 2, for example. This applies not just to the rms velocity but to the entire velocity distribution. The punchline is, looking at a video of a hot ideal gas is not distinguishable from looking at a sped-up video of a cold ideal gas, keeping the volume fixed.
Continuing this scaling investigation, for a gas with collisions, slowing down the playback of a video has the effect of increasing the time between collisions, and as discussed, slowing down the video should look like lowering the temperature. And given a hard sphere-like collision of two particles, scaling up the velocities of the particles involved also scales up the energy exchanged in the collision. So, just from kinetic theory, we see that the rate of heat transfer between two gases must increase if the temperature of both gases were increased by the same proportionality. This is what Newton's law of cooling says, and it is the opposite of what your proposed law says.
Here is a further oddity: your law predicts that an infinitely hot heat bath has a bounded rate of heat exchange with any system with a finite, non-zero temperature, which similar to the above, doesn't agree with how would understand it from the kinetic theory of gases.
This post is wrong. See this comment.
A high-school physics textbook will tell you that, according to Newton's law of cooling, a hot thing will cool down according to
dQdt∝T∞−T
where Q is heat, t is time, T is temperature, and T∞ is the temperature of the environment. But this is slightly wrong! It turns out that the correct cooling rate, which I'll call Lambert's law, is
dQdt∝1T−1T∞.
(Newton's law approximates this pretty well.) Let's derive this from first principles, and try to test it empirically.
Epistemic status: Logical deductions from a speculative physical model. I don't actually know nonequilibrium thermodynamics.
Rates of change
First, a lemma: I claim that if X is any macroscopic variable, then the rate of change of X is proportional to the derivative of the entropy with respect to X:
dXdt∝dSdX
Why? Suppose our system changes in discrete timesteps, and X can only change in steps of size ΔX. And suppose the probability of a transition X↦X+ΔX is proportional to the number of microstates Ω(X+ΔX). Then you can calculate the expected change in X per timestep, and from there prove the lemma.[1]
The Lambert cooling law
Now let our system comprise two subsystems. There's a fixed amount of heat split between the subsystems. Apply our lemma, taking X to be Q1, the amount of heat in subsystem 1. We find that[2]
dQ1dt∝dS1dQ1−dS2dQ2=1T1−1T2
Here the inverse of temperature, or coldness, shows up. The coldness of a system is defined to be the derivative of entropy with respect to energy. (The only energy here is heat.)
Now if subsystem 2 is very big, its temperature doesn't change much. So let's rename Q1 to Q, T1 to T; and T2 to T∞, because that's the temperature subsystem 1 will have in the limit:
dQdt∝1T−1T∞
We can use this to get an expression for dT/dt involving the second derivative of entropy, which is related to the heat capacity:
dTdt∝−d2SdQ2T2(1T−1T∞)
However, textbooks tend to assume that the heat capacity is temperature-independent, in which case heat is roughly proportional to temperature. So we have
dTdt∝1T−1T∞
This differential equation has a neat solution after a change of variables. Let
τ=(TT∞−1)eT/T∞−1
Then the solution is exponential decay:
τ=τ0e−rt
where τ0 is the value of τ at t=0, and r is a constant. If you want to express the solution in terms of T, you'll need the Lambert W function.
Experimental support?
Newton's law is a linear approximation of the Lambert law. If T is close to T∞ (and both are far from absolute zero), then the two laws are hard to distinguish.
I tried to distinguish them anyway, by measuring a hot baking sheet with a cooking thermometer. A best-fit Newton law predicts the data pretty well. The best-fit Lambert law is quite close to Newton (within 3 degrees Kelvin), but is a *worse* fit than Newton. I think my experimental setup has too much error to distinguish the two laws.
I'd guess a proper experiment would have to control convection, prevent the environment from heating up, and explicitly account for changing heat capacity. Cooling down to liquid nitrogen temperatures would make Lambert easier to distinguish from Newton, but that might be harder to work with.
Conclusion
If this cooling law is described somewhere in the nonequilibrium thermodynamics literature, I'd love to see a reference to it. Also, I'd love to know more about the assumption I started with: That a macroscopic variable must change gradually, and that the transition probabilities are proportional to the number of microstates; it seems relevant to dynamics in the same way that ergodicity is relevant to equilibrium.
Hints: Remember that S=logΩ. Also, to first order, the denominator Ω(X−ΔX)+Ω(X)+Ω(X+ΔX) is approximately 3Ω(X).
Hints: The entropy of the whole system is the sum of the entropies of the subsystems, S=S1+S2. And dQ2dQ1=−1.