Notes on Dissipation: the Phase-Space Perspective
[What follows is not scientifically rigorous; it’s just the author’s personal opinion, not necessarily well-informed]
Paper by R. Kawai, J. M. R. Parrondo, and C. Van den Broeck, published in PRL 2007. They obtain a simple new result about the thermodynamics of far-from-equilibrium systems with the usual Jarzynski statistical-mechanical techniques.
Preliminaries
In an irreversible process the dissipated work is given by
and has the meaning of amount of work done on the system that does not directly contribute to a change in the macroscopic state-function, the thermodynamical potential useful for later transformations. This is why it is "dissipated". The authors make the assumption that the system be initially in contact with a bath at temperature T so that the initial microstates can be picked up from a canonical distribution
(1)
They further assume that the bath is then removed and the system is mantained in isolation, so that the only energy exchange with the environment is through the work performed, and the first law of (microscopic) thermodynamics* assures that all the work done is gained in the form of internal energy:
This also implies that there is no phase-space contraction and therefore that Lioville theorem holds along a trajectory
where we used the notation z = (q,p) as the collective vector of positions and momenta.** The physical meaning is simply determinism and conservation of information: if a given initial condition is picked up with probability \rho than its evolute at time t still has probability \rho to show up. Owing to the fact that there is no heat exchange (or, better, that the heat flux is known) we do not loose information about the system.
We also consider an inverse process, starting with the final energy and performing the inverse transformation, pulling out work from the system and lowering its internal energy to the initial value. The initial state of this inverse process is also sampled from a canonical distribution \rho_B, meaning that the system is re-chilled to temperature T before going back. Esponentiating (1) yealds
and if one averages above all the trajectories one obtains Jarzynski equality:
More interestingly, one has (setting \beta=1)
(2)
Average dissipated work is the relative entropy between the two distributions \rho_A and \rho_B, a measure of how much they differ. It is always positive (Gibbs’s inequality). This is a very nice relationship but really adds little to what previously known, as it also makes its appearance in a (badly written) section of the wikipedia article about Kullback-Leibler divergence.
* The first law IS microscopic; it is the second law that has both macroscopic and microscopic interpretations.
** Remember that
and that dz = dz(t) by phase-space volume conservation.
The new result
Now let’s follow one single trajectory from the initial state to the final one and back. Due to phase-space volume conservation we have
where we have assumed that the hamiltonian is time-reversible, that is even under inversion of momenta. Substituting into (2) one gets
where the \dagger means time inversion. This is eq. (5) in the paper, with the only difference that we are expliciting time dependence of the densities. Here the paper is unclear about what it is being integrated over, I get confused about whether the argument of the densities are state variables or trajectories. But one can of course perform a change in the integration variables and gets
Because of hamiltonian dynamics the jacobian determinant is 1 and we obtain the desired result
(3)
where we have made use of the Kullback-Leibler notation.
Coarse graining
An important property of relative entropy is that it decreases upon coarse graining. Let’s consider a partition of the state space in cells \omega_i and define the coarse-grained densities:
Relative entropy is given by
The article displays a straighforward demonstration of this fact. Calling W = \ln \rho/\mu one has
Now \exp -W is a convex function and therefore by Jensen’s inequality one has
Summing over j yealds the relative entropy in the LHS and the coarse-grained relative entropy in the RHS. If one has limited knowledge of the probability distributions over the state space, one still can get a lower bound on dissipated work. In any case this result refines the second law of thermodynamics. Examples follow in the article.
Discussion and considerations
The nice thing is that (3) compares the forward and backward distributions at the same arbitrary time: in fact such distributions encode all the information we need from a full trajectory, which is splitted into two contributions one of which is reinterpreted as backward. Theoretically one could determine the average dissipated work just by measuring the phase space distributions for forward and backward trajectories. This does not really add a powerful tool, since all we needed in (2) was just knowledge of the initial and final distributions (that is, the energy function).
In any case, dissipation is put in relation with a measure of time-symmetry breaking, as is usually done in such articles. We usually expressed heat in terms of probabilities of forward over backward paths (path-dependent). Here we have dissipated work in terms of the state of the forward and backward path at agiven time (state-dependent).
The problem is that the result works in a very limited situation, that of adiabatic processes, for which exact dissipated work is very easily calculated. Jarzynski’s equality instead holds in greater generality also in the case of an heat flux. Can we think of a more general result? Otherwise we might be absording heat flux into the definition of work. Is this a satisfactory modelization of a physical system or a rather peculiar one? Moreover, can we always do that? What kind of renormalization/coarse graining procedure can we think of to transform the microscopic radom heat flux into a macroscopic deterministic work.




