Nonequilibrium work and fluctuation relations

Consider a prescribed time-dependent control parameter taking a Hamiltonian from an initial form to a final form. Work $W$ is defined as work done on the system, and the process starts from a canonical equilibrium ensemble at positive temperature $T>0$ and inverse temperature $\beta=1/(k_BT)>0$. The free-energy difference $\Delta F$ compares the equilibrium ensembles of the endpoint Hamiltonians at that same temperature, even if the actual final distribution has not equilibrated. Mechanical driving, thermal exchange, and work measurement must be defined consistently.

Under the usual microreversible classical Hamiltonian or suitable heat-bath dynamics assumptions, Jarzynski's equality relates the exponential work average to that free-energy difference. A quantum formulation needs a specified work-measurement protocol, commonly two projective energy measurements, rather than treating work as an arbitrary Hermitian observable. Feedback, postselection, noncanonical initial conditions, and strong coupling can require modified relations. The equality is powerful because it allows far-from-equilibrium driving under its assumptions, not because it removes those assumptions.

Crooks's work relation compares a forward protocol with the appropriately time-reversed protocol, including reversal of odd variables and external fields when required. The reverse process starts in equilibrium for the forward final Hamiltonian. The ratio of forward probability for work $W$ to reverse probability for work $-W$ is an exponential of the dissipated work. This is a statement about paired ensemble probabilities, not a claim that one observed path has reversed spontaneously with equal likelihood.

Integrating the probability-ratio relation gives Jarzynski's equality when the distributions and supports satisfy the required conditions. The forward and reverse work distributions cross at $W=\Delta F$ where a well-defined common density or matching discrete masses exist. Rare negative dissipated-work outcomes are allowed; their relative probabilities are constrained. A small system can temporarily deliver more work than the free-energy drop without violating the average bound, because the distribution, rather than every realization separately, carries the thermodynamic statement.

The factor $e^{-\beta W}$ gives unusually large weight to low-work realizations. Under strongly dissipative driving those realizations may be very rare, so a finite sample can underestimate the exponential mean and bias an inferred free energy. Increasing the number of typical high-work trajectories does not guarantee that the influential rare tail has been sampled. Forward–reverse methods, slower protocols, repeated uncertainty analysis, and overlap checks can be more informative than one apparently converged running average.

The endpoint free-energy difference is an equilibrium state function, while the work distribution depends on the driving protocol. Different protocols can therefore share the same exact exponential average and have very different mean dissipation and sampling difficulty. Finite systems are especially convenient for checking this distinction because every state can sometimes be enumerated. Macroscopic extensivity usually makes dissipation and rare-event costs grow with size, so an equality that remains formally valid can become impractical to estimate by naive repeated driving.