Equilibrium free energy differences are given by exponential averages of nonequilibrium work values; such averages, however, often converge poorly, as they are dominated by rare realizations. I show that there is a simple and intuitively appealing description of these rare but dominant realizations. This description is expressed as a duality between "forward" and "reverse" processes, and provides both heuristic insights and quantitative estimates regarding the number of realizations needed for convergence of the exponential average. Analogous results apply to the equilibrium perturbation method of estimating free energy differences. The pedagogical example of a piston and gas [R.C. Lua and A.Y. Grosberg, J. Phys. Chem. B 109, 6805 (2005)] is used to illustrate the general discussion.The nonequilibrium work theorem,relates the work performed on a system during a nonequilibrium process, to the free energy difference between two equilibrium states of that system. The angular brackets denote an average over an ensemble of realizations (repetitions) of a thermodynamic process, during which a system evolves in time as a control parameter λ is varied from an initial value A to a final value B. W is the external work performed on the system during one realization; ∆F = F B − F A is the free energy difference between two equilibrium states of the system, corresponding to λ = A and B; and β is the inverse temperature of a heat reservoir with which the system is equilibrated prior to the start of each realization of the process. A sample of derivations of Eq. 1 can be found in Refs. [1,2,3,4,5,6,7,8,9, 10]; pedagogical and review treatments are given in Refs. [11,12,13,14,15,16]; for experimental tests of this and closely related results, see Refs. [17,18,19,20]; finally, Refs. [21,22,23,24,25] discuss quantal versions of the theorem. In principle, Eq.1 implies that ∆F can be estimated using nonequilibrim experiments or numerical simulations. If we repeat the thermodynamic process N times, and observe work valueswhere the approximation becomes an equality in the limit of infinitely many realizations, N → ∞. In practice, the average of e −βW is often dominated by very rare realizations, leading to poor convergence with N . The aim of this paper is develop an understanding of these rare but important realizations. I will argue that there is a simple description of these dominant realizations, which leads to both quantitative estimates and useful heuristic insights regarding the number of realizations needed for convergence of the average of e −βW . The organization of this paper is as follows. In Section I the central result is summarized, then illustrated using a simple example. Section II contains a derivation of this result. Section III discusses the number of realizations needed for convergence of the exponential average. Section IV focuses on the free energy perturbation method (a limiting case of Eq.1), and Section V concludes with a brief discussion.