Non-Equilibrium Statistical Mechanics

All preceding chapters have been concerned, ultimately, with equilibrium: configurations drawn from the Boltzmann distribution $\pi(\mathbf{q}) \propto e^{-\beta U(\mathbf{q})}$, ensemble averages that do not depend on time, and free energies that are state functions. Equilibrium statistical mechanics is a complete and elegant theory. But most processes of scientific interest are not equilibrium processes. A protein folds under non-equilibrium conditions; an enzyme operates in a cell that is constantly consuming ATP; a single molecule is pulled by an atomic force microscope tip at a finite rate. Even the enhanced sampling methods of Chapter 5 — steered MD, metadynamics — drive the system out of equilibrium as the bias is applied. This chapter develops the theoretical tools for connecting non-equilibrium observations to equilibrium quantities.

What Changes Out of Equilibrium

In an equilibrium simulation, the Hamiltonian is time-independent and the distribution of configurations is stationary. In a non-equilibrium process, the system is driven by a protocol $\lambda(t)$: a time-dependent parameter that appears in the Hamiltonian, changing it from $H(\mathbf{q}; \lambda_0)$ at time $t = 0$ to $H(\mathbf{q}; \lambda_\tau)$ at time $t = \tau$. The parameter $\lambda$ might be the position of an AFM tip, the end-to-end distance of a polymer pulled at constant velocity, or the coupling parameter of an alchemical transformation ramped from 0 to 1.

[!definition] 8.1 — Non-Equilibrium Work The work done on the system by driving the protocol from $\lambda(0) = \lambda_0$ to $\lambda(\tau) = \lambda_1$ along a trajectory $\mathbf{q}(t)$ is:

$$W = \int_0^\tau \frac{\partial H(\mathbf{q}(t); \lambda(t))}{\partial \lambda}\, \dot{\lambda}(t)\, dt = H(\mathbf{q}(\tau); \lambda_1) - H(\mathbf{q}(0); \lambda_0) - Q,$$

where $Q$ is the heat exchanged with the thermostat. For a switching process at finite speed, $W$ fluctuates from one trajectory realization to the next — even if the protocol $\lambda(t)$ is fixed — because the initial conditions are drawn from the equilibrium distribution of $H(\cdot; \lambda_0)$ and the trajectory is stochastic.

The second law of thermodynamics states that the mean work satisfies $\langle W \rangle \geq \Delta F$, where $\Delta F = F(\lambda_1) - F(\lambda_0)$ is the equilibrium free energy difference between the two endpoints. Equality holds only in the quasi-static limit $\tau \to \infty$, where the protocol is so slow that the system remains in equilibrium at every instant.

The dissipated work $W_{\mathrm{diss}} = W - \Delta F \geq 0$ measures the irreversibility of the process. Fast switching dissipates more work than slow switching; the excess is converted to heat. The central question of non-equilibrium statistical mechanics is whether one can extract $\Delta F$ — an equilibrium quantity — from measurements of $W$ made under finite-speed, irreversible conditions. The surprising answer, established by the work theorems of the 1990s, is yes.

The Jarzynski Equality

The Jarzynski equality (Jarzynski, 1997) is one of the most striking results in modern statistical mechanics. It states that the equilibrium free energy difference can be recovered exactly from an average over non-equilibrium work values:

[!definition] 8.2 — The Jarzynski Equality For a system initially at equilibrium with Hamiltonian $H(\cdot; \lambda_0)$ at temperature $T$, driven by an arbitrary protocol $\lambda(t)$ of duration $\tau$:

$$e^{-\beta \Delta F} = \left\langle e^{-\beta W} \right\rangle,$$

where the average is over all realizations of the stochastic trajectory (different initial conditions drawn from $\pi_{\lambda_0}$, different noise realizations in the thermostat), and $\Delta F = F(\lambda_1) - F(\lambda_0)$ is the equilibrium free energy difference between the final and initial states.

The equality (8.2) is remarkable: the left side is an equilibrium quantity; the right side is an average over arbitrarily far-from-equilibrium trajectories. The protocol can be as fast as desired — instantaneous quenches are allowed — and the equality remains exact. The only requirement is that the system starts in equilibrium.

The derivation in the overdamped limit follows from a change of measure. For any trajectory $\mathbf{q}(t)$ generated by Langevin dynamics at protocol $\lambda(t)$, define the path weight $\mathcal{P}[\mathbf{q}]$ as the probability density of observing that particular trajectory. The ratio of forward to reverse path weights satisfies:

$\ln \frac{\mathcal{P}[\mathbf{q}]}{\mathcal{P}^{\dagger}[\mathbf{q}^{\dagger}]} = -\beta W[\mathbf{q}] + \beta \Delta F,$

where $\mathcal{P}^{\dagger}$ is the path weight under the time-reversed protocol and $\mathbf{q}^{\dagger}(t) = \mathbf{q}(\tau - t)$ is the time-reversed trajectory. Exponentiating and integrating over all trajectories gives (8.2) immediately.

[!remark] 8.1 The exponential average in (8.2) faces the same convergence problem as free energy perturbation (Chapter 6): it is dominated by rare trajectories where $W \ll \Delta F$, i.e., trajectories where the system absorbs anomalously little work and the process happens to be nearly reversible. Under fast switching, such trajectories are exponentially rare, and the estimator $\hat{\Delta F} = -k_BT \ln (N^{-1} \sum_i e^{-\beta W_i})$ has exponentially large variance. In practice, the Jarzynski equality is useful when the switching is not too fast — when the work distribution $P(W)$ is narrow enough that its lower tail is well sampled with a tractable number of trajectories. The Bennett acceptance ratio can be applied here too: by combining forward and reverse work measurements, one obtains the Crooks fluctuation theorem estimator, which has optimal variance for a given number of trajectories.

The Crooks Fluctuation Theorem

The Crooks fluctuation theorem (Crooks, 1999) is a stronger result from which the Jarzynski equality follows as a corollary. It relates the full distribution of work values under the forward protocol to the full distribution under the reverse protocol.

[!definition] 8.3 — The Crooks Fluctuation Theorem Let $P_F(W)$ be the distribution of work done on the system under the forward protocol (from $\lambda_0$ to $\lambda_1$), with the system starting in equilibrium at $\lambda_0$. Let $P_R(-W)$ be the distribution of the negative work (i.e., work done on the system during the time-reversed protocol, from $\lambda_1$ to $\lambda_0$), with the system starting in equilibrium at $\lambda_1$. Then:

$$\frac{P_F(W)}{P_R(-W)} = e^{\beta(W - \Delta F)}.$$

The forward and reverse work distributions cross at $W = \Delta F$: the intersection point of $P_F$ and $P_R$ gives $\Delta F$ directly, without exponential averaging.

The Crooks theorem is more general than the Jarzynski equality in several respects. It provides information about the entire distribution of work, not just its exponential average. It establishes that the irreversibility of a process — measured by the ratio $P_F(W)/P_R(-W)$ — is directly related to the work in units of $k_BT$. And it enables a practically superior estimator: the crossing point of $P_F(W)$ and $P_R(-W)$ converges faster than the Jarzynski exponential average because it relies on the bulk of both distributions rather than their rare low-work tails.

Integrating both sides of (8.3) over $W$ and using the normalization of $P_R$ gives:

$\int P_F(W)\, dW = e^{-\beta \Delta F} \int P_R(-W)\, e^{\beta W}\, dW = e^{-\beta \Delta F} \cdot \langle e^{\beta W} \rangle_R,$

which, since the left side equals 1, gives the Jarzynski equality applied to the reverse process. The Jarzynski equality in the forward direction follows by symmetry. The Crooks theorem is thus more fundamental.

The Second Law as a Consequence

The second law, $\langle W \rangle \geq \Delta F$, follows from the Jarzynski equality by Jensen’s inequality. Because the exponential is a convex function,

$e^{-\beta \Delta F} = \langle e^{-\beta W} \rangle \geq e^{-\beta \langle W \rangle},$

taking logarithms gives $\langle W \rangle \geq \Delta F$. Equality holds if and only if $W$ is a constant — i.e., every trajectory does the same work, which happens only in the quasi-static limit.

The fluctuation theorems thus place the second law in a broader context: it is not an absolute prohibition on certain processes but a statement about the typical behavior of work fluctuations. Atypical fluctuations — trajectories where $W < \Delta F$, violating the second law locally — occur with probability proportional to $e^{-\beta(W - \Delta F)}$, which is small but nonzero for any finite-time process. In a macroscopic system at room temperature, such fluctuations are negligible (the suppression is $e^{-N}$ for $N$ degrees of freedom); in a nanoscale or single-molecule system, they are observable and experimentally relevant.

Stochastic Thermodynamics

The work theorems above apply at the level of ensembles: averages over many trajectory realizations. Stochastic thermodynamics (Seifert, 2005, 2012) defines thermodynamic quantities — work, heat, entropy production — along individual stochastic trajectories and derives fluctuation theorems at the trajectory level.

For a single trajectory $\mathbf{q}(t)$ at $0 \leq t \leq \tau$, the total entropy production is:

where $\Delta s_{\mathrm{sys}} = \ln \pi(\mathbf{q}(0); \lambda_0) - \ln \pi(\mathbf{q}(\tau); \lambda_1)$ is the change in the system’s Shannon entropy along the trajectory. The total entropy production $\Sigma$ satisfies the integral fluctuation theorem:

$\langle e^{-\Sigma} \rangle = 1,$

which implies $\langle \Sigma \rangle \geq 0$ by Jensen’s inequality — the ensemble-averaged entropy production is non-negative, as required by the second law.

The trajectory-level perspective reveals that entropy production is not merely a macroscopic bookkeeping quantity but a property of individual realizations of the stochastic process. Configurations can be assigned an instantaneous entropy production rate $\dot{\sigma}(\mathbf{q})$; the global entropy production rate in a non-equilibrium steady state is $\langle \dot{\sigma} \rangle$, which is zero if and only if the system satisfies detailed balance (i.e., is in equilibrium or a potential-driven system). Measuring $\langle \dot{\sigma} \rangle$ from trajectory data gives a quantitative measure of how far from equilibrium a system is operating.

[!remark] 8.2 Stochastic thermodynamics has important implications for biological molecular machines — ATP synthase, myosin, kinesin — which operate as non-equilibrium engines that convert chemical free energy (from ATP hydrolysis) into mechanical work. The efficiency of such machines is bounded by the Carnot limit only when heat flows between reservoirs at different temperatures; for isothermal machines driven by chemical potential differences, the relevant bound is the free energy available from ATP hydrolysis ($\sim 20\, k_BT$ per molecule under physiological conditions). Stochastic thermodynamics provides the theoretical framework for analyzing the efficiency of these machines from single-molecule trajectory data, connecting the observed fluctuations in step size and dwell time to the underlying free energy landscape and dissipation.

Linear Response and the Green–Kubo Relations

When the system is driven only weakly from equilibrium — the perturbation is small — the response is linear in the perturbation strength, and the linear response coefficients can be expressed as equilibrium correlation functions. This is linear response theory (Kubo, 1957).

Consider a system at equilibrium perturbed at $t = 0$ by a small external force $f$ coupling to an observable $A(\mathbf{q})$: the Hamiltonian becomes $H \to H - f\, A$. The response of another observable $B$ is, to first order in $f$:

$\langle B(t) \rangle_f - \langle B \rangle_0 = \beta f \int_0^t C_{BA}(t')\, dt',$

where $C_{BA}(t) = \langle B(t)\, A(0) \rangle_0 - \langle B \rangle_0 \langle A \rangle_0$ is the equilibrium time-correlation function. All non-equilibrium response information is encoded in the equilibrium fluctuations.

The Green–Kubo relations are a consequence: transport coefficients — diffusivity, viscosity, thermal conductivity, electrical conductivity — can be computed from the integral of the appropriate equilibrium autocorrelation function:

where $D$ is the diffusion coefficient (from the velocity autocorrelation function), $\eta$ is the shear viscosity (from the off-diagonal stress tensor autocorrelation $P_{xy}$), and $V$ is the volume. These relations allow transport properties to be computed from equilibrium MD simulations: run without any external force, record the appropriate correlation function, and integrate. The Green–Kubo approach is exact within the linear response regime and avoids the statistical problems of non-equilibrium simulations driven by large external perturbations.

[!remark] 8.3 The Einstein relations provide an equivalent route to transport coefficients via mean-squared displacements rather than autocorrelation functions. The diffusion coefficient, for example, satisfies

$D = \lim_{t \to \infty} \frac{\langle |\mathbf{r}(t) - \mathbf{r}(0)|^2 \rangle}{6t}.$

In practice, the Green–Kubo and Einstein approaches give identical results in the long-time limit but have different finite-time convergence properties. The velocity autocorrelation function decays to zero on the timescale of a few picoseconds for atomic liquids, making Green–Kubo efficient; the mean-squared displacement converges more slowly because it accumulates errors from the entire trajectory. For slow processes — diffusion of large molecules, viscosity of viscous fluids — the MSD converges faster because the integrand of the Green–Kubo formula is large at long times, requiring very long trajectories to converge.

Non-Equilibrium Steady States and Entropy Production

The fluctuation theorems above describe transient non-equilibrium processes: a protocol that starts and ends at definite states. Many biological and physical systems operate in non-equilibrium steady states (NESS): driven continuously by an external force or chemical gradient, they settle into a stationary distribution that is not the Boltzmann distribution and does not satisfy detailed balance.

In a NESS, the probability current $\mathbf{J}(\mathbf{q})$ is nonzero but divergence-free: $\nabla \cdot \mathbf{J} = 0$. The distribution $\rho_{\mathrm{ss}}(\mathbf{q})$ is stationary but is supported by a steady dissipation of free energy. The entropy production rate of the steady state is:

$\dot{S}_{\mathrm{tot}} = k_B \int \frac{|\mathbf{J}(\mathbf{q})|^2}{\rho_{\mathrm{ss}}(\mathbf{q})\, D}\, d\mathbf{q} \geq 0,$

which vanishes if and only if $\mathbf{J} = 0$ — i.e., if the system is in equilibrium. Computing $\dot{S}_{\mathrm{tot}}$ from simulation requires estimating the probability current from trajectory data, which in high dimensions requires either a good low-dimensional parametrization or density estimation methods capable of operating in $3N$ dimensions. Both remain active areas of research at the intersection of statistical mechanics and machine learning.

---

The Jarzynski equality was derived by Jarzynski (1997a, 1997b) and independently by Hummer and Szabo (2001) in the context of pulling experiments; its experimental verification in single-molecule RNA unfolding was provided by Liphardt et al. (2002). The Crooks fluctuation theorem was derived by Crooks (1999). Stochastic thermodynamics at the trajectory level was developed by Seifert (2005) and reviewed comprehensively by Seifert (2012). The Green–Kubo relations were derived by Green (1954) and Kubo (1957); the Einstein relations connecting diffusivity to the mean-squared displacement are due to Einstein (1905). The application of fluctuation theorems to non-equilibrium free energy calculations in molecular simulation was developed by Hummer and Szabo (2001) and Park and Schulten (2004). The connection between entropy production and irreversibility in non-equilibrium steady states was established by the work of Lebowitz and Spohn (1999) and reviewed in the context of molecular simulation by Seifert (2012) and Crooks (1999).