Jarzynski equality and Crooks fluctuation theorem

Intuition Beginner

The second law of thermodynamics says no engine extracts more useful work from a heat bath than the free energy allows. Pull a molecule slowly and the work you spend equals the free energy change. Pull it fast and you spend extra, dissipated as heat. Or do you? A single fast pull now and then gives back more than expected. The second law is a statement about averages, and a single measurement can violate it.

Repeating the same fast pull thousands of times produces a spread of work values. Most pulls cost more than the slow limit; a rare few cost less. The Jarzynski equality is the exact recipe for combining these scattered values. Weight each measurement with a decaying exponential of the work, average those weights, and take a logarithm. The answer is the free energy difference, recovered exactly from data that never reached equilibrium.

The Crooks theorem sharpens the picture. Run the protocol forward and then in reverse, collect the two work distributions, and compare. The distributions cross at exactly the free energy value. The further a process sits from reversible, the more lopsided the crossing becomes. This unit develops both relations as the second law lifted to the scale of individual fluctuations.

Visual Beginner

A schematic shows a single RNA hairpin held between an optical trap and a fixed bead. The trap is pulled away from the bead at a chosen speed, unfolding the hairpin. A histogram panel below plots two work distributions side by side: a forward distribution $P_F$ (fast pulling) and a reverse distribution $P_R$ (fast refolding). The two bell-shaped curves overlap and cross at a single work value, marked as the equilibrium free energy $\Delta F$. A vertical dashed line at the crossing makes the Crooks relation visible.

The picture captures the essential content: dissipated work measures how distinguishable the forward and reverse distributions are, and the crossing point is the equilibrium quantity recovered from nonequilibrium data.

Worked example Beginner

Consider an RNA hairpin driven out of equilibrium by a fast optical-trap pull. The equilibrium folding free energy is $\Delta F=10$ in units of $k_{B}T$. You repeat the fast pull three times and record the work values $W=8$, $W=12$, and $W=15$ (same units).

Step 1. Compute the arithmetic mean of the three work values: $(8+12+15)/3=35/3\approx 11.67$. This is greater than $\Delta F=10$, so the second law holds on average, as it must.

Step 2. Compute the Jarzynski average. For each measurement, evaluate $e^{-W}$: $e^{-8}\approx 0.000335$, $e^{-12}\approx 0.0000061$, $e^{-15}\approx 0.00000031$. The total of these three numbers is approximately $0.000342$.

Step 3. Divide by three to get the average: $0.000342/3\approx 0.000114$. Take the negative of the natural logarithm: $-\ln (0.000114)\approx 9.08$.

Step 4. The Jarzynski estimate of the free energy is $9.08$ in units of $k_{B}T$. The true value is $10$. With only three samples the estimate lands a little low, because the rare small-work events that dominate the exponential average were undersampled. Repeat the pull a thousand times and the estimate converges to $10$.

What this tells us: individual pulls may undershoot the equilibrium limit, the arithmetic mean always respects the second law, and the exponential average recovers the equilibrium free energy exactly once enough samples are collected.

Check your understanding Beginner

Formal definition Intermediate+

Let $\mathcal{S}$ be a classical system with microstate $x$ in a phase space $\Gamma$ and a Hamiltonian $H(x;\lambda )$ that depends on an externally controllable parameter $\lambda$ (trap position, field strength, volume). A protocol $\lambda (t):[0,\tau ]\to \mathbb{R}$ drives the system from ${\lambda }_A=\lambda (0)$ to ${\lambda }_B=\lambda (\tau )$. The system is prepared in equilibrium at temperature $T$ before the protocol begins, with initial density ${\rho }_A^{\mathrm{eq}}(x)=e^{-\beta H(x;{\lambda }_A)}/Z_A$ where $\beta =1/(k_{B}T)$ and $Z_A={\int }_{\Gamma }{e}^{-\beta H(x;{\lambda }_A)}dx$. The dynamics in between may be Hamiltonian, Langevin, or any Markov process satisfying detailed balance with respect to the instantaneous equilibrium distribution.

The work performed on the system along a trajectory $x(t)$ is the accumulated change in the Hamiltonian due to the explicit time dependence of the protocol ^{[Jarzynski 1997]}:

$$W[x]={\int }_0^{\tau }\frac{\partial H(x(t);\lambda (t))}{\partial t}dt={\int }_0^{\tau }\dot{\lambda }(t)\frac{\partial H}{\partial \lambda }{|}_{x(t),\lambda (t)}dt.$$

The heat exchanged with the bath is $Q[x]={\int }_0^{\tau }\dot{x}(t)\cdot {\nabla }_{x}Hdt$, and the first law along each trajectory reads $\Delta H[x]=W[x]+Q[x]$.

The equilibrium free energy difference is $\Delta F=F_B-F_A=-{\beta }^{-1}\ln (Z_B/Z_A)$ with $F(\lambda )=-{\beta }^{-1}\ln Z(\lambda )$. The dissipated work along a trajectory is $W_{\mathrm{diss}}[x]=W[x]-\Delta F$. A protocol run quasistatically ($\tau \to \infty$) gives $W=\Delta F$ for every trajectory; a finite-time protocol gives $W_{\mathrm{diss}}>0$ on average.

The two principal objects of this unit are the Jarzynski equality ^{[Jarzynski 1997]}

$$⟨e^{-\beta W}{⟩}_F=e^{-\beta \Delta F},$$

where $⟨\cdot {⟩}_F$ denotes the average over trajectories initiated at equilibrium at ${\lambda }_A$ and driven by the forward protocol $\lambda (t)$, and the Crooks fluctuation theorem ^{[Crooks 1999]}

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

where $P_F(W)$ is the probability density of work in the forward process ${\lambda }_A\to {\lambda }_B$ and $P_R(-W)$ is the density of work in the time-reversed process ${\lambda }_R(t)=\lambda (\tau -t)$ driving ${\lambda }_B\to {\lambda }_A$, evaluated at $-W$.

Counterexamples to common slips

• Confusing heat and work. The work $W$ is the change in $H$ due to the protocol's explicit $\lambda$-dependence; heat is the change due to the bath-driven motion of $x$. The Jarzynski equality concerns $W$ only, not $Q$. Mixing the two is the most common error when first reading trajectory-level thermodynamics.
• Assuming the equality holds without equilibrium initial conditions. The system must begin each repetition in equilibrium at $T$. Start it out of equilibrium and the Jarzynski equality fails; an additional term accounting for the relaxation of the initial state appears. The Crooks ratio also requires the reverse process to begin in equilibrium at ${\lambda }_B$.
• Treating $W$ as a state function. Work is a functional of the trajectory, not a function of the endpoints. Two repetitions of the same protocol produce different $W$ values; only the exponential average is fixed by $\Delta F$. The free energy $F$ is a state function; the work $W$ is not.
• Expecting finite-sample convergence. The estimator $-{\beta }^{-1}\ln ⟨e^{-\beta W}{⟩}_n$ is biased by the rare events that dominate the exponential. With $n$ finite samples the estimate systematically undershoots $\Delta F$; convergence is slow (the variance of the exponential weight is governed by the dissipated work itself). This is the experimental difficulty addressed by Hummer-Szabo and by the Collin-Ritort protocol.

Key theorem with proof Intermediate+

Theorem (Crooks fluctuation theorem and Jarzynski equality). Let $\mathcal{S}$ be a system with Hamiltonian $H(x;\lambda )$ driven by a forward protocol ${\lambda }_F(t):{\lambda }_A\to {\lambda }_B$ over time $\tau$, with the system initially in equilibrium at temperature $T$. Let ${\lambda }_R(t)={\lambda }_F(\tau -t)$ denote the time-reversed protocol. Assume the dynamics satisfy microscopic reversibility (detailed balance at the trajectory level). Then for every trajectory $x(t)$ with time-reverse $\tilde{x}(t)=x(\tau -t)$,

$$\frac{{\mathcal{P}}_F[x]}{{\mathcal{P}}_R[\tilde{x}]}=e^{\beta (W[x]-\Delta F)},$$

and marginalising over all trajectories producing a given work $W$ yields

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

Integrating against $e^{-\beta W}$ gives

$$⟨e^{-\beta W}{⟩}_F=e^{-\beta \Delta F}\text{(Jarzynski)}.$$

By Jensen's inequality applied to the convex function $e^{-\beta (\cdot )}$,

$$⟨W{⟩}_F\ge \Delta F,$$

which is the second law of thermodynamics for the average work performed on an isothermal system.

Proof. Decompose the trajectory probability into an initial-condition factor and a dynamical factor. For the forward process,

$${\mathcal{P}}_F[x]={\rho }_A^{\mathrm{eq}}(x(0))\cdot {\mathcal{M}}_F[x|x(0)],$$

where ${\rho }_A^{\mathrm{eq}}(x)=e^{-\beta H(x;{\lambda }_A)}/Z_A$ is the equilibrium density at ${\lambda }_A$ and ${\mathcal{M}}_F[x|x(0)]$ is the conditional probability density of the trajectory given the initial point, generated by the Markov dynamics (Hamiltonian flow, or Langevin/Glauber/Metropolis dynamics satisfying detailed balance with respect to the instantaneous Boltzmann distribution). The reverse trajectory probability is

$${\mathcal{P}}_R[\tilde{x}]={\rho }_B^{\mathrm{eq}}(\tilde{x}(0))\cdot {\mathcal{M}}_R[\tilde{x}|\tilde{x}(0)]={\rho }_B^{\mathrm{eq}}(x(\tau ))\cdot {\mathcal{M}}_R[\tilde{x}|x(\tau )],$$

with ${\rho }_B^{\mathrm{eq}}(x)=e^{-\beta H(x;{\lambda }_B)}/Z_B$.

Take the ratio and split it into an endpoint factor and a dynamical factor:

$$\frac{{\mathcal{P}}_F[x]}{{\mathcal{P}}_R[\tilde{x}]}=\frac{{\rho }_A^{\mathrm{eq}}(x(0))}{{\rho }_B^{\mathrm{eq}}(x(\tau ))}\cdot \frac{{\mathcal{M}}_F[x|x(0)]}{{\mathcal{M}}_R[\tilde{x}|x(\tau )]}.$$

The endpoint factor evaluates to

$$\frac{{\rho }_A^{\mathrm{eq}}(x(0))}{{\rho }_B^{\mathrm{eq}}(x(\tau ))}=\frac{Z_B}{Z_A}\exp [-\beta H(x(0);{\lambda }_A)+\beta H(x(\tau );{\lambda }_B)]=e^{-\beta \Delta F}\cdot e^{\beta \Delta H[x]},$$

using $Z_B/Z_A=e^{-\beta \Delta F}\cdot e^{\beta F_B}/e^{\beta F_A}$ rearranged through $\Delta F=-{\beta }^{-1}\ln (Z_B/Z_A)$, which gives $Z_B/Z_A=e^{-\beta \Delta F}$.

The dynamical factor is governed by microscopic reversibility. For Hamiltonian dynamics this is the ratio of phase-space volume elements, which equals $1$ by Liouville's theorem. For Langevin or Metropolis dynamics satisfying detailed balance with the instantaneous equilibrium ${\rho }_t^{\mathrm{eq}}(x)\propto e^{-\beta H(x;\lambda (t))}$, the ratio of the forward path probability to the reverse path probability is

$$\frac{{\mathcal{M}}_F[x|x(0)]}{{\mathcal{M}}_R[\tilde{x}|x(\tau )]}=\exp (-\beta Q[x]),$$

where $Q[x]=\Delta H[x]-W[x]$ is the heat exchanged with the bath along the trajectory (the energy change not due to the explicit $\lambda$-dependence). This is the trajectory form of detailed balance: each infinitesimal transition satisfies $m(x\to x^{\prime };\lambda )/m(x^{\prime }\to x;\lambda )=e^{-\beta [H(x^{\prime };\lambda )-H(x;\lambda )]}$, and multiplying over the trajectory gives the path ratio $e^{-\beta Q}$. Substituting both factors,

$$\frac{{\mathcal{P}}_F[x]}{{\mathcal{P}}_R[\tilde{x}]}=e^{-\beta \Delta F}\cdot e^{\beta \Delta H[x]}\cdot e^{-\beta Q[x]}=e^{\beta (\Delta H[x]-Q[x]-\Delta F)}=e^{\beta (W[x]-\Delta F)},$$

using $\Delta H=W+Q$. This establishes the trajectory-level Crooks ratio ^{[Crooks 1998 thesis]}.

To obtain the work-distribution form, collect all forward trajectories with work value $W$ into the set ${\Gamma }_F(W)$ and all reverse trajectories with work $-W$ into ${\Gamma }_R(-W)$. The time-reversal map $x\mapsto \tilde{x}$ is a bijection between ${\Gamma }_F(W)$ and ${\Gamma }_R(-W)$, so marginalising the trajectory ratio gives

$$P_F(W)={\int }_{{\Gamma }_F(W)}{\mathcal{P}}_F[x]\mathcal{D}x={\int }_{{\Gamma }_R(-W)}{e}^{\beta (W-\Delta F)}{\mathcal{P}}_R[\tilde{x}]\mathcal{D}\tilde{x}=e^{\beta (W-\Delta F)}{P}_R(-W),$$

yielding the Crooks fluctuation theorem ^{[Crooks 1999]}.

The Jarzynski equality follows on integrating the Crooks relation against $e^{-\beta W}$:

$$⟨e^{-\beta W}{⟩}_F={\int }_{-\infty }^{\infty }{e}^{-\beta W}{P}_F(W)dW={\int }_{-\infty }^{\infty }{e}^{-\beta W}{e}^{\beta (W-\Delta F)}{P}_R(-W)dW=e^{-\beta \Delta F}{\int }_{-\infty }^{\infty }{P}_R(-W)dW.$$

The final integral equals $1$ by normalisation of $P_R$, giving $⟨e^{-\beta W}{⟩}_F=e^{-\beta \Delta F}$.

For the second law, apply Jensen's inequality to the convex function $\phi (w)=e^{-\beta w}$: $\phi (⟨W⟩)\le ⟨\phi (W)⟩$, i.e. $e^{-\beta ⟨W⟩}\le e^{-\beta \Delta F}$, and since the exponential is monotone increasing, $⟨W⟩\ge \Delta F$. Equality holds if and only if $W$ is deterministic, which occurs only for a quasistatic reversible protocol. $\square$

Bridge. This builds toward the general theory of nonequilibrium work relations and appears again in 08.12.01 (the fluctuation-dissipation theorem), whose linear-response kernel governs the leading-order dissipation that the present theorem bounds exactly. The foundational reason the exponential average suppresses the rare dissipative trajectories is microscopic reversibility — this is exactly the Crooks trajectory ratio $P_F[x]/P_R[\tilde{x}]=e^{\beta (W-\Delta F)}$, and putting these together identifies the dissipated work with a Kullback-Leibler divergence. The bridge is that the second law is the Jensen-inequality shadow of an exact identity: the central insight is that fluctuation theorems promote the inequality $⟨W⟩\ge \Delta F$ to an equality at the cost of exponentiating each trajectory.

Exercises Intermediate+

Advanced results Master

Theorem (dissipated work as relative entropy; Kawai-Parrondo-van den Broeck 2007). For a system driven by a protocol $\lambda (t)$ with forward phase-space density ${\rho }_F(x,t)$ and its time-reversed counterpart ${\tilde{\rho }}_R(x,t)$ under the reverse protocol, the average dissipated work equals

$$⟨W_{\mathrm{diss}}⟩=k_{B}T{\int }_0^{\tau }{\int }_{\Gamma }{\rho }_F(x,t)\ln \frac{{\rho }_F(x,t)}{{\tilde{\rho }}_R(\tilde{x},\tau -t)}dxdt=k_{B}T\cdot D_{\mathrm{KL}}({\rho }_F\Vert {\tilde{\rho }}_R),$$

where $\tilde{x}$ is the phase-space point reached by applying the momentum-reversal map to $x$. In particular $⟨W_{\mathrm{diss}}⟩\ge 0$, with equality only when ${\rho }_F={\tilde{\rho }}_R$ at every time, which is the condition for a quasistatic reversible protocol ^{[Kawai Parrondo van den Broeck 2007]}.

The result refines Exercise 5: the work-distribution KL divergence is the data-processing projection of the underlying phase-space relative entropy. Dissipation measures how distinguishable the actual forward evolution is from its time reverse. The bridge to information theory is immediate — a protocol that dissipates $5k_{B}T$ does work that is, in bits, $\beta ⟨W_{\mathrm{diss}}⟩/\ln 2$ distinguishable from its time reverse.

Theorem (stochastic entropy production and Seifert's integral fluctuation theorem; Seifert 2005). For a Markov process with trajectory $x(t)$ and instantaneous phase-space density $\rho (x,t)$, the total entropy produced along the trajectory is $\Delta s_{\mathrm{tot}}[x]=\ln ({\mathcal{P}}_F[x]/{\mathcal{P}}_R[\tilde{x}])=\ln ({\rho }_F(x(0),0)/{\rho }_F(x(\tau ),\tau ))+\beta Q[x]$. It satisfies the integral fluctuation theorem

$$⟨e^{-\Delta s_{\mathrm{tot}}}⟩=1,$$

from which $⟨\Delta s_{\mathrm{tot}}⟩\ge 0$ and $\mathrm{Pr}(\Delta s_{\mathrm{tot}}<-\sigma )\le e^{-\sigma }$ follow by Jensen and the tail bound ^{[Seifert 2005]}.

Seifert's theorem extends the framework to systems that begin out of equilibrium, where there is no equilibrium free energy to reference. The total entropy production is the log ratio of forward to reverse trajectory probabilities; the Jarzynski equality and the Crooks theorem are the specialisation to the case where the initial and final distributions are equilibrium Boltzmann distributions.

Theorem (transient and steady-state fluctuation theorems; Evans-Searles 1993, Gallavotti-Cohen 1995). Let ${\Omega }_t$ denote the time-averaged entropy-production rate over a window $[0,t]$ for an anoergodic system driven by a thermostatted nonequilibrium protocol. The Evans-Searles transient fluctuation theorem gives

$$\frac{\mathrm{Pr}({\Omega }_t=A)}{\mathrm{Pr}({\Omega }_t=-A)}=e^{At},$$

for every $A>0$. For chaotic dynamical systems satisfying the chaotic hypothesis, the Gallavotti-Cohen fluctuation theorem states that the large-deviation rate function $\zeta (\sigma )$ for the entropy-production rate satisfies the symmetry $\zeta (-\sigma )-\zeta (\sigma )=\sigma$ ^{[Evans Searles 1993; Gallavotti Cohen 1995]}.

These theorems precede Jarzynski and Crooks historically and situate them within a broader programme: the second law is one member of a family of exact fluctuation symmetries. The Gallavotti-Cohen symmetry is a statement about the large-deviation rate function (cf. 08.12.02 on Cramér-Sanov theory); the Crooks ratio is its sharp, finite-time form specialised to the work observable.

Theorem (thermodynamic length and minimum dissipation; Salamon-Nulton 1984, Sivak-Crooks 2012-2016). Let $\lambda \mapsto {\rho }_{\lambda }^{\mathrm{eq}}$ be a path in the manifold of equilibrium states parametrised by the control parameter $\lambda$. Equip this manifold with the Fisher-information metric $g_{\mu \nu }(\lambda )=⟨{\partial }_{\mu }\ln {\rho }_{\lambda }^{\mathrm{eq}}{\partial }_{\nu }\ln {\rho }_{\lambda }^{\mathrm{eq}}⟩$. The thermodynamic length of a protocol $\lambda (t)$ over time $\tau$ is

$$\mathcal{L}[\lambda ]={\int }_0^{\tau }\sqrt{\dot{\lambda }(t{)}^{\top }g(\lambda (t))\dot{\lambda }(t)}dt.$$

For a protocol with fixed endpoints and total time $\tau$, the excess dissipation satisfies the lower bound

$$⟨W_{\mathrm{diss}}⟩\ge \frac{1}{2}{k}_{B}T\frac{{\mathcal{L}}_{\min }^2}{\tau },$$

where ${\mathcal{L}}_{\min }$ is the thermodynamic length of the geodesic connecting the endpoints. The bound is saturated by protocols that traverse the equilibrium manifold at constant thermodynamic speed ^{[Sivak Crooks 2016]}.

The Fisher metric converts the space of equilibrium states into a Riemannian manifold, and dissipation into a geometric cost. This is the bridge between fluctuation theorems and optimal-control theory: the protocols that minimise dissipation are the geodesics of this manifold, and finite-time thermodynamics becomes a problem in Riemannian geometry.

Theorem (quantum fluctuation theorems; Kurchan 2000, Tasaki 2000). For a quantum system prepared in the Gibbs state ${\rho }_A^{\mathrm{eq}}=e^{-\beta H_A}/Z_A$, driven by a unitary protocol $U=\mathcal{T}\exp (-i\int H(\lambda (t))dt)$, the work performed in a single realisation is defined by projective energy measurements at the start and end: $W=E_B-E_A$ where $E_A$ ($E_B$) is the outcome of measuring $H_A$ ($H_B$). The two-point measurement fluctuation theorem reads

$$⟨e^{-\beta W}⟩=\frac{\mathrm{Tr}{e}^{-\beta H_B}}{\mathrm{Tr}{e}^{-\beta H_A}}=e^{-\beta \Delta F},$$

recovering the Jarzynski equality in the quantum regime. Coherences generated between the two measurements do not contribute to the work average in the two-point scheme ^{[Tasaki 2000; Campisi Hänggi Talkner 2011]}.

The quantum case is subtle: work is not an observable but a property of a pair of measurements. Alternative schemes (full-counting statistics, Bayesian) handle coherence differently; the quantum-classical correspondence is recovered in the limit of commuting Hamiltonians.

Theorem (information-theoretic second law and Maxwell's demon; Sagawa-Ueda 2010). For a protocol with feedback based on a measurement outcome $m$ carrying mutual information $I=⟨\ln (p(m|x)/p(m))⟩$ with the system microstate, the work extractable is bounded by

$$⟨W⟩\ge \Delta F-k_{B}T\cdot I.$$

Acquired information lowers the free-energy bound by $k_{B}T$ per nat; measurement and erasure costs (Landauer's principle) restore the unconditional second law when the demon's memory is reset ^{[Sagawa Ueda 2010]}.

This theorem closes the loop opened by Maxwell's demon in 1867: information is thermodynamically fungible with work, and the fluctuation-theorem framework quantifies the exchange rate. Single-molecule implementations (Toyabe et al. 2010) verified the bound directly.

Synthesis. The foundational reason the second law survives at the fluctuation scale is microscopic reversibility: the trajectory ratio $P_F[x]/P_R[\tilde{x}]=e^{\beta (W-\Delta F)}$ forces the forward and reverse work distributions to cross at $\Delta F$, and this is exactly the Crooks theorem. The central insight is that dissipated work measures how distinguishable the two distributions are — putting these together identifies $⟨W_{\mathrm{diss}}⟩$ with $k_{B}T\cdot D_{\mathrm{KL}}(P_F\Vert P_R)$, a non-negative quantity that vanishes only for a quasistatic protocol. This builds toward the geometric theory of thermodynamic length, where the Fisher-information metric on the manifold of equilibrium states fixes the minimum dissipation, and appears again in the linear-response regime of 08.12.01 whose kernel governs the leading-order dissipation that the present theorem bounds exactly. The bridge is that every nonequilibrium work relation — Jarzynski, Crooks, Evans-Searles, Gallavotti-Cohen, Seifert, Hatano-Sasa — is a facet of the same trajectory-level detailed balance; the pattern generalises to quantum two-point fluctuation theorems (Tasaki, Kurchan) and to feedback-driven demons (Sagawa-Ueda), where acquired information enters the bound as a negative entropy term.

Full proof set Master

Proposition (Kullback-Leibler identity for dissipated work). Under the Crooks fluctuation theorem, $D_{\mathrm{KL}}(P_F\Vert P_R\circ (-))=\beta ⟨W_{\mathrm{diss}}⟩$.

Proof. By definition $D_{\mathrm{KL}}(P_F\Vert P_R\circ (-))=\int P_F(W)\ln (P_F(W)/P_R(-W))dW$, whenever the integral exists. The Crooks theorem gives $\ln (P_F(W)/P_R(-W))=\beta (W-\Delta F)$ for every $W$ in the support. Substituting,

$$D_{\mathrm{KL}}(P_F\Vert P_R\circ (-))=\int P_F(W)\beta (W-\Delta F)dW=\beta (⟨W{⟩}_F-\Delta F)\cdot \int P_F(W)dW=\beta ⟨W_{\mathrm{diss}}⟩,$$

using $\int P_F(W)dW=1$. Non-negativity of relative entropy yields the second law $⟨W_{\mathrm{diss}}⟩\ge 0$. Equality holds iff $P_F(W)=P_R(-W)$ almost everywhere, which under Crooks forces $e^{\beta (W-\Delta F)}=1$ on the support, i.e. $W=\Delta F$ almost surely, the quasistatic limit. $\square$

Proposition (Jarzynski equality via Hamiltonian ergodicity). For an isolated Hamiltonian system initially in equilibrium at ${\lambda }_A$ and driven by $\lambda (t)$, the Jarzynski equality $⟨e^{-\beta W}⟩=e^{-\beta \Delta F}$ holds, with the average taken over the equilibrium initial distribution.

Proof. Work equals the change in the Hamiltonian: $W[x]=H(x(\tau );{\lambda }_B)-H(x(0);{\lambda }_A)=\Delta H[x]$, since an isolated Hamiltonian system exchanges no heat ($Q=0$). The map $x(0)\mapsto x(\tau )$ generated by Hamilton's equations preserves phase-space volume (Liouville's theorem), so the distribution of $x(\tau )$ is the push-forward of the equilibrium distribution ${\rho }_A^{\mathrm{eq}}$ under the Hamiltonian flow. Therefore

$$⟨e^{-\beta W}⟩={\int }_{\Gamma }{e}^{-\beta [H(x(\tau );{\lambda }_B)-H(x(0);{\lambda }_A)]}{\rho }_A^{\mathrm{eq}}(x(0))dx(0).$$

Using ${\rho }_A^{\mathrm{eq}}(x)=e^{-\beta H(x;{\lambda }_A)}/Z_A$, the integrand becomes $e^{-\beta H(x(\tau );{\lambda }_B)}\cdot e^{\beta H(x(0);{\lambda }_A)}\cdot e^{-\beta H(x(0);{\lambda }_A)}/Z_A=e^{-\beta H(x(\tau );{\lambda }_B)}/Z_A$. By Liouville's theorem the change of variables $x(0)\to x(\tau )$ has unit Jacobian, so

$$⟨e^{-\beta W}⟩=\frac{1}{Z_A}{\int }_{\Gamma }{e}^{-\beta H(x(\tau );{\lambda }_B)}dx(\tau )=\frac{Z_B}{Z_A}=e^{-\beta \Delta F}.$$

The same form holds for systems coupled to a heat bath via a thermostatted or Langevin dynamics that preserves the equilibrium distribution in the adiabatic limit, with an additional Jacobian factor that cancels by microscopic reversibility. $\square$

Proposition (thermodynamic-length lower bound on dissipation). For a slow protocol $\lambda (t)$ with total time $\tau$ connecting ${\lambda }_A$ to ${\lambda }_B$ in the manifold of equilibrium states equipped with the Fisher metric $g_{\mu \nu }$, the excess dissipation satisfies $⟨W_{\mathrm{diss}}⟩\ge \frac{1}{2}{k}_{B}T{\mathcal{L}}_{\min }^2/\tau$.

Proof sketch. In the slow-protocol regime the system remains near equilibrium, $\rho (x,t)={\rho }_{\lambda (t)}^{\mathrm{eq}}(x)+\delta \rho$ with $\delta \rho =O(1/\tau )$. Linear response (cf. 08.12.01) gives $\delta \rho =-{\int }_{-\infty }^t{\partial }_s{\rho }_{\lambda (s)}^{\mathrm{eq}}ds$, and the dissipated power is ${\dot{W}}_{\mathrm{diss}}={\dot{\lambda }}^{\top }g(\lambda )\dot{\lambda }/k_{B}T$ plus corrections of order $1/{\tau }^2$ (the symmetric Onsager matrix equals the Fisher metric up to a bath factor, by the fluctuation-dissipation relation of 08.12.01). Integrating over $t\in [0,\tau ]$ gives $⟨W_{\mathrm{diss}}⟩=(k_{B}T{)}^{-1}{\int }_0^{\tau }{\dot{\lambda }}^{\top }g\dot{\lambda }dt$. By the Cauchy-Schwarz inequality applied with the metric $g$,

$${\int }_0^{\tau }{\dot{\lambda }}^{\top }g\dot{\lambda }dt\ge \frac{1}{\tau }{\left({\int }_0^{\tau }\sqrt{{\dot{\lambda }}^{\top }g\dot{\lambda }}dt\right)}^2=\frac{\mathcal{L}[\lambda {]}^2}{\tau }\ge \frac{{\mathcal{L}}_{\min }^2}{\tau },$$

and multiplying by $(k_{B}T{)}^{-1}$ gives the bound up to the bath prefactor; rescaling conventions for $g$ yields $⟨W_{\mathrm{diss}}⟩\ge \frac{1}{2}{k}_{B}T{\mathcal{L}}_{\min }^2/\tau$. Equality is approached by geodesic protocols traversed at constant thermodynamic speed. $\square$

Connections Master

• Equilibrium fluctuations of thermodynamic quantities 08.12.02. The present unit is the nonequilibrium companion to the equilibrium Gaussian fluctuation theory of the sibling; the equilibrium variances $⟨(\Delta E{)}^2⟩=k_B{T}^2{C}_V$ of 08.12.02 set the Fisher metric that governs the minimum dissipation derived here, and the Einstein formula $W\propto e^{\Delta S/k_B}$ of the sibling is the static shadow of the trajectory-level Crooks ratio.

• Fluctuation-dissipation theorem (Landau-Callen-Welton) 08.12.01. Linear response near equilibrium is the leading-order limit of the present theorem: the FDT kernel fixes the dissipated work to first order in the protocol speed, while the Jarzynski equality and Crooks theorem give the exact, all-orders relation. The Fisher metric of thermodynamic length equals the zero-frequency limit of the FDT response function.

• Free energy 08.01.04. The Helmholtz free-energy difference $\Delta F=F_B-F_A$ between equilibrium states is the equilibrium quantity recovered exactly by nonequilibrium work averages; every protocol, however irreversible, encodes $\Delta F$ in its work distribution through the Jarzynski equality.

• Partition function 08.01.01. The ratio $Z_B/Z_A=e^{-\beta \Delta F}$ that defines the free-energy difference in 08.01.01 is the endpoint factor that enters the trajectory-level Crooks ratio; the partition function is the bridge between the equilibrium thermodynamics of the endpoints and the nonequilibrium work done in between.

• Boltzmann distribution 08.01.03. The equilibrium initial density ${\rho }^{\mathrm{eq}}=e^{-\beta H}/Z$ of 08.01.03 is the boundary condition required by both the Jarzynski equality and the Crooks theorem; drop it (start out of equilibrium) and the theorems acquire the Seifert entropy-production correction.

Historical & philosophical context Master

The trajectory from Loschmidt's 1876 Umkehreinwand (reversibility objection) against Boltzmann's $H$-theorem to the modern fluctuation relations took just over a century. Loschmidt observed that Newton's equations are time-reversal symmetric, so any trajectory that decreases the Boltzmann entropy has a time-reversed partner that increases it; Boltzmann's response was probabilistic — the increasing-entropy trajectories overwhelm the decreasing ones statistically. Quantifying that statistical imbalance precisely waited for the work of Denis Evans, Edith Searles, and Giovanni Gallavotti in the 1990s. Evans and Searles established the transient fluctuation theorem $\mathrm{Pr}({\Omega }_t=A)/\mathrm{Pr}({\Omega }_t=-A)=e^{At}$ in 1993-1994 ^{[Evans Searles 1993]}; Gallavotti and Cohen, building on the chaotic hypothesis, proved the steady-state fluctuation theorem for the entropy-production rate function in 1995 ^{[Gallavotti Cohen 1995]}.

Christopher Jarzynski's 1997 paper Nonequilibrium equality for free energy differences (Phys. Rev. Lett. 78, 2690) ^{[Jarzynski 1997]} condensed the programme into a single exact identity. By exponentiating the work performed on a system driven out of equilibrium and averaging over repetitions begun in equilibrium, one recovers the equilibrium free-energy difference exactly: $⟨e^{-\beta W}⟩=e^{-\beta \Delta F}$. The equality holds for arbitrary protocol speed and arbitrary distance from equilibrium; the second law drops out as the Jensen-inequality shadow. Gavin Crooks, in his 1998 Berkeley PhD thesis and the 1999 paper Entropy production fluctuation theorem and the nonequilibrium work relation (Phys. Rev. E 60, 2721) ^{[Crooks 1999]}, supplied the deeper trajectory-level statement: the forward and reverse work distributions satisfy $P_F(W)/P_R(-W)=e^{\beta (W-\Delta F)}$, and the Jarzynski equality is the integral of this ratio. The Crooks theorem makes the role of microscopic reversibility explicit, tracing the second law to a ratio of trajectory probabilities.

Experimental verification came from single-molecule biophysics. Jan Liphardt, Carlos Bustamante and collaborators pulled individual RNA hairpins with optical tweezers and verified the Jarzynski equality on single molecules (Science 296, 1832, 2002) ^{[Liphardt 2002]}, reconstructing the equilibrium folding free energy from irreversible pulls. Felix Ritort, Bustamante, and collaborators then verified the Crooks fluctuation theorem directly in 2005 (Nature 437, 231) ^{[Collin 2005]}, showing that the forward and reverse work distributions cross at exactly the equilibrium $\Delta F$. Udo Seifert's 2005 stochastic-entropy-production formulation ^{[Seifert 2005]} extended the framework to arbitrarily driven systems without an equilibrium reference, and the Sagawa-Ueda information-theoretic second law of 2010 ^{[Sagawa Ueda 2010]} closed the loop with Maxwell's demon by pricing information in units of $k_{B}T$. The Kawai-Parrondo-van den Broeck 2007 identification of dissipated work with Kullback-Leibler divergence ^{[Kawai Parrondo van den Broeck 2007]} gave the information-theoretic reframing its sharpest form.

Bibliography Master

@article{Jarzynski1997,
  author  = {Jarzynski, C.},
  title   = {Nonequilibrium equality for free energy differences},
  journal = {Physical Review Letters},
  volume  = {78},
  number  = {14},
  year    = {1997},
  pages   = {2690--2693}
}

@article{Crooks1999,
  author  = {Crooks, Gavin E.},
  title   = {Entropy production fluctuation theorem and the nonequilibrium work relation for free energy differences},
  journal = {Physical Review E},
  volume  = {60},
  number  = {3},
  year    = {1999},
  pages   = {2721--2726}
}

@phdthesis{Crooks1998,
  author  = {Crooks, Gavin E.},
  title   = {Nonequilibrium Measurements of Free Energy Differences for Microscopically Reversible {M}arkovian Systems},
  school  = {University of California, Berkeley},
  year    = {1998}
}

@article{EvansSearles1994,
  author  = {Evans, Denis J. and Searles, Debra J.},
  title   = {Equilibrium microstates which generate second law violating steady states},
  journal = {Physical Review E},
  volume  = {50},
  number  = {2},
  year    = {1994},
  pages   = {1645--1648}
}

@article{GallavottiCohen1995,
  author  = {Gallavotti, Giovanni and Cohen, E. G. D.},
  title   = {Dynamical ensembles in nonequilibrium statistical mechanics},
  journal = {Physical Review Letters},
  volume  = {74},
  number  = {15},
  year    = {1995},
  pages   = {2694--2697}
}

@article{KawaiParrondoBroeck2007,
  author  = {Kawai, Ryoichi and Parrondo, Juan M. R. and van den Broeck, Christian},
  title   = {Dissipation: The phase-space perspective},
  journal = {Physical Review Letters},
  volume  = {98},
  number  = {8},
  year    = {2007},
  pages   = {080602}
}

@article{Seifert2005,
  author  = {Seifert, Udo},
  title   = {Entropy production along a stochastic trajectory and an integral fluctuation theorem},
  journal = {Physical Review Letters},
  volume  = {95},
  number  = {4},
  year    = {2005},
  pages   = {040602}
}

@article{Liphardt2002,
  author  = {Liphardt, Jan and Dumont, Sophie and Smith, Steven B. and Tinoco, Ignacio and Bustamante, Carlos},
  title   = {Equilibrium information from nonequilibrium measurements in experimental realizations of the {J}arzynski equality},
  journal = {Science},
  volume  = {296},
  number  = {5574},
  year    = {2002},
  pages   = {1832--1835}
}

@article{Collin2005,
  author  = {Collin, Damond and Ritort, Felix and Jarzynski, Christopher and Smith, Steven B. and Tinoco, Ignacio and Bustamante, Carlos},
  title   = {Verification of the {C}rooks fluctuation theorem and recovery of {RNA} folding free energies},
  journal = {Nature},
  volume  = {437},
  number  = {7056},
  year    = {2005},
  pages   = {231--234}
}

@article{HummerSzabo2001,
  author  = {Hummer, Gerhard and Szabo, Attila},
  title   = {Free energy reconstruction from nonequilibrium single-molecule pulling experiments},
  journal = {Proceedings of the National Academy of Sciences USA},
  volume  = {98},
  number  = {7},
  year    = {2001},
  pages   = {3658--3661}
}

@article{Tasaki2000,
  author  = {Tasaki, Hal},
  title   = {Jarzynski relations for quantum systems and their applications to thermodynamics of quantum states},
  journal = {arXiv:cond-mat/0009244},
  year    = {2000}
}

@article{Kurchan2000,
  author  = {Kurchan, Jorge},
  title   = {Quantum detailed balance conditions},
  journal = {Physical Review E},
  volume  = {61},
  number  = {5},
  year    = {2000},
  pages   = {5533--5539}
}

@article{CampisiHanggiTalkner2011,
  author  = {Campisi, Michele and H{\"a}nggi, Peter and Talkner, Peter},
  title   = {Colloquium: Quantum fluctuation relations: Foundations and applications},
  journal = {Reviews of Modern Physics},
  volume  = {83},
  number  = {3},
  year    = {2011},
  pages   = {771--791}
}

@article{SagawaUeda2010,
  author  = {Sagawa, Takahiro and Ueda, Masahito},
  title   = {Minimal energy cost for thermodynamic information processing: Measurement and memory erasure},
  journal = {Physical Review Letters},
  volume  = {104},
  number  = {9},
  year    = {2010},
  pages   = {090602}
}

@article{SivakCrooks2012,
  author  = {Sivak, David A. and Crooks, Gavin E.},
  title   = {Thermodynamic metrics and optimal paths},
  journal = {Physical Review Letters},
  volume  = {108},
  number  = {19},
  year    = {2012},
  pages   = {190602}
}

@book{Sekimoto2010,
  author    = {Sekimoto, Ken},
  title     = {Stochastic Energetics},
  publisher = {Springer},
  series    = {Lecture Notes in Physics},
  volume    = {799},
  year      = {2010}
}